Datasets:

memray

/

nus

like 0

1

2-Source Dispersers for Sub-Polynomial Entropy and Ramsey Graphs Beating the Frankl-Wilson Construction

The main result of this paper is an explicit disperser for two independent sources on n bits, each of entropy k = n o(1). Put differently, setting N = 2n and K = 2k , we construct explicit N N Boolean matrices for which no K K sub-matrix is monochromatic. Viewed as adjacency matrices of bipartite graphs, this gives an explicit construction of K-Ramsey bipartite graphs of size N . This greatly improves the previous bound of k = o(n) of Barak, Kindler, Shaltiel, Sudakov and Wigderson [4]. It also significantly improves the 25-year record of k = ~ O(n) on the special case of Ramsey graphs, due to Frankl and Wilson [9]. The construction uses (besides "classical" extractor ideas) almost all of the machinery developed in the last couple of years for extraction from independent sources, including: Bourgain's extractor for 2 independent sources of some entropy rate < 1/2 [5] Raz's extractor for 2 independent sources, one of which has any entropy rate > 1/2 [18] Rao's extractor for 2 independent block-sources of entropy n (1) [17] The "Challenge-Response" mechanism for detecting "entropy concentration" of [4]. The main novelty comes in a bootstrap procedure which allows the Challenge-Response mechanism of [4] to be used with sources of less and less entropy, using recursive calls to itself. Subtleties arise since the success of this mechanism depends on restricting the given sources, and so recursion constantly changes the original sources. These are resolved via a new construct, in between a disperser and an extractor, which behaves like an extractor on sufficiently large subsources of the given ones. This version is only an extended abstract, please see the full version, available on the authors' homepages, for more details.

INTRODUCTION This paper deals with randomness extraction from weak random sources. Here a weak random source is a distribution which contains some entropy. The extraction task is to design efficient algorithms (called extractors) to convert this entropy into useful form, namely a sequence of independent unbiased bits. Beyond the obvious motivations (potential use of physical sources in pseudorandom generators and in derandomization), extractors have found applications in a variety of areas in theoretical computer science where randomness does not seem an issue, such as in efficient constructions of communication networks [24, 7], error correcting codes [22, 12], data structures [14] and more. Most work in this subject over the last 20 years has focused on what is now called seeded extraction, in which the extractor is given as input not only the (sample from the) defective random source, but also a few truly random bits (called the seed). A comprehensive survey of much of this body of work is [21]. Another direction, which has been mostly dormant till about two years ago, is (seedless, deterministic) extraction from a few independent weak sources. This kind of extraction is important in several applications where it is unrealis-tic to have a short random seed or deterministically enumerate over its possible values. However, it is easily shown to be impossible when only one weak source is available. When at least 2 independent sources are available extraction becomes possible in principle. The 2-source case is the one we will focus on in this work. The rest of the introduction is structured as follows. We'll start by describing our main result in the context of Ramsey graphs. We then move to the context of extractors and disperser , describing the relevant background and stating our result in this language. Then we give an overview of the construction of our dispersers, describing the main building blocks we construct along the way. As the construction is quite complex and its analysis quite subtle, in this proceedings version we try to abstract away many of the technical difficulties so that the main ideas, structure and tools used are highlighted. For that reason we also often state definitions and theorems somewhat informally. 1.1 Ramsey Graphs Definition 1.1. A graph on N vertices is called a K-Ramsey Graph if it contains no clique or independent set of size K. In 1947 Erd os published his paper inaugurating the Prob-abilistic Method with a few examples, including a proof that most graphs on N = 2 n vertices are 2n-Ramsey. The quest for constructing such graphs explicitly has existed ever since and lead to some beautiful mathematics. The best record to date was obtained in 1981 by Frankl and Wilson [9], who used intersection theorems for set systems to construct N -vertex graphs which are 2 n log n -Ramsey. This bound was matched by Alon [1] using the Polynomial Method, by Grolmusz [11] using low rank matrices over rings, and also by Barak [2] boosting Abbot's method with almost k-wise independent random variables (a construction that was independently discovered by others as well). Remark-ably all of these different approaches got stuck at essentially the same bound. In recent work, Gopalan [10] showed that other than the last construction, all of these can be viewed as coming from low-degree symmetric representations of the OR function. He also shows that any such symmetric representation cannot be used to give a better Ramsey graph, which gives a good indication of why these constructions had similar performance. Indeed, as we will discuss in a later section, the n entropy bound initially looked like a natural obstacle even for our techniques, though eventually we were able to surpass it. The analogous question for bipartite graphs seemed much harder. Definition 1.2. A bipartite graph on two sets of N vertices is a K-Ramsey Bipartite Graph if it has no K K complete or empty bipartite subgraph. While Erd os' result on the abundance of 2n-Ramsey graphs holds as is for bipartite graphs, until recently the best explicit construction of bipartite Ramsey graphs was 2 n/2 Ramsey , using the Hadamard matrix. This was improved last year, first to o(2 n/2 ) by Pudlak and R odl [16] and then to 2 o(n) by Barak, Kindler, Shaltiel, Sudakov and Wigderson [4]. It is convenient to view such graphs as functions f : ({0, 1} n ) 2 {0, 1}. This then gives exactly the definition of a disperser. Definition 1.3. A function f : ({0, 1} n ) 2 {0, 1} is called a 2-source disperser for entropy k if for any two sets X, Y {0, 1} n with |X| = |Y | = 2 k , we have that the image f (X, Y ) is {0, 1}. This allows for a more formal definition of explicitness: we simply demand that the function f is computable in polynomial time. Most of the constructions mentioned above are explicit in this sense. 1 Our main result (stated informally) significantly improves the bounds in both the bipartite and non-bipartite settings: Theorem 1.4. For every N we construct polynomial time computable bipartite graphs which are 2 n o (1) -Ramsey. A standard transformation of these graphs also yields polynomial time computable ordinary Ramsey Graphs with the same parameters . 1.2 Extractors and Dispersers from independent sources Now we give a brief review of past relevant work (with the goal of putting this paper in proper context) and describe some of the tools from these past works that we will use. We start with the basic definitions of k-sources by Nisan and Zuckerman [15] and of extractors and dispersers for independent sources by Santha and Vazirani [20]. Definition 1.5 ([15], see also [8]). The min-entropy of a distribution X is the maximum k such that for every element x in its support, Pr[X = x] 2 -k . If X is a distribution on strings with min-entropy at least k, we will call X a k-source 2 . To simplify the presentation, in this version of the paper we will assume that we are working with entropy as opposed to min-entropy. Definition 1.6 ([20]). A function f : ({0, 1} n ) c {0, 1} m is a c-source (k, ) extractor if for every family of c independent k-sources X 1 , , X c , the output f (X 1 , , X c ) 1 The Abbot's product based Ramsey-graph construction of [3] and the bipartite Ramsey construction of [16] only satisfy a weaker notion of explicitness. 2 It is no loss of generality to imagine that X is uniformly distributed over some (unknown) set of size 2 k . 672 is a -close 3 to uniformly distributed on m bits. f is a disperser for the same parameters if the output is simply required to have a support of relative size (1 - ). To simplify the presentation, in this version of the paper, we will assume that = 0 for all of our constructions. In this language, Erd os' theorem says that most functions f : ({0, 1} n ) 2 {0,1} are dispersers for entropy 1 + log n (treating f as the characteristic function for the set of edges of the graph). The proof easily extends to show that indeed most such functions are in fact extractors. This naturally challenges us to find explicit functions f that are 2-source extractors. Until one year ago, essentially the only known explicit construction was the Hadamard extractor Had defined by Had (x, y) = x, y ( mod 2). It is an extractor for entropy k > n/2 as observed by Chor and Goldreich [8] and can be extended to give m = (n) output bits as observed by Vazirani [23]. Over 20 years later, a recent breakthrough of Bourgain [5] broke this "1/2 barrier" and can handle 2 sources of entropy .4999n, again with linear output length m = (n). This seemingly minor improvement will be crucial for our work! Theorem 1.7 ([5]). There is a polynomial time computable 2-source extractor f : ({0, 1} n ) 2 {0, 1} m for entropy .4999n and m = (n). No better bounds are known for 2-source extractors. Now we turn our attention to 2-source dispersers. It turned out that progress for building good 2-source dispersers came via progress on extractors for more than 2 sources, all happening in fast pace in the last 2 years. The seminal paper of Bourgain , Katz and Tao [6] proved the so-called "sum-product theorem" in prime fields, a result in arithmetic combinatorics . This result has already found applications in diverse areas of mathematics, including analysis, number theory, group theory and ... extractor theory. Their work implic-itly contained dispersers for c = O(log(n/k)) independent sources of entropy k (with output m = (k)). The use of the "sum-product" theorem was then extended by Barak et al. [3] to give extractors with similar parameters. Note that for linear entropy k = (n), the number of sources needed for extraction c is a constant! Relaxing the independence assumptions via the idea of repeated condensing, allowed the reduction of the number of independent sources to c = 3, for extraction from sources of any linear entropy k = (n), by Barak et al. [4] and independently by Raz [18]. For 2 sources Barak et al. [4] were able to construct dispersers for sources of entropy o(n). To do this, they first showed that if the sources have extra structure (block-source structure, defined below), even extraction is possible from 2 sources. The notion of block-sources, capturing "semi inde-pendence" of parts of the source, was introduced by Chor and Goldreich [8]. It has been fundamental in the development of seeded extractors and as we shall see, is essential for us as well. Definition 1.8 ([8]). A distribution X = X 1 , . . . , X c is a c-block-source of (block) entropy k if every block X i has entropy k even conditioned on fixing the previous blocks X 1 , , X i-1 to arbitrary constants. 3 The error is usually measured in terms of 1 distance or variation distance. This definition allowed Barak et al. [4] to show that their extractor for 4 independent sources, actually performs as well with only 2 independent sources, as long as both are 2-block-sources. Theorem 1.9 ([4]). There exists a polynomial time computable extractor f : ({0, 1} n ) 2 {0, 1} for 2 independent 2-block-sources with entropy o(n). There is no reason to assume that the given sources are block-sources, but it is natural to try and reduce to this case. This approach has been one of the most successful in the extractor literature. Namely try to partition a source X into two blocks X = X 1 , X 2 such that X 1 , X 2 form a 2-block-source. Barak et al. introduced a new technique to do this reduction called the Challenge-Response mechanism, which is crucial for this paper. This method gives a way to "find" how entropy is distributed in a source X, guiding the choice of such a partition. This method succeeds only with small probability, dashing the hope for an extractor, but still yielding a disperser. Theorem 1.10 ([4]). There exists a polynomial time computable 2-source disperser f : ({0, 1} n ) 2 {0, 1} for entropy o(n). Reducing the entropy requirement of the above 2-source disperser, which is what we achieve in this paper, again needed progress on achieving a similar reduction for extractors with more independent sources. A few months ago Rao [?] was able to significantly improve all the above results for c 3 sources. Interestingly, his techniques do not use arithmetic combinatorics, which seemed essential to all the papers above. He improves the results of Barak et al. [3] to give c = O((log n)/(log k))-source extractors for entropy k. Note that now the number c of sources needed for extraction is constant, even when the entropy is as low as n for any constant ! Again, when the input sources are block-sources with sufficiently many blocks, Rao proves that 2 independent sources suffice (though this result does rely on arithmetic combinatorics , in particular, on Bourgain's extractor). Theorem 1.11 ([?]). There is a polynomial time computable extractor f : ({0, 1} n ) 2 {0, 1} m for 2 independent c-block-sources with block entropy k and m = (k), as long as c = O((log n)/(log k)). In this paper (see Theorem 2.7 below) we improve this result to hold even when only one of the 2 sources is a c-block -source. The other source can be an arbitrary source with sufficient entropy. This is a central building block in our construction. This extractor, like Rao's above, critically uses Bourgain's extractor mentioned above. In addition it uses a theorem of Raz [18] allowing seeded extractors to have "weak" seeds, namely instead of being completely random they work as long as the seed has entropy rate > 1/2. MAIN NOTIONS AND NEW RESULTS The main result of this paper is a polynomial time computable disperser for 2 sources of entropy n o(1) , significantly improving both the results of Barak et al. [4] (o(n) entropy). It also improves on Frankl and Wilson [9], who only built Ramsey Graphs and only for entropy ~ O(n). 673 Theorem 2.1 (Main theorem, restated). There exists a polynomial time computable 2-source disperser D : ({0, 1} n ) 2 {0, 1} for entropy n o(1) . The construction of this disperser will involve the construction of an object which in some sense is stronger and in another weaker than a disperser: a subsource somewhere extractor. We first define a related object: a somewhere extractor , which is a function producing several outputs, one of which must be uniform. Again we will ignore many technical issues such as error, min-entropy vs. entropy and more, in definitions and results, which are deferred to the full version of this paper. Definition 2.2. A function f : ({0, 1} n ) 2 ({0, 1} m ) is a 2-source somewhere extractor with outputs, for entropy k, if for every 2 independent k-sources X, Y there exists an i [] such the ith output f(X,Y ) i is a uniformly distributed string of m bits. Here is a simple construction of such a somewhere extractor with as large as poly(n) (and the p in its name will stress the fact that indeed the number of outputs is that large). It will nevertheless be useful to us (though its description in the next sentence may be safely skipped). Define pSE (x, y) i = V(E(x, i), E(y, i)) where E is a "strong" logarithmic seed extractor, and V is the Hadamard/Vazirani 2-source extractor. Using this construction, it is easy to see that: Proposition 2.3. For every n, k there is a polynomial time computable somewhere extractor pSE : ({0, 1} n ) 2 ({0, 1} m ) with = poly(n) outputs, for entropy k, and m = (k). Before we define subsource somewhere extractor, we must first define a subsource. Definition 2.4 (Subsources). Given random variables Z and ^ Z on {0, 1} n we say that ^ Z is a deficiency d subsource of Z and write ^ Z Z if there exists a set A {0,1} n such that (Z|Z A) = ^Z and Pr[Z A] 2 -d . A subsource somewhere extractor guarantees the "some-where extractor" property only on subsources X , Y of the original input distributions X, Y (respectively). It will be extremely important for us to make these subsources as large as possible (i.e. we have to lose as little entropy as possible). Controlling these entropy deficiencies is a major technical complication we have to deal with. However we will be informal with it here, mentioning it only qualitatively when needed. We discuss this issue a little more in Section 6. Definition 2.5. A function f : ({0, 1} n ) 2 ({0, 1} m ) is a 2-source subsource somewhere extractor with outputs for entropy k, if for every 2 independent k-sources X, Y there exists a subsource ^ X of X, a subsource ^ Y of Y and an i [] such the i th output f ( ^ X, ^ Y ) i is a uniformly distributed string of m bits. A central technical result for us is that with this "sub-source" relaxation, we can have much fewer outputs indeed we'll replace poly(n) outputs in our first construction above with n o(1) outputs. Theorem 2.6 (Subsource somewhere extractor). For every > 0 there is a polynomial time computable subsource somewhere extractor SSE : ({0, 1} n ) 2 ({0,1} m ) with = n o(1) outputs, for entropy k = n , with output m = k. We will describe the ideas used for constructing this important object and analyzing it in the next section, where we will also indicate how it is used in the construction of the final disperser. Here we state a central building block, mentioned in the previous section (as an improvement of the work of Rao [?]). We construct an extractor for 2 independent sources one of which is a block-sources with sufficient number of blocks. Theorem 2.7 (Block Source Extractor). There is a polynomial time computable extractor B : ({0, 1} n ) 2 {0, 1} m for 2 independent sources, one of which is a c-block-sources with block entropy k and the other a source of entropy k, with m = (k), and c = O((log n)/(log k)). A simple corollary of this block-source extractor B, is the following weaker (though useful) somewhere block-source extractor SB. A source Z = Z 1 , Z 2 , , Z t is a somewhere c-block-source of block entropy k if for some c indices i 1 < i 2 < < i c the source Z i 1 , Z i 2 , , Z i c is a c-block-source. Collecting the outputs of B on every c-subset of blocks results in that somewhere extractor. Corollary 2.8. There is a polynomial time computable somewhere extractor SB : ({0, 1} n ) 2 ({0, 1} m ) for 2 independent sources, one of which is a somewhere c-block-sources with block entropy k and t blocks total and the other a source of entropy k, with m = (k), c = O((log n)/(log k)), and t c . In both the theorem and corollary above, the values of entropy k we will be interested in are k = n (1) . It follows that a block-source with a constant c = O(1) suffices. THE CHALLENGE-RESPONSE MECHANISM We now describe abstractly a mechanism which will be used in the construction of the disperser as well as the subsource somewhere extractor. Intuitively, this mechanism allows us to identify parts of a source which contain large amounts of entropy. One can hope that using such a mechanism one can partition a given source into blocks in a way which make it a block-source, or alternatively focus on a part of the source which is unusually condensed with entropy two cases which may simplify the extraction problem. The reader may decide, now or in the middle of this section, to skip ahead to the next section which describes the construction of the subsource somewhere extractor SSE, which extensively uses this mechanism. Then this section may seem less abstract, as it will be clearer where this mechanism is used. This mechanism was introduced by Barak et al. [4], and was essential in their 2-source disperser. Its use in this paper is far more involved (in particular it calls itself recursively, a fact which creates many subtleties). However, at a high level, the basic idea behind the mechanism is the same: Let Z be a source and Z a part of Z (Z projected on a subset of the coordinates). We know that Z has entropy k, 674 and want to distinguish two possibilities: Z has no entropy (it is fixed) or it has at least k entropy. Z will get a pass or fail grade, hopefully corresponding to the cases of high or no entropy in Z . Anticipating the use of this mechanism, it is a good idea to think of Z as a "parent" of Z , which wants to check if this "child" has sufficient entropy. Moreover, in the context of the initial 2 sources X, Y we will operate on, think of Z as a part of X, and thus that Y is independent of Z and Z . To execute this "test" we will compute two sets of strings (all of length m, say): the Challenge C = C(Z , Y ) and the Response R = R(Z, Y ). Z fails if C R and passes otherwise. The key to the usefulness of this mechanism is the following lemma, which states that what "should" happen, indeed happens after some restriction of the 2 sources Z and Y . We state it and then explain how the functions C and R are defined to accommodate its proof. Lemma 3.1. Assume Z, Y are sources of entropy k. 1. If Z has entropy k + O(m), then there are subsources ^ Z of Z and ^ Y of Y , such that Pr[ ^ Z passes] = Pr[C( ^ Z , ^ Y ) R ( ^ Z, ^ Y )] 1-n O(1) 2 -m 2. If Z is fixed (namely, has zero entropy), then for some subsources ^ Z of Z and ^ Y of Y , we have Pr[Z fails] = Pr[C( ^ Z , ^ Y ) R( ^Z, ^Y)] = 1 Once we have such a mechanism, we will design our disperser algorithm assuming that the challenge response mechanism correctly identifies parts of the source with high or low levels of entropy. Then in the analysis, we will ensure that our algorithm succeeds in making the right decisions, at least on subsources of the original input sources. Now let us explain how to compute the sets C and R. We will use some of the constructs above with parameters which don't quite fit. The response set R(Z, Y ) = pSE(Z, Y ) is chosen to be the output of the somewhere extractor of Proposition 2.3. The challenge set C(Z , Y ) = SSE(Z , Y ) is chosen to be the output of the subsource somewhere extractor of Theorem 2.6. Why does it work? We explain each of the two claims in the lemma in turn (and after each comment on the important parameters and how they differ from Barak et al. [4]). 1. Z has entropy. We need to show that Z passes the test with high probability. We will point to the output string in C( ^ Z , ^ Y ) which avoids R( ^ Z, ^ Y ) with high probability as follows. In the analysis we will use the union bound on several events, one associated with each (poly(n) many) string in pSE( ^ Z, ^ Y ). We note that by the definition of the response function, if we want to fix a particular element in the response set to a particular value, we can do this by fixing E(Z, i) and E (Y, i). This fixing keeps the restricted sources independent and loses only O(m) entropy. In the subsource of Z guaranteed to exist by Theorem 2.6 we can afford to lose this entropy in Z . Thus we conclude that one of its outputs is uniform. The probability that this output will equal any fixed value is thus 2 -m , completing the argument. We note that we can handle the polynomial output size of pSE, since the uniform string has length m = n (1) (something which could not be done with the technology available to Barak et al. [4]). 2. Z has no entropy. We now need to guarantee that in the chosen subsources (which we choose) ^ Z, ^ Y , all strings in C = C( ^ Z , ^ Y ) are in R( ^ Z, ^ Y ). First notice that as Z is fixed, C is only a function of Y . We set ~ Y to be the subsource of Y that fixes all strings in C = C(Y ) to their most popular values (losing only m entropy from Y ). We take care of including these fixed strings in R(Z, ~ Y ) one at a time, by restricting to subsources assuring that. Let be any m-bit string we want to appear in R(Z, ~ Y ). Recall that R (z, y) = V(E(z, i), E(y, i)). We pick a "good" seed i, and restrict Z, ~ Y to subsources with only O(m) less entropy by fixing E(Z, i) = a and E( ~ Y , i) = b to values (a, b) for which V(a, b) = . This is repeated suc-cessively times, and results in the final subsources ^ Z, ^ Y on which ^ Z fails with probability 1. Note that we keep reducing the entropy of our sources times, which necessitates that this be tiny (here we could not tolerate poly(n), and indeed can guarantee n o(1) , at least on a subsource this is one aspect of how crucial the subsource somewhere extractor SSE is to the construction. We note that initially it seemed like the Challenge-Response mechanism as used in [4] could not be used to handle entropy that is significantly less than n (which is approxi-mately the bound that many of the previous constructions got stuck at). The techniques of [4] involved partitioning the sources into t pieces of length n/t each, with the hope that one of those parts would have a significant amount of entropy, yet there'd be enough entropy left over in the rest of the source (so that the source can be partitioned into a block source). However it is not clear how to do this when the total entropy is less than n. On the one hand we will have to partition our sources into blocks of length significantly more than n (or the adversary could distribute a negligible fraction of entropy in all blocks). On the other hand, if our blocks are so large, a single block could contain all the entropy. Thus it was not clear how to use the challenge response mechanism to find a block source. THE SUBSOURCE SOMEWHERE EXTRACTOR SSE We now explain some of the ideas behind the construction of the subsource somewhere extractor SSE of Theorem 2.6. Consider the source X. We are seeking to find in it a somewhere c-block-source, so that we can use it (together with Y ) in the block-source extractor of Theorem 2.8. Like in previous works in the extractor literature (e.g. [19, 13]) we use a "win-win" analysis which shows that either X is already a somewhere c-block-source, or it has a condensed part which contains a lot of the entropy of the source. In this case we proceed recursively on that part. Continuing this way we eventually reach a source so condensed that it must be a somewhere block source. Note that in [4], the challenge response mechanism was used to find a block source also, but there the entropy was so high that they could afford to use 675 t blocks low high med n bits total t blocks med med low high responded Challenge Challenge responded Challenge Unresponded med med n/t bits total SB SB Outputs Somewhere Block Source! Not Somewhere block source X Random Row < k' 0< low < k'/t k'/c < high < k' k'/t < med < k'/c Figure 1: Analysis of the subsource somewhere extractor. a tree of depth 1. They did not need to recurse or condense the sources. Consider the tree of parts of the source X evolved by such recursion. Each node in the tree corresponds to some interval of bit locations of the source, with the root node corresponding to the entire source. A node is a child of another if its interval is a subinterval of the parent. It can be shown that some node in the tree is "good"; it corresponds to a somewhere c-source, but we don't know which node is good. Since we only want a somewhere extractor, we can apply to each node the somewhere block-source extractor of Corollary 2.8 this will give us a random output in every "good" node of the tree. The usual idea is output all these values (and in seeded extractors, merge them using the ex-ternally given random seed). However, we cannot afford to do that here as there is no external seed and the number of these outputs (the size of the tree) is far too large. Our aim then will be to significantly prune this number of candidates and in fact output only the candidates on one path to a canonical "good" node. First we will give a very informal description of how to do this (Figure 1). Before calling SSE recursively on a subpart of a current part of X, we'll use the "Challenge-Response" mechanism described above to check if "it has entropy". 4 We will recurse only with the first (in left-to-right order) part which passes the "entropy test". Thus note that we will follow a single path on this tree. The algorithm SSE will output only the sets of strings produced by applying the somewhere c-block-extractor SB on the parts visited along this path. Now let us describe the algorithm for SSE. SSE will be initially invoked as SSE(x, y), but will recursively call itself with different inputs z which will always be substrings of x. 4 We note that we ignore the additional complication that SSE will actually use recursion also to compute the challenge in the challenge-response mechanism. Algorithm: SSE (z, y) Let pSE(., .) be the somewhere extractor with a polynomial number of outputs of Proposition 2.3. Let SB be the somewhere block source extractor of Corollary 2.8. Global Parameters: t, the branching factor of the tree. k the original entropy of the sources. Output will be a set of strings. 1. If z is shorter than k, return the empty set, else continue. 2. Partition z into t equal parts z = z 1 , z 2 , . . . , z t . 3. Compute the response set R(z, y) which is the set of strings output by pSE(z, y). 4. For i [t], compute the challenge set C(z i , y), which is the set of outputs of SSE(z i , y). 5. Let h be the smallest index for which the challenge set C (z h , y) is not contained in the response set (set h = t if no such index exists). 6. Output SB(z, y) concatenated with SSE(z h , y). Proving that indeed there are subsources on which SSE will follow a path to a "good" (for these subsources) node, is the heart of the analysis. It is especially complex due to the fact that the recursive call to SSE on subparts of the current part is used to generate the Challenges for the Challenge-Response mechanism. Since SSE works only on a subsources we have to guarantee that restriction to these does not hamper the behavior of SSE in past and future calls to it. Let us turn to the highlights of the analysis, for the proof of Theorem 2.6. Let k be the entropy of the source Z at some place in this recursion. Either one of its blocks Z i has 676 entropy k /c, in which case it is very condensed, since its size is n/t for t c), or it must be that c of its blocks form a c-block source with block entropy k /t (which is sufficient for the extractor B used by SB). In the 2nd case the fact that SB(z, y) is part of the output of of our SSE guarantees that we are somewhere random. If the 2nd case doesn't hold, let Z i be the leftmost condensed block. We want to ensure that (on appropriate subsources) SSE calls itself on that ith subpart. To do so, we fix all Z j for j < i to constants z j . We are now in the position described in the Challenge-Response mechanism section, that (in each of the first i parts) there is either no entropy or lots of entropy. We further restrict to subsources as explained there which make all first i - 1 blocks fail the "entropy test", and the fact that Z i still has lots of entropy after these restrictions (which we need to prove) ensures that indeed SSE will be recursively applied to it. We note that while the procedure SSE can be described recursively , the formal analysis of fixing subsources is actually done globally, to ensure that indeed all entropy requirements are met along the various recursive calls. Let us remark on the choice of the branching parameter t. On the one hand, we'd like to keep it small, as it dominates the number of outputs t c of SB, and thus the total number of outputs (which is t c log t n). For this purpose, any t = n o(1) will do. On the other hand, t should be large enough so that condensing is faster than losing entropy. Here note that if Z is of length n, its child has length n/t, while the entropy shrinks only from k to k /c. A simple calculation shows that if k (log t)/ log c) > n 2 then a c block-source must exist along such a path before the length shrinks to k. Note that for k = n (1) a (large enough) constant t suffices (resulting in only logarithmic number of outputs of SSE). This analysis is depicted pictorially in Figure 1. THE FINAL DISPERSER D Following is a rough description of our disperser D proving Theorem 2.1. The high level structure of D will resemble the structure of SSE - we will recursively split the source X and look for entropy in the parts. However now we must output a single value (rather than a set) which can take both values 0 and 1. This was problematic in SSE, even knowing where the "good" part (containing a c-block-source) was! How can we do so now? We now have at our disposal a much more powerful tool for generating challenges (and thus detecting entropy), namely the subsource somewhere disperser SSE. Note that in constructing SSE we only had essentially the somewhere c-block-source extractor SB to (recursively) generate the challenges, but it depended on a structural property of the block it was applied on. Now SSE does not assume any structure on its input sources except sufficient entropy 5 . Let us now give a high level description of the disperser D . It too will be a recursive procedure. If when processing some part Z of X it "realizes" that a subpart Z i of Z has entropy, but not all the entropy of Z (namely Z i , Z is a 2-block-source) then we will halt and produce the output of D. Intuitively, thinking about the Challenge-Response mechanism described above, the analysis implies that we 5 There is a catch it only works on subsources of them! This will cause us a lot of head ache; we will elaborate on it later. can either pass or fail Z i (on appropriate subsources). But this means that the outcome of this "entropy test" is a 1-bit disperser! To capitalize on this idea, we want to use SSE to identify such a block-source in the recursion tree. As before, we scan the blocks from left to right, and want to distinguish three possibilities. low Z i has low entropy. In this case we proceed to i + 1. medium Z i has "medium" entropy (Z i , Z is a block-source). In which case we halt and produce an output (zero or one). high Z i has essentially all entropy of Z. In this case we recurse on the condensed block Z i . As before, we use the Challenge-Response mechanism (with a twist). We will compute challenges C(Z i , Y ) and responses R (Z, Y ), all strings of length m. The responses are computed exactly as before, using the somewhere extractor pSE. The Challenges are computed using our subsource somewhere extractor SSE. We really have 4 possibilities to distinguish, since when we halt we also need to decide which output bit we give. We will do so by deriving three tests from the above challenges and responses: (C H , R H ), (C M , R M ), (C L , R L ) for high, medium and low respectively, as follows. Let m m H >> m M >> m L be appropriate integers: then in each of the tests above we restrict ourselves to prefixes of all strings of the appropriate lengths only. So every string in C M will be a prefix of length m M of some string in C H . Similarly, every string in R L is the length m L prefix of some string in R H . Now it is immediately clear that if C M is contained in R M , then C L is contained in R L . Thus these tests are monotone, if our sample fails the high test, it will definitely fail all tests. Algorithm: D (z, y) Let pSE(., .) be the somewhere extractor with a polynomial number of outputs of Proposition 2.3. Let SSE(., .) be the subsource somewhere extractor of Theorem 2.6. Global Parameters: t, the branching factor of the tree. k the original entropy of the sources. Local Parameters for recursive level: m L m M m H . Output will be an element of {0, 1}. 1. If z is shorter than k, return 0. 2. Partition z into t equal parts z = z 1 , z 2 , . . . , z t . 3. Compute three response sets R L , R M , R H using pSE(z, y). R j will be the prefixes of length m j of the strings in pSE (z, y). 4. For each i [t], compute three challenge sets C i L , C i M , C i H using SSE(z i , y). C i j will be the prefixes of length m j of the strings in SSE(z i , y). 5. Let h be the smallest index for which the challenge set C L is not contained in the response set R L , if there is no such index, output 0 and halt. 6. If C h H is contained in R H and C h H is contained in R M , output 0 and halt. If C h H is contained in R H but C h H is not contained in R M , output 1 and halt. 677 t blocks t blocks t blocks fail fail fail pass pass pass fail fail fail fail fail fail fail fail fail fail fail fail pass pass fail pass fail fail low low high low low low high low med n bits total n/t bits total X low low Output 0 Output 1 n/t^2 bits total X_3 (X_3)_4 Figure 2: Analysis of the disperser. 7. Output D(z h , y), First note the obvious monotonicity of the tests. If Z i fails one of the tests it will certainly fail for shorter strings. Thus there are only four outcomes to the three tests, written in the order (low, medium, high): (pass, pass, pass), (pass, pass, fail), (pass, fail, fail) and (fail, fail, fail). Conceptually, the algorithm is making the following decisions using the four tests: 1. (fail, fail, fail): Assume Z i has low entropy and proceed to block i + 1. 2. (pass, fail, fail): Assume Z i is medium, halt and output 0. 3. (pass, pass, fail): Assume Z i is medium, halt and output 1. 4. (pass, pass, pass): Assume Z i is high and recurse on Z i . The analysis of this idea (depicted in Figure 2).turns out to be more complex than it seems. There are two reasons for that. Now we briefly explain them and the way to overcome them in the construction and analysis. The first reason is the fact mentioned above, that SSE which generates the challenges, works only on a subsources of the original sources. Restricting to these subsources at some level of the recursion (as required by the analysis of of the test) causes entropy loss which affects both definitions (such as these entropy thresholds for decisions) and correct-ness of SSE in higher levels of recursion. Controlling this entropy loss is achieved by calling SSE recursively with smaller and smaller entropy requirements, which in turn limits the entropy which will be lost by these restrictions. In order not to lose all the entropy for this reason alone, we must work with special parameters of SSE, essentially requiring that at termination it has almost all the entropy it started with. The second reason is the analysis of the test when we are in a medium block. In contrast with the above situation, we cannot consider the value of Z i fixed when we need it to fail on the Medium and Low tests. We need to show that for these two tests (given a pass for High), they come up both (pass, fail) and (fail, fail) each with positive probability. Since the length of Medium challenges and responses is m M , the probability of failure is at least exp(-(m M )) (this follows relatively easily from the fact that the responses are somewhere random). If the Medium test fails so does the Low test, and thus (fail, fail) has a positive probability and our disperser D outputs 0 with positive probability. To bound (pass, fail) we first observe (with a similar reasoning) that the low test fails with probability at least exp(-(m L )). But we want the medium test to pass at the same time. This probability is at least the probability that low fails minus the probability that medium fails. We already have a bound on the latter: it is at most poly(n)exp(-m M ). Here comes our control of the different length into play - we can make the m L sufficiently smaller than m M to yield this difference positive. We conclude that our disperser D outputs 1 with positive probability as well. Finally, we need to take care of termination: we have to ensure that the recurrence always arrives at a medium subpart , but it is easy to chose entropy thresholds for low, medium and high to ensure that this happens. 678 RESILIENCY AND DEFICIENCY In this section we will breifly discuss an issue which arises in our construction that we glossed over in the previous sections . Recall our definition of subsources: Definition 6.1 (Subsources). Given random variables Z and ^ Z on {0, 1} n we say that ^ Z is a deficiency d subsource of Z and write ^ Z Z if there exists a set A {0,1} n such that (Z|A) = ^Z and Pr[Z A] 2 -d . Recall that we were able to guarantee that our algorithms made the right decisions only on subsources of the original source. For example, in the construction of our final disperser , to ensure that our algorithms correctly identify the right high block to recurse on, we were only able to guarantee that there are subsources of the original sources in which our algorithm makes the correct decision with high probability. Then, later in the analysis we had to further restrict the source to even smaller subsources. This leads to complications, since the original event of picking the correct high block, which occurred with high probability, may become an event which does not occur with high probability in the current subsource. To handle these kinds of issues, we will need to be very careful in measuring how small our subsources are. In the formal analysis we introduce the concept of resiliency to deal with this. To give an idea of how this works, here is the actual definition of somewhere subsource extractor that we use in the formal analysis. Definition 6.2 (subsource somewhere extractor). A function SSE : {0, 1} n {0, 1} n ({0, 1} m ) is a subsource somewhere extractor with nrows output rows, entropy threshold k, deficiency def, resiliency res and error if for every (n, k)-sources X, Y there exist a deficiency def subsource X good of X and a deficiency def subsource Y good of Y such that for every deficiency res subsource X of X good and deficiency res subsource Y of Y good , the random variable SSE(X , Y ) is -close to a m somewhere random distribution. It turns out that our subsource somewhere extractor does satisfy this stronger definition. The advantage of this definition is that it says that once we restrict our attention to the good subsources X good , Y good , we have the freedom to further restrict these subsources to smaller subsources, as long as our final subsources do not lose more entropy than the resiliency permits. This issue of managing the resiliency for the various objects that we construct is one of the major technical challenges that we had to overcome in our construction. OPEN PROBLEMS Better Independent Source Extractors A bottleneck to improving our disperser is the block versus general source extractor of Theorem 2.7. A good next step would be to try to build an extractor for one block source (with only a constant number of blocks) and one other independent source which works for polylog-arithmic entropy, or even an extractor for a constant number of sources that works for sub-polynomial entropy . Simple Dispersers While our disperser is polynomial time computable, it is not as explicit as one might have hoped. For instance the Ramsey Graph construction of Frankl-Wilson is extremely simple: For a prime p, let the vertices of the graph be all subsets of [p 3 ] of size p 2 - 1. Two vertices S,T are adjacent if and only if |S T| -1 mod p. It would be nice to find a good disperser that beats the Frankl-Wilson construction, yet is comparable in simplicity. REFERENCES [1] N. Alon. The shannon capacity of a union. Combinatorica, 18, 1998. [2] B. Barak. A simple explicit construction of an n ~ o(log n) -ramsey graph. Technical report, Arxiv, 2006. http://arxiv.org/abs/math.CO/0601651 . [3] B. Barak, R. Impagliazzo, and A. Wigderson. Extracting randomness using few independent sources. In Proceedings of the 45th Annual IEEE Symposium on Foundations of Computer Science, pages 384393, 2004. [4] B. Barak, G. Kindler, R. Shaltiel, B. Sudakov, and A. Wigderson. Simulating independence: New constructions of condensers, Ramsey graphs, dispersers, and extractors. In Proceedings of the 37th Annual ACM Symposium on Theory of Computing, pages 110, 2005. [5] J. Bourgain. More on the sum-product phenomenon in prime fields and its applications. International Journal of Number Theory, 1:132, 2005. [6] J. Bourgain, N. Katz, and T. Tao. A sum-product estimate in finite fields, and applications. Geometric and Functional Analysis, 14:2757, 2004. [7] M. Capalbo, O. Reingold, S. Vadhan, and A. Wigderson. Randomness conductors and constant-degree lossless expanders. In Proceedings of the 34th Annual ACM Symposium on Theory of Computing, pages 659668, 2002. [8] B. Chor and O. Goldreich. Unbiased bits from sources of weak randomness and probabilistic communication complexity. SIAM Journal on Computing, 17(2):230261, 1988. [9] P. Frankl and R. M. Wilson. Intersection theorems with geometric consequences. Combinatorica, 1(4):357368, 1981. [10] P. Gopalan. Constructing ramsey graphs from boolean function representations. In Proceedings of the 21th Annual IEEE Conference on Computational Complexity, 2006. [11] V. Grolmusz. Low rank co-diagonal matrices and ramsey graphs. Electr. J. Comb, 7, 2000. [12] V. Guruswami. Better extractors for better codes? Electronic Colloquium on Computational Complexity (ECCC), (080), 2003. [13] C. J. Lu, O. Reingold, S. Vadhan, and A. Wigderson. Extractors: Optimal up to constant factors. In Proceedings of the 35th Annual ACM Symposium on Theory of Computing, pages 602611, 2003. [14] P. Miltersen, N. Nisan, S. Safra, and A. Wigderson. On data structures and asymmetric communication complexity. Journal of Computer and System Sciences, 57:3749, 1 1998. 679 [15] N. Nisan and D. Zuckerman. More deterministic simulation in logspace. In Proceedings of the 25th Annual ACM Symposium on Theory of Computing, pages 235244, 1993. [16] P. Pudlak and V. Rodl. Pseudorandom sets and explicit constructions of ramsey graphs. Submitted for publication, 2004. [17] A. Rao. Extractors for a constant number of polynomially small min-entropy independent sources. In Proceedings of the 38th Annual ACM Symposium on Theory of Computing, 2006. [18] R. Raz. Extractors with weak random seeds. In Proceedings of the 37th Annual ACM Symposium on Theory of Computing, pages 1120, 2005. [19] O. Reingold, R. Shaltiel, and A. Wigderson. Extracting randomness via repeated condensing. In Proceedings of the 41st Annual IEEE Symposium on Foundations of Computer Science, pages 2231, 2000. [20] M. Santha and U. V. Vazirani. Generating quasi-random sequences from semi-random sources. Journal of Computer and System Sciences, 33:7587, 1986. [21] R. Shaltiel. Recent developments in explicit constructions of extractors. Bulletin of the European Association for Theoretical Computer Science, 77:6795, 2002. [22] A. Ta-Shma and D. Zuckerman. Extractor codes. IEEE Transactions on Information Theory, 50, 2004. [23] U. Vazirani. Towards a strong communication complexity theory or generating quasi-random sequences from two communicating slightly-random sources (extended abstract). In Proceedings of the 17th Annual ACM Symposium on Theory of Computing, pages 366378, 1985. [24] A. Wigderson and D. Zuckerman. Expanders that beat the eigenvalue bound: Explicit construction and applications. Combinatorica, 19(1):125138, 1999. 680

sum-product theorem;distribution;explicit disperser;construction of disperser;Extractors;recursion;subsource somewhere extractor;structure;bipartite graph;extractors;independent sources;extractor;tools;Ramsey Graphs;disperser;polynomial time computable disperser;resiliency;Theorem;Ramsey graphs;block-sources;deficiency;termination;entropy;Ramsey graph;Independent Sources;algorithms;independent source;subsource;Dispersers;randomness extraction

10

A Frequency-based and a Poisson-based Definition of the Probability of Being Informative

This paper reports on theoretical investigations about the assumptions underlying the inverse document frequency (idf ). We show that an intuitive idf -based probability function for the probability of a term being informative assumes disjoint document events. By assuming documents to be independent rather than disjoint, we arrive at a Poisson-based probability of being informative. The framework is useful for understanding and deciding the parameter estimation and combination in probabilistic retrieval models.

INTRODUCTION AND BACKGROUND The inverse document frequency (idf ) is one of the most successful parameters for a relevance-based ranking of retrieved objects. With N being the total number of documents , and n(t) being the number of documents in which term t occurs, the idf is defined as follows: idf(t) := - log n(t) N , 0 <= idf(t) < Ranking based on the sum of the idf -values of the query terms that occur in the retrieved documents works well, this has been shown in numerous applications. Also, it is well known that the combination of a document-specific term Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SIGIR'03, July 28August 1, 2003, Toronto, Canada. Copyright 2003 ACM 1-58113-646-3/03/0007 ... $ 5.00. weight and idf works better than idf alone. This approach is known as tf-idf , where tf(t, d) (0 <= tf(t, d) <= 1) is the so-called term frequency of term t in document d. The idf reflects the discriminating power (informativeness) of a term, whereas the tf reflects the occurrence of a term. The idf alone works better than the tf alone does. An explanation might be the problem of tf with terms that occur in many documents; let us refer to those terms as "noisy" terms. We use the notion of "noisy" terms rather than "fre-quent" terms since frequent terms leaves open whether we refer to the document frequency of a term in a collection or to the so-called term frequency (also referred to as within-document frequency) of a term in a document. We associate "noise" with the document frequency of a term in a collection, and we associate "occurrence" with the within-document frequency of a term. The tf of a noisy term might be high in a document, but noisy terms are not good candidates for representing a document. Therefore, the removal of noisy terms (known as "stopword removal") is essential when applying tf . In a tf-idf approach, the removal of stopwords is conceptually obsolete, if stopwords are just words with a low idf . From a probabilistic point of view, tf is a value with a frequency-based probabilistic interpretation whereas idf has an "informative" rather than a probabilistic interpretation. The missing probabilistic interpretation of idf is a problem in probabilistic retrieval models where we combine uncertain knowledge of different dimensions (e.g.: informativeness of terms, structure of documents, quality of documents, age of documents, etc.) such that a good estimate of the probability of relevance is achieved. An intuitive solution is a normalisation of idf such that we obtain values in the interval [0; 1]. For example, consider a normalisation based on the maximal idf -value. Let T be the set of terms occurring in a collection. P freq (t is informative) := idf(t) maxidf maxidf := max( {idf(t)|t T }), maxidf <= - log(1/N) minidf := min( {idf(t)|t T }), minidf >= 0 minidf maxidf P freq (t is informative) 1.0 This frequency-based probability function covers the interval [0; 1] if the minimal idf is equal to zero, which is the case if we have at least one term that occurs in all documents. Can we interpret P freq , the normalised idf , as the probability that the term is informative? When investigating the probabilistic interpretation of the 227 normalised idf , we made several observations related to disjointness and independence of document events. These observations are reported in section 3. We show in section 3.1 that the frequency-based noise probability n(t) N used in the classic idf -definition can be explained by three assumptions: binary term occurrence, constant document containment and disjointness of document containment events. In section 3.2 we show that by assuming independence of documents, we obtain 1 - e -1 1 - 0.37 as the upper bound of the noise probability of a term. The value e -1 is related to the logarithm and we investigate in section 3.3 the link to information theory. In section 4, we link the results of the previous sections to probability theory. We show the steps from possible worlds to binomial distribution and Poisson distribution. In section 5, we emphasise that the theoretical framework of this paper is applicable for both idf and tf . Finally, in section 6, we base the definition of the probability of being informative on the results of the previous sections and compare frequency-based and Poisson-based definitions. BACKGROUND The relationship between frequencies, probabilities and information theory (entropy) has been the focus of many researchers. In this background section, we focus on work that investigates the application of the Poisson distribution in IR since a main part of the work presented in this paper addresses the underlying assumptions of Poisson. [4] proposes a 2-Poisson model that takes into account the different nature of relevant and non-relevant documents, rare terms (content words) and frequent terms (noisy terms, function words, stopwords). [9] shows experimentally that most of the terms (words) in a collection are distributed according to a low dimension n-Poisson model. [10] uses a 2-Poisson model for including term frequency-based probabilities in the probabilistic retrieval model. The non-linear scaling of the Poisson function showed significant improvement compared to a linear frequency-based probability. The Poisson model was here applied to the term frequency of a term in a document. We will generalise the discussion by pointing out that document frequency and term frequency are dual parameters in the collection space and the document space, respectively. Our discussion of the Poisson distribution focuses on the document frequency in a collection rather than on the term frequency in a document. [7] and [6] address the deviation of idf and Poisson, and apply Poisson mixtures to achieve better Poisson-based estimates . The results proved again experimentally that a one-dimensional Poisson does not work for rare terms, therefore Poisson mixtures and additional parameters are proposed. [3], section 3.3, illustrates and summarises comprehen-sively the relationships between frequencies, probabilities and Poisson. Different definitions of idf are put into context and a notion of "noise" is defined, where noise is viewed as the complement of idf . We use in our paper a different notion of noise: we consider a frequency-based noise that corresponds to the document frequency, and we consider a term noise that is based on the independence of document events. [11], [12], [8] and [1] link frequencies and probability estimation to information theory. [12] establishes a framework in which information retrieval models are formalised based on probabilistic inference. A key component is the use of a space of disjoint events, where the framework mainly uses terms as disjoint events. The probability of being informative defined in our paper can be viewed as the probability of the disjoint terms in the term space of [12]. [8] address entropy and bibliometric distributions. Entropy is maximal if all events are equiprobable and the frequency -based Lotka law (N/i is the number of scientists that have written i publications, where N and are distribution parameters), Zipf and the Pareto distribution are related . The Pareto distribution is the continuous case of the Lotka and Lotka and Zipf show equivalences. The Pareto distribution is used by [2] for term frequency normalisation. The Pareto distribution compares to the Poisson distribution in the sense that Pareto is "fat-tailed", i. e. Pareto assigns larger probabilities to large numbers of events than Poisson distributions do. This makes Pareto interesting since Poisson is felt to be too radical on frequent events. We restrict in this paper to the discussion of Poisson, however , our results show that indeed a smoother distribution than Poisson promises to be a good candidate for improving the estimation of probabilities in information retrieval. [1] establishes a theoretical link between tf-idf and information theory and the theoretical research on the meaning of tf-idf "clarifies the statistical model on which the different measures are commonly based". This motivation matches the motivation of our paper: We investigate theoretically the assumptions of classical idf and Poisson for a better understanding of parameter estimation and combination. FROM DISJOINT TO INDEPENDENT We define and discuss in this section three probabilities: The frequency-based noise probability (definition 1), the total noise probability for disjoint documents (definition 2). and the noise probability for independent documents (definition 3). 3.1 Binary occurrence, constant containment and disjointness of documents We show in this section, that the frequency-based noise probability n(t) N in the idf definition can be explained as a total probability with binary term occurrence, constant document containment and disjointness of document containments . We refer to a probability function as binary if for all events the probability is either 1.0 or 0.0. The occurrence probability P (t|d) is binary, if P (t|d) is equal to 1.0 if t d, and P (t|d) is equal to 0.0, otherwise. P (t|d) is binary : P (t|d) = 1.0 P (t|d) = 0.0 We refer to a probability function as constant if for all events the probability is equal. The document containment probability reflect the chance that a document occurs in a collection. This containment probability is constant if we have no information about the document containment or we ignore that documents differ in containment. Containment could be derived, for example, from the size, quality, age, links, etc. of a document. For a constant containment in a collection with N documents, 1 N is often assumed as the containment probability. We generalise this definition and introduce the constant where 0 N. The containment of a document d depends on the collection c, this is reflected by the notation P (d|c) used for the containment 228 of a document. P (d|c) is constant : d : P (d|c) = N For disjoint documents that cover the whole event space, we set = 1 and obtain d P (d|c) = 1.0. Next, we define the frequency-based noise probability and the total noise probability for disjoint documents. We introduce the event notation t is noisy and t occurs for making the difference between the noise probability P (t is noisy|c) in a collection and the occurrence probability P (t occurs|d) in a document more explicit, thereby keeping in mind that the noise probability corresponds to the occurrence probability of a term in a collection. Definition 1. The frequency-based term noise probability : P freq (t is noisy|c) := n(t) N Definition 2. The total term noise probability for disjoint documents: P dis (t is noisy|c) := d P (t occurs|d) P (d|c) Now, we can formulate a theorem that makes assumptions explicit that explain the classical idf . Theorem 1. IDF assumptions: If the occurrence probability P (t|d) of term t over documents d is binary, and the containment probability P (d|c) of documents d is constant , and document containments are disjoint events, then the noise probability for disjoint documents is equal to the frequency-based noise probability. P dis (t is noisy|c) = P freq (t is noisy|c) Proof. The assumptions are: d : (P (t occurs|d) = 1 P (t occurs|d) = 0) P (d|c) = N d P (d|c) = 1.0 We obtain: P dis (t is noisy|c) = d|td 1 N = n(t) N = P freq (t is noisy|c) The above result is not a surprise but it is a mathematical formulation of assumptions that can be used to explain the classical idf . The assumptions make explicit that the different types of term occurrence in documents (frequency of a term, importance of a term, position of a term, document part where the term occurs, etc.) and the different types of document containment (size, quality, age, etc.) are ignored, and document containments are considered as disjoint events. From the assumptions, we can conclude that idf (frequency-based noise, respectively) is a relatively simple but strict estimate. Still, idf works well. This could be explained by a leverage effect that justifies the binary occurrence and constant containment: The term occurrence for small documents tends to be larger than for large documents, whereas the containment for small documents tends to be smaller than for large documents. From that point of view, idf means that P (t d|c) is constant for all d in which t occurs, and P (t d|c) is zero otherwise. The occurrence and containment can be term specific. For example, set P (t d|c) = 1/N D (c) if t occurs in d, where N D (c) is the number of documents in collection c (we used before just N). We choose a document-dependent occurrence P (t|d) := 1/N T (d), i. e. the occurrence probability is equal to the inverse of N T (d), which is the total number of terms in document d. Next, we choose the containment P (d|c) := N T (d)/N T (c)N T (c)/N D (c) where N T (d)/N T (c) is a document length normalisation (number of terms in document d divided by the number of terms in collection c), and N T (c)/N D (c) is a constant factor of the collection (number of terms in collection c divided by the number of documents in collection c). We obtain P (td|c) = 1/N D (c). In a tf-idf -retrieval function, the tf -component reflects the occurrence probability of a term in a document. This is a further explanation why we can estimate the idf with a simple P (t|d), since the combined tf-idf contains the occurrence probability. The containment probability corresponds to a document normalisation (document length normalisation , pivoted document length) and is normally attached to the tf -component or the tf-idf -product. The disjointness assumption is typical for frequency-based probabilities. From a probability theory point of view, we can consider documents as disjoint events, in order to achieve a sound theoretical model for explaining the classical idf . But does disjointness reflect the real world where the containment of a document appears to be independent of the containment of another document? In the next section, we replace the disjointness assumption by the independence assumption . 3.2 The upper bound of the noise probability for independent documents For independent documents, we compute the probability of a disjunction as usual, namely as the complement of the probability of the conjunction of the negated events: P (d 1 . . . d N ) = 1 - P (d 1 . . . d N ) = 1 d (1 - P (d)) The noise probability can be considered as the conjunction of the term occurrence and the document containment. P (t is noisy|c) := P (t occurs (d 1 . . . d N ) |c) For disjoint documents, this view of the noise probability led to definition 2. For independent documents, we use now the conjunction of negated events. Definition 3. The term noise probability for independent documents: P in (t is noisy|c) := d (1 - P (t occurs|d) P (d|c)) With binary occurrence and a constant containment P (d|c) := /N, we obtain the term noise of a term t that occurs in n(t) documents: P in (t is noisy|c) = 1 - 1 N n(t) 229 For binary occurrence and disjoint documents, the containment probability was 1/N. Now, with independent documents , we can use as a collection parameter that controls the average containment probability. We show through the next theorem that the upper bound of the noise probability depends on . Theorem 2. The upper bound of being noisy: If the occurrence P (t|d) is binary, and the containment P (d|c) is constant, and document containments are independent events, then 1 - e is the upper bound of the noise probability . t : P in (t is noisy|c) < 1 - e Proof . The upper bound of the independent noise probability follows from the limit lim N (1 + x N ) N = e x (see any comprehensive math book, for example, [5], for the convergence equation of the Euler function). With x = -, we obtain: lim N 1 N N = e For the term noise, we have: P in (t is noisy|c) = 1 - 1 N n(t) P in (t is noisy|c) is strictly monotonous: The noise of a term t n is less than the noise of a term t n+1 , where t n occurs in n documents and t n+1 occurs in n + 1 documents. Therefore , a term with n = N has the largest noise probability. For a collection with infinite many documents, the upper bound of the noise probability for terms t N that occur in all documents becomes: lim N P in (t N is noisy) = lim N 1 - 1 N N = 1 - e By applying an independence rather a disjointness assumption , we obtain the probability e -1 that a term is not noisy even if the term does occur in all documents. In the disjoint case, the noise probability is one for a term that occurs in all documents. If we view P (d|c) := /N as the average containment, then is large for a term that occurs mostly in large documents , and is small for a term that occurs mostly in small documents. Thus, the noise of a term t is large if t occurs in n(t) large documents and the noise is smaller if t occurs in small documents. Alternatively, we can assume a constant containment and a term-dependent occurrence. If we assume P (d|c) := 1, then P (t|d) := /N can be interpreted as the average probability that t represents a document. The common assumption is that the average containment or occurrence probability is proportional to n(t). However, here is additional potential: The statistical laws (see [3] on Luhn and Zipf) indicate that the average probability could follow a normal distribution, i. e. small probabilities for small n(t) and large n(t), and larger probabilities for medium n(t). For the monotonous case we investigate here, the noise of a term with n(t) = 1 is equal to 1 - (1 - /N) = /N and the noise of a term with n(t) = N is close to 1 - e . In the next section, we relate the value e to information theory. 3.3 The probability of a maximal informative signal The probability e -1 is special in the sense that a signal with that probability is a signal with maximal information as derived from the entropy definition. Consider the definition of the entropy contribution H(t) of a signal t. H(t) := P (t) - ln P (t) We form the first derivation for computing the optimum. H(t) P (t) = - ln P (t) + -1 P (t) P (t) = -(1 + ln P (t)) For obtaining optima, we use: 0 = -(1 + ln P (t)) The entropy contribution H(t) is maximal for P (t) = e -1 . This result does not depend on the base of the logarithm as we see next: H(t) P (t) = - log b P (t) + -1 P (t) ln b P (t) = 1 ln b + log b P (t) = 1 + ln P (t) ln b We summarise this result in the following theorem: Theorem 3. The probability of a maximal informative signal: The probability P max = e -1 0.37 is the probability of a maximal informative signal. The entropy of a maximal informative signal is H max = e -1 . Proof. The probability and entropy follow from the derivation above. The complement of the maximal noise probability is e and we are looking now for a generalisation of the entropy definition such that e is the probability of a maximal informative signal. We can generalise the entropy definition by computing the integral of + ln P (t), i. e. this derivation is zero for e . We obtain a generalised entropy: -( + ln P (t)) d(P (t)) = P (t) (1 - - ln P (t)) The generalised entropy corresponds for = 1 to the classical entropy. By moving from disjoint to independent documents , we have established a link between the complement of the noise probability of a term that occurs in all documents and information theory. Next, we link independent documents to probability theory. THE LINK TO PROBABILITY THEORY We review for independent documents three concepts of probability theory: possible worlds, binomial distribution and Poisson distribution. 4.1 Possible Worlds Each conjunction of document events (for each document, we consider two document events: the document can be true or false) is associated with a so-called possible world. For example, consider the eight possible worlds for three documents (N = 3). 230 world w conjunction w 7 d 1 d 2 d 3 w 6 d 1 d 2 d 3 w 5 d 1 d 2 d 3 w 4 d 1 d 2 d 3 w 3 d 1 d 2 d 3 w 2 d 1 d 2 d 3 w 1 d 1 d 2 d 3 w 0 d 1 d 2 d 3 With each world w, we associate a probability (w), which is equal to the product of the single probabilities of the document events. world w probability (w) w 7 N 3 1 N 0 w 6 N 2 1 N 1 w 5 N 2 1 N 1 w 4 N 1 1 N 2 w 3 N 2 1 N 1 w 2 N 1 1 N 2 w 1 N 1 1 N 2 w 0 N 0 1 N 3 The sum over the possible worlds in which k documents are true and N -k documents are false is equal to the probability function of the binomial distribution, since the binomial coefficient yields the number of possible worlds in which k documents are true. 4.2 Binomial distribution The binomial probability function yields the probability that k of N events are true where each event is true with the single event probability p. P (k) := binom(N, k, p) := N k p k (1 - p)N -k The single event probability is usually defined as p := /N, i. e. p is inversely proportional to N, the total number of events. With this definition of p, we obtain for an infinite number of documents the following limit for the product of the binomial coefficient and p k : lim N N k p k = = lim N N (N -1) . . . (N -k +1) k! N k = k k! The limit is close to the actual value for k << N. For large k, the actual value is smaller than the limit. The limit of (1 -p)N -k follows from the limit lim N (1+ x N ) N = e x . lim N (1 - p) N-k = lim N 1 N N -k = lim N e 1 N -k = e Again , the limit is close to the actual value for k << N. For large k, the actual value is larger than the limit. 4.3 Poisson distribution For an infinite number of events, the Poisson probability function is the limit of the binomial probability function. lim N binom(N, k, p) = k k! e P (k) = poisson(k, ) := k k! e The probability poisson (0, 1) is equal to e -1 , which is the probability of a maximal informative signal. This shows the relationship of the Poisson distribution and information theory. After seeing the convergence of the binomial distribution, we can choose the Poisson distribution as an approximation of the independent term noise probability. First, we define the Poisson noise probability: Definition 4. The Poisson term noise probability: P poi (t is noisy|c) := e n(t) k=1 k k! For independent documents, the Poisson distribution approximates the probability of the disjunction for large n(t), since the independent term noise probability is equal to the sum over the binomial probabilities where at least one of n(t) document containment events is true. P in (t is noisy|c) = n(t) k=1 n(t) k p k (1 - p)N -k P in (t is noisy|c) P poi (t is noisy|c) We have defined a frequency-based and a Poisson-based probability of being noisy, where the latter is the limit of the independence-based probability of being noisy. Before we present in the final section the usage of the noise probability for defining the probability of being informative, we emphasise in the next section that the results apply to the collection space as well as to the the document space. THE COLLECTION SPACE AND THE DOCUMENT SPACE Consider the dual definitions of retrieval parameters in table 1. We associate a collection space D T with a collection c where D is the set of documents and T is the set of terms in the collection. Let N D := |D| and N T := |T | be the number of documents and terms, respectively. We consider a document as a subset of T and a term as a subset of D. Let n T (d) := |{t|d t}| be the number of terms that occur in the document d, and let n D (t) := |{d|t d}| be the number of documents that contain the term t. In a dual way, we associate a document space L T with a document d where L is the set of locations (also referred to as positions, however, we use the letters L and l and not P and p for avoiding confusion with probabilities) and T is the set of terms in the document. The document dimension in a collection space corresponds to the location (position) dimension in a document space. The definition makes explicit that the classical notion of term frequency of a term in a document (also referred to as the within-document term frequency) actually corresponds to the location frequency of a term in a document. For the 231 space collection document dimensions documents and terms locations and terms document/location frequency n D (t, c): Number of documents in which term t occurs in collection c n L (t, d): Number of locations (positions) at which term t occurs in document d N D (c): Number of documents in collection c N L (d): Number of locations (positions) in document d term frequency n T (d, c): Number of terms that document d contains in collection c n T (l, d): Number of terms that location l contains in document d N T (c): Number of terms in collection c N T (d): Number of terms in document d noise/occurrence P (t|c) (term noise) P (t|d) (term occurrence) containment P (d|c) (document) P (l|d) (location) informativeness - ln P (t|c) - ln P (t|d) conciseness - ln P (d|c) - ln P (l|d) P(informative) ln(P (t|c))/ ln(P (t min , c)) ln(P (t|d))/ ln(P (t min , d)) P(concise) ln(P (d|c))/ ln(P (d min |c)) ln(P (l|d))/ ln(P (l min |d)) Table 1: Retrieval parameters actual term frequency value, it is common to use the maximal occurrence (number of locations; let lf be the location frequency). tf(t, d) := lf(t, d) := P freq (t occurs|d) P freq (t max occurs |d) = n L (t, d) n L (t max , d) A further duality is between informativeness and conciseness (shortness of documents or locations): informativeness is based on occurrence (noise), conciseness is based on containment . We have highlighted in this section the duality between the collection space and the document space. We concentrate in this paper on the probability of a term to be noisy and informative. Those probabilities are defined in the collection space. However, the results regarding the term noise and informativeness apply to their dual counterparts: term occurrence and informativeness in a document. Also, the results can be applied to containment of documents and locations THE PROBABILITY OF BEING INFORMATIVE We showed in the previous sections that the disjointness assumption leads to frequency-based probabilities and that the independence assumption leads to Poisson probabilities. In this section, we formulate a frequency-based definition and a Poisson-based definition of the probability of being informative and then we compare the two definitions. Definition 5. The frequency-based probability of being informative: P freq (t is informative|c) := - ln n(t) N - ln 1 N = - log N n(t) N = 1 - log N n(t) = 1 - ln n(t) ln N We define the Poisson-based probability of being informative analogously to the frequency-based probability of being informative (see definition 5). Definition 6. The Poisson-based probability of being informative: P poi (t is informative|c) := ln e n(t) k=1 k k! - ln(e ) = - ln n(t) k=1 k k! - ln For the sum expression, the following limit holds: lim n(t) n(t) k=1 k k! = e - 1 For >> 1, we can alter the noise and informativeness Poisson by starting the sum from 0, since e >> 1. Then, the minimal Poisson informativeness is poisson(0, ) = e . We obtain a simplified Poisson probability of being informative: P poi (t is informative|c) - ln n(t) k=0 k k! = 1 - ln n(t) k=0 k k! The computation of the Poisson sum requires an optimi-sation for large n(t). The implementation for this paper exploits the nature of the Poisson density: The Poisson density yields only values significantly greater than zero in an interval around . Consider the illustration of the noise and informativeness definitions in figure 1. The probability functions displayed are summarised in figure 2 where the simplified Poisson is used in the noise and informativeness graphs. The frequency-based noise corresponds to the linear solid curve in the noise figure. With an independence assumption, we obtain the curve in the lower triangle of the noise figure. By changing the parameter p := /N of the independence probability , we can lift or lower the independence curve. The noise figure shows the lifting for the value := ln N 9.2. The setting = ln N is special in the sense that the frequency-based and the Poisson-based informativeness have the same denominator, namely ln N, and the Poisson sum converges to . Whether we can draw more conclusions from this setting is an open question. We can conclude, that the lifting is desirable if we know for a collection that terms that occur in relatively few doc-232 0 0.2 0.4 0.6 0.8 1 0 2000 4000 6000 8000 10000 Probability of being noisy n(t): Number of documents with term t frequency independence: 1/N independence: ln(N)/N poisson: 1000 poisson: 2000 poisson: 1000,2000 0 0.2 0.4 0.6 0.8 1 0 2000 4000 6000 8000 10000 Probability of being informative n(t): Number of documents with term t frequency independence: 1/N independence: ln(N)/N poisson: 1000 poisson: 2000 poisson: 1000,2000 Figure 1: Noise and Informativeness Probability function Noise Informativeness Frequency P freq Def n(t)/N ln(n(t)/N)/ ln(1/N) Interval 1/N P freq 1.0 0.0 P freq 1.0 Independence P in Def 1 - (1 - p) n(t) ln(1 - (1 - p) n(t) )/ ln(p) Interval p P in < 1 - e ln (p) P in 1.0 Poisson P poi Def e n(t) k=1 k k! ( - ln n(t) k=1 k k! )/( - ln ) Interval e P poi < 1 - e ( - ln(e - 1))/( - ln ) P poi 1.0 Poisson P poi simplified Def e n(t) k=0 k k! ( - ln n(t) k=0 k k! )/ Interval e P poi < 1.0 0.0 < P poi 1.0 Figure 2: Probability functions uments are no guarantee for finding relevant documents, i. e. we assume that rare terms are still relatively noisy. On the opposite, we could lower the curve when assuming that frequent terms are not too noisy, i. e. they are considered as being still significantly discriminative. The Poisson probabilities approximate the independence probabilities for large n(t); the approximation is better for larger . For n(t) < , the noise is zero whereas for n(t) > the noise is one. This radical behaviour can be smoothened by using a multi-dimensional Poisson distribution. Figure 1 shows a Poisson noise based on a two-dimensional Poisson: poisson(k, 1 , 2 ) := e 1 k 1 k! + (1 - ) e 2 k 2 k! The two dimensional Poisson shows a plateau between 1 = 1000 and 2 = 2000, we used here = 0.5. The idea behind this setting is that terms that occur in less than 1000 documents are considered to be not noisy (i.e. they are informative ), that terms between 1000 and 2000 are half noisy, and that terms with more than 2000 are definitely noisy. For the informativeness, we observe that the radical behaviour of Poisson is preserved. The plateau here is approximately at 1/6, and it is important to realise that this plateau is not obtained with the multi-dimensional Poisson noise using = 0.5. The logarithm of the noise is normalised by the logarithm of a very small number, namely 0.5 e -1000 + 0.5 e -2000 . That is why the informativeness will be only close to one for very little noise, whereas for a bit of noise, informativeness will drop to zero. This effect can be controlled by using small values for such that the noise in the interval [ 1 ; 2 ] is still very little. The setting = e -2000/6 leads to noise values of approximately e -2000/6 in the interval [ 1 ; 2 ], the logarithms lead then to 1/6 for the informativeness. The indepence-based and frequency-based informativeness functions do not differ as much as the noise functions do. However, for the indepence-based probability of being informative , we can control the average informativeness by the definition p := /N whereas the control on the frequency-based is limited as we address next. For the frequency-based idf , the gradient is monotonously decreasing and we obtain for different collections the same distances of idf -values, i. e. the parameter N does not affect the distance. For an illustration, consider the distance between the value idf(t n+1 ) of a term t n+1 that occurs in n+1 documents, and the value idf(t n ) of a term t n that occurs in n documents. idf(t n+1 ) - idf(t n ) = ln n n + 1 The first three values of the distance function are: idf(t 2 ) - idf(t 1 ) = ln(1/(1 + 1)) = 0.69 idf(t 3 ) - idf(t 2 ) = ln(1/(2 + 1)) = 0.41 idf(t 4 ) - idf(t 3 ) = ln(1/(3 + 1)) = 0.29 For the Poisson-based informativeness, the gradient decreases first slowly for small n(t), then rapidly near n(t) and then it grows again slowly for large n(t). In conclusion, we have seen that the Poisson-based definition provides more control and parameter possibilities than 233 the frequency-based definition does. Whereas more control and parameter promises to be positive for the personalisa-tion of retrieval systems, it bears at the same time the danger of just too many parameters. The framework presented in this paper raises the awareness about the probabilistic and information-theoretic meanings of the parameters. The parallel definitions of the frequency-based probability and the Poisson-based probability of being informative made the underlying assumptions explicit. The frequency-based probability can be explained by binary occurrence, constant containment and disjointness of documents. Independence of documents leads to Poisson, where we have to be aware that Poisson approximates the probability of a disjunction for a large number of events, but not for a small number. This theoretical result explains why experimental investigations on Poisson (see [7]) show that a Poisson estimation does work better for frequent (bad, noisy) terms than for rare (good, informative) terms. In addition to the collection-wide parameter setting, the framework presented here allows for document-dependent settings, as explained for the independence probability. This is in particular interesting for heterogeneous and structured collections, since documents are different in nature (size, quality, root document, sub document), and therefore, binary occurrence and constant containment are less appropriate than in relatively homogeneous collections. SUMMARY The definition of the probability of being informative transforms the informative interpretation of the idf into a probabilistic interpretation, and we can use the idf -based probability in probabilistic retrieval approaches. We showed that the classical definition of the noise (document frequency) in the inverse document frequency can be explained by three assumptions: the term within-document occurrence probability is binary, the document containment probability is constant, and the document containment events are disjoint. By explicitly and mathematically formulating the assumptions , we showed that the classical definition of idf does not take into account parameters such as the different nature (size, quality, structure, etc.) of documents in a collection, or the different nature of terms (coverage, importance, position , etc.) in a document. We discussed that the absence of those parameters is compensated by a leverage effect of the within-document term occurrence probability and the document containment probability. By applying an independence rather a disjointness assumption for the document containment, we could establish a link between the noise probability (term occurrence in a collection), information theory and Poisson. From the frequency-based and the Poisson-based probabilities of being noisy, we derived the frequency-based and Poisson-based probabilities of being informative. The frequency-based probability is relatively smooth whereas the Poisson probability is radical in distinguishing between noisy or not noisy, and informative or not informative, respectively. We showed how to smoothen the radical behaviour of Poisson with a multi-dimensional Poisson. The explicit and mathematical formulation of idf - and Poisson-assumptions is the main result of this paper. Also, the paper emphasises the duality of idf and tf , collection space and document space, respectively. Thus, the result applies to term occurrence and document containment in a collection, and it applies to term occurrence and position containment in a document. This theoretical framework is useful for understanding and deciding the parameter estimation and combination in probabilistic retrieval models. The links between indepence-based noise as document frequency, probabilistic interpretation of idf , information theory and Poisson described in this paper may lead to variable probabilistic idf and tf definitions and combinations as required in advanced and personalised information retrieval systems. Acknowledgment: I would like to thank Mounia Lalmas, Gabriella Kazai and Theodora Tsikrika for their comments on the as they said "heavy" pieces. My thanks also go to the meta-reviewer who advised me to improve the presentation to make it less "formidable" and more accessible for those "without a theoretic bent". This work was funded by a research fellowship from Queen Mary University of London. REFERENCES [1] A. Aizawa. An information-theoretic perspective of tf-idf measures. Information Processing and Management, 39:4565, January 2003. [2] G. Amati and C. J. Rijsbergen. Term frequency normalization via Pareto distributions. In 24th BCS-IRSG European Colloquium on IR Research, Glasgow, Scotland, 2002. [3] R. K. Belew. Finding out about. Cambridge University Press, 2000. [4] A. Bookstein and D. Swanson. Probabilistic models for automatic indexing. Journal of the American Society for Information Science, 25:312318, 1974. [5] I. N. Bronstein. Taschenbuch der Mathematik. Harri Deutsch, Thun, Frankfurt am Main, 1987. [6] K. Church and W. Gale. Poisson mixtures. Natural Language Engineering, 1(2):163190, 1995. [7] K. W. Church and W. A. Gale. Inverse document frequency: A measure of deviations from poisson. In Third Workshop on Very Large Corpora, ACL Anthology, 1995. [8] T. Lafouge and C. Michel. Links between information construction and information gain: Entropy and bibliometric distribution. Journal of Information Science, 27(1):3949, 2001. [9] E. Margulis. N-poisson document modelling. In Proceedings of the 15th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 177189, 1992. [10] S. E. Robertson and S. Walker. Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval. In Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 232241, London, et al., 1994. Springer-Verlag. [11] S. Wong and Y. Yao. An information-theoric measure of term specificity. Journal of the American Society for Information Science, 43(1):5461, 1992. [12] S. Wong and Y. Yao. On modeling information retrieval with probabilistic inference. ACM Transactions on Information Systems, 13(1):3868, 1995. 234

inverse document frequency (idf);independent and disjoint documents;computer science;information search;probability theories;Poisson based probability;Term frequency;probabilistic retrieval models;Probability of being informative;Independent documents;Disjoint documents;Normalisation;relevance-based ranking of retrieved objects;information theory;Noise probability;frequency-based term noise probability;Poisson-based probability of being informative;Assumptions;Collection space;Poisson distribution;Probabilistic information retrieval;Document space;document retrieval;entropy;Frequency-based probability;Document frequency;Inverse document frequency;Information theory;independence assumption;inverse document frequency;maximal informative signal

100

High Performance Crawling System

In the present paper, we will describe the design and implementation of a real-time distributed system of Web crawling running on a cluster of machines. The system crawls several thousands of pages every second, includes a high-performance fault manager, is platform independent and is able to adapt transparently to a wide range of configurations without incurring additional hardware expenditure. We will then provide details of the system architecture and describe the technical choices for very high performance crawling. Finally, we will discuss the experimental results obtained, comparing them with other documented systems.

INTRODUCTION With the World Wide Web containing the vast amount of information (several thousands in 1993, 3 billion today) that it does and the fact that it is ever expanding, we need a way to find the right information (multimedia of textual). We need a way to access the information on specific subjects that we require. To solve the problems above several programs and algorithms were designed that index the web, these various designs are known as search engines, spiders, crawlers, worms or knowledge robots graph in its simplest terms. The pages are the nodes on the graph and the links are the arcs on the graph. What makes this so difficult is the vast amount of data that we have to handle, and then we must also take into account the fact that the World Wide Web is constantly growing and the fact that people are constantly updating the content of their web pages. Any High performance crawling system should offer at least the following two features. Firstly, it needs to be equipped with an intelligent navigation strategy, i.e. enabling it to make decisions regarding the choice of subsequent actions to be taken (pages to be downloaded etc). Secondly, its supporting hardware and software architecture should be optimized to crawl large quantities of documents per unit of time (generally per second). To this we may add fault tolerance (machine crash, network failure etc.) and considerations of Web server resources. Recently we have seen a small interest in these two field. Studies on the first point include crawling strategies for important pages [9, 17], topic-specific document downloading [5, 6, 18, 10], page recrawling to optimize overall refresh frequency of a Web archive [8, 7] or scheduling the downloading activity according to time [22]. However, little research has been devoted to the second point, being very difficult to implement [20, 13]. We will focus on this latter point in the rest of this paper. Indeed, only a few crawlers are equipped with an optimized scalable crawling system, yet details of their internal workings often remain obscure (the majority being proprietary solutions). The only system to have been given a fairly in-depth description in existing literature is Mercator by Heydon and Najork of DEC/Compaq [13] used in the AltaVista search engine (some details also exist on the first version of the Google [3] and Internet Archive [4] robots). Most recent studies on crawling strategy fail to deal with these features, contenting themselves with the solution of minor issues such as the calculation of the number of pages to be downloaded in order to maximize/minimize some functional objective. This may be acceptable in the case of small applications, but for real time 1 applications the system must deal with a much larger number of constraints. We should also point out that little academic research is concerned with high performance search engines, as compared with their commercial counterparts (with the exception of the WebBase project [14] at Stanford). In the present paper, we will describe a very high availability, optimized and distributed crawling system. We will use the system on what is known as breadth-first crawling, though this may be easily adapted to other navigation strategies. We will first focus on input/output, on management of network traffic and robustness when changing scale. We will also discuss download policies in 1 "Soft" real time 299 Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MIR'04, October 1516, 2004, New York, New York, USA. Copyright 2004 ACM 1-58113-940-3/04/0010...$5.00. terms of speed regulation, fault management by supervisors and the introduction/suppression of machine nodes without system restart during a crawl. Our system was designed within the experimental framework of the D ep^ ot L egal du Web Fran cais (French Web Legal Deposit). This consists of archiving only multimedia documents in French available on line, indexing them and providing ways for these archives to be consulted. Legal deposit requires a real crawling strategy in order to ensure site continuity over time. The notion of registration is closely linked to that of archiving, which requires a suitable strategy to be useful. In the course of our discussion, we will therefore analyze the implication and impact of this experimentation for system construction. STATE OF THE ART In order to set our work in this field in context, listed below are definitions of services that should be considered the minimum requirements for any large-scale crawling system. Flexibility: as mentioned above, with some minor adjustments our system should be suitable for various scenarios. However, it is important to remember that crawling is established within a specific framework: namely, Web legal deposit. High Performance: the system needs to be scalable with a minimum of one thousand pages/second and extending up to millions of pages for each run on low cost hardware. Note that here, the quality and efficiency of disk access are crucial to maintaining high performance. Fault Tolerance: this may cover various aspects. As the system interacts with several servers at once, specific problems emerge. First, it should at least be able to process invalid HTML code, deal with unexpected Web server behavior, and select good communication protocols etc. The goal here is to avoid this type of problem and, by force of circumstance, to be able to ignore such problems completely. Second, crawling processes may take days or weeks, and it is imperative that the system can handle failure, stopped processes or interruptions in network services, keeping data loss to a minimum. Finally, the system should be persistent, which means periodically switching large data structures from memory to the disk (e.g. restart after failure). Maintainability and Configurability: an appropriate interface is necessary for monitoring the crawling process, including download speed, statistics on the pages and amounts of data stored. In online mode, the administrator may adjust the speed of a given crawler, add or delete processes, stop the system, add or delete system nodes and supply the black list of domains not to be visited, etc. 2.2 General Crawling Strategies There are many highly accomplished techniques in terms of Web crawling strategy. We will describe the most relevant of these here. Breadth-first Crawling: in order to build a wide Web archive like that of the Internet Archive [15], a crawl is carried out from a set of Web pages (initial URLs or seeds). A breadth-first exploration is launched by following hypertext links leading to those pages directly connected with this initial set. In fact, Web sites are not really browsed breadth-first and various restrictions may apply, e.g. limiting crawling processes to within a site, or downloading the pages deemed most interesting first 2 Repetitive Crawling: once pages have been crawled, some systems require the process to be repeated periodically so that indexes are kept updated. In the most basic case, this may be achieved by launching a second crawl in parallel. A variety of heuristics exist to overcome this problem: for example, by frequently relaunching the crawling process of pages, sites or domains considered important to the detriment of others. A good crawling strategy is crucial for maintaining a constantly updated index list. Recent studies by Cho and Garcia-Molina [8, 7] have focused on optimizing the update frequency of crawls by using the history of changes recorded on each site. Targeted Crawling: more specialized search engines use crawling process heuristics in order to target a certain type of page, e.g. pages on a specific topic or in a particular language, images, mp3 files or scientific papers. In addition to these heuristics, more generic approaches have been suggested. They are based on the analysis of the structures of hypertext links [6, 5] and techniques of learning [9, 18]: the objective here being to retrieve the greatest number of pages relating to a particular subject by using the minimum bandwidth. Most of the studies cited in this category do not use high performance crawlers, yet succeed in producing acceptable results. Random Walks and Sampling: some studies have focused on the effect of random walks on Web graphs or modified versions of these graphs via sampling in order to estimate the size of documents on line [1, 12, 11]. Deep Web Crawling: a lot of data accessible via the Web are currently contained in databases and may only be downloaded through the medium of appropriate requests or forms. Recently, this often-neglected but fascinating problem has been the focus of new interest. The Deep Web is the name given to the Web containing this category of data [9]. Lastly, we should point out the acknowledged differences that exist between these scenarios. For example, a breadth-first search needs to keep track of all pages already crawled. An analysis of links should use structures of additional data to represent the graph of the sites in question, and a system of classifiers in order to assess the pages' relevancy [6, 5]. However, some tasks are common to all scenarios, such as 2 See [9] for the heuristics that tend to find the most important pages first and [17] for experimental results proving that breadth-first crawling allows the swift retrieval of pages with a high PageRank. 300 respecting robot exclusion files (robots.txt), crawling speed, resolution of domain names . . . In the early 1990s, several companies claimed that their search engines were able to provide complete Web coverage. It is now clear that only partial coverage is possible at present. Lawrence and Giles [16] carried out two experiments in order to measure coverage performance of data established by crawlers and of their updates. They adopted an approach known as overlap analysis to estimate the size of the Web that may be indexed (See also Bharat and Broder 1998 on the same subject). Let W be the total set of Web pages and W a W and W b W the pages downloaded by two different crawlers a and b. What is the size of W a and W b as compared with W ? Let us assume that uniform samples of Web pages may be taken and their membership of both sets tested. Let P (W a ) and P (W b ) be the probability that a page is downloaded by a or b respectively. We know that: P (W a W b |W b ) = W a W b |W b | (1) Now, if these two crawling processes are assumed to be independent, the left side of equation 1may be reduced to P (W a ), that is data coverage by crawler a. This may be easily obtained by the intersection size of the two crawling processes. However, an exact calculation of this quantity is only possible if we do not really know the documents crawled. Lawrence and Giles used a set of controlled data of 575 requests to provide page samples and count the number of times that the two crawlers retrieved the same pages. By taking the hypothesis that the result P (W a ) is correct, we may estimate the size of the Web as |W a |/P (W a ). This approach has shown that the Web contained at least 320 million pages in 1997 and that only 60% was covered by the six major search engines of that time. It is also interesting to note that a single search engine would have covered only 1/3 of the Web. As this approach is based on observation, it may reflect a visible Web estimation, excluding for instance pages behind forms, databases etc. More recent experiments assert that the Web contains several billion pages. 2.2.1 Selective Crawling As demonstrated above, a single crawler cannot archive the whole Web. The fact is that the time required to carry out the complete crawling process is very long, and impossible given the technology currently available. Furthermore, crawling and indexing very large amounts of data implies great problems of scalability, and consequently entails not inconsiderable costs of hardware and maintenance. For maximum optimization, a crawling system should be able to recognize relevant sites and pages, and restrict itself to downloading within a limited time. A document or Web page's relevancy may be officially recognized in various ways. The idea of selective crawling may be introduced intuitively by associating each URL u with a score calculation function s () respecting relevancy criterion and parameters . In the most basic case, we may assume a Boolean relevancy function, i.e. s(u) = 1 if the document designated by u is relevant and s(u) = 0 if not. More generally, we may think of s(d) as a function with real values, such as a conditional probability that a document belongs to a certain category according to its content. In all cases, we should point out that the score calculation function depends only on the URL and and not on the time or state of the crawler. A general approach for the construction of a selective crawler consists of changing the URL insertion and extraction policy in the queue Q of the crawler. Let us assume that the URLs are sorted in the order corresponding to the value retrieved by s(u). In this case, we obtain the best-first strategy (see [19]) which consists of downloading URLs with the best scores first). If s(u) provides a good relevancy model, we may hope that the search process will be guided towards the best areas of the Web. Various studies have been carried out in this direction: for example, limiting the search depth in a site by specifying that pages are no longer relevant after a certain depth. This amounts to the following equation: s (depth) (u) = 1, if |root(u) u| < 0, else (2) where root(u) is the root of the site containing u. The interest of this approach lies in the fact that maximizing the search breadth may make it easier for the end-user to retrieve the information. Nevertheless, pages that are too deep may be accessed by the user, even if the robot fails to take them into account. A second possibility is the estimation of a page's popularity . One method of calculating a document's relevancy would relate to the number of backlinks. s (backlinks) (u) = 1, if indegree(u) > 0, else (3) where is a threshold. It is clear that s (backlinks) (u) may only be calculated if we have a complete site graph (site already downloaded beforehand). In practice, we make take an approximate value and update it incrementally during the crawling process. A derivative of this technique is used in Google's famous PageRank calculation. OUR APPROACH THE DOMINOS SYSTEM As mentioned above, we have divided the system into two parts: workers and supervisors. All of these processes may be run on various operating systems (Windows, MacOS X, Linux, FreeBSD) and may be replicated if need be. The workers are responsible for processing the URL flow coming from their supervisors and for executing crawling process tasks in the strict sense. They also handle the resolution of domain names by means of their integrated DNS resolver, and adjust download speed in accordance with node policy. A worker is a light process in the Erlang sense, acting as a fault tolerant and highly available HTTP client. The process-handling mode in Erlang makes it possible to create several thousands of workers in parallel. In our system, communication takes place mainly by sending asynchronous messages as described in the specifications for Erlang language. The type of message varies according to need: character string for short messages and binary format for long messages (large data structures or files). Disk access is reduced to a minimum as far as possible and structures are stored in the real-time Mnesia 3 database that forms 3 http://www.erlang.org/doc/r9c/lib/mnesia-4 .1.4/doc/html/ 301 a standard part of the Erlang development kit. Mnesia's features give it a high level of homogeneity during the base's access, replication and deployment. It is supported by two table management modules ETS and DETS. ETS allows tables of values to be managed by random access memory, while DETS provides a persistent form of management on the disk. Mnesia's distribution faculty provides an efficient access solution for distributed data. When a worker moves from one node to another (code migration), it no longer need be concerned with the location of the base or data. It simply has to read and write the information transparently. 1 loop(InternalState) -> % Supervisor main 2 % loop 3 receive {From,{migrate,Worker,Src,Dest}} -> 4 % Migrate the Worker process from 5 % Src node to Dest node 6 spawn(supervisor,migrate, 7 [Worker,Src,Dest]), 8 % Infinite loop 9 loop(InternalState); 10 11 {From,{replace,OldPid,NewPid,State}} -> 12 % Add the new worker to 13 % the supervisor state storage 14 NewInternalState = 15 replace(OldPid,NewPid,InternalState), 16 % Infinite loop 17 loop(NewInternalState); 18 ... 19 end. 20 21 migrate(Pid,Src,Dest) -> % Migration 22 % process 23 receive 24 Pid ! {self(), stop}, 25 receive 26 {Pid,{stopped,LastState}} -> 27 NewPid = spawn{Dest,worker,proc, 28 [LastState]}, 29 self() ! {self(), {replace,Pid, 30 NewPid,LastState}}; 31 {Pid,Error} -> ... 32 end. Listing 1: Process Migration Code 1describes the migration of a worker process from one node Src to another Dest. 4 The supervisor receives the migration order for process P id (line 4). The migration action is not blocking and is performed in a different Erlang process (line 7). The supervisor stops the worker with the identifier P id (line 25) and awaits the operation result (line 26). It then creates a remote worker in the node Dest with the latest state of the stopped worker (line 28) and updates its internal state (lines 30 and 12). 3.1 Dominos Process The Dominos system is different from all the other crawling systems cited above. Like these, the Dominos offering is on distributed architecture, but with the difference of being totally dynamic. The system's dynamic nature allows its architecture to be changed as required. If, for instance, one of the cluster's nodes requires particular maintenance, all of the processes on it will migrate from this node to another. When servicing is over, the processes revert automatically 4 The character % indicates the beginning of a comment in Erlang. to their original node. Crawl processes may change pool so as to reinforce one another if necessary. The addition or deletion of a node in the cluster is completely transparent in its execution. Indeed, each new node is created containing a completely blank system. The first action to be undertaken is to search for the generic server in order to obtain the parameters of the part of the system that it is to belong to. These parameters correspond to a limited view of the whole system. This enables Dominos to be deployed more easily, the number of messages exchanged between processes to be reduced and allows better management of exceptions. Once the generic server has been identified, binaries are sent to it and its identity is communicated to the other nodes concerned. Dominos Generic Server (GenServer): Erlang process responsible for managing the process identifiers on the whole cluster. To ensure easy deployment of Dominos, it was essential to mask the denominations of the process identifiers. Otherwise, a minor change in the names of machines or their IP would have required complete reorganization of the system. GenServer stores globally the identifiers of all processes existing at a given time. Dominos RPC Concurrent (cRPC): as its name suggests , this process is responsible for delegating the execution of certain remote functions to other processes . Unlike conventional RPCs where it is necessary to know the node and the object providing these functions (services), our RPCC completely masks the information. One need only call the function, with no concern for where it is located in the cluster or for the name of the process offering this function. Moreover, each RPCC process is concurrent, and therefore manages all its service requests in parallel. The results of remote functions are governed by two modes: blocking or non-blocking. The calling process may therefore await the reply of the remote function or continue its execution. In the latter case, the reply is sent to its mailbox. For example, no worker knows the process identifier of its own supervisor. In order to identify it, a worker sends a message to the process called supervisor. The RPCC deals with the message and searches the whole cluster for a supervisor process identifier, starting with the local node. The address is therefore resolved without additional network overhead, except where the supervisor does not exist locally. Dominos Distributed Database (DDB): Erlang process responsible for Mnesia real-time database management . It handles the updating of crawled information, crawling progress and the assignment of URLs to be downloaded to workers. It is also responsible for replicating the base onto the nodes concerned and for the persistency of data on disk. Dominos Nodes: a node is the physical representation of a machine connected (or disconnected as the case may be) to the cluster. This connection is considered in the most basic sense of the term, namely a simple plugging-in (or unplugging) of the network outlet. Each node clearly reflects the dynamic character of the Dominos system. 302 Dominos Group Manager: Erlang process responsible for controlling the smooth running of its child processes (supervisor and workers). Dominos Master-Supervisor Processes: each group manager has a single master process dealing with the management of crawling states of progress. It therefore controls all the slave processes (workers) contained within it. Dominos Slave-Worker Processes: workers are the lowest-level elements in the crawling process. This is the very heart of the Web client wrapping the libCURL. With Dominos architecture being completely dynamic and distributed, we may however note the hierarchical character of processes within a Dominos node. This is the only way to ensure very high fault tolerance. A group manager that fails is regenerated by the node on which it depends. A master process (supervisor) that fails is regenerated by its group manager. Finally, a worker is regenerated by its supervisor. As for the node itself, it is controlled by the Dominos kernel (generally on another remote machine). The following code describes the regeneration of a worker process in case of failure. 1 % Activate error handling 2 process_flag(trap_exit, true ), 3 ... 4 loop(InternalState) -> % Supervisor main loop 5 receive 6 {From,{job, Name ,finish}, State} -> 7 % Informe the GenServer that the download is ok 8 ?ServerGen ! {job, Name ,finish}, 9 10 % Save the new worker state 11 NewInternalState=save_state(From,State,InternalState), 12 13 % Infinite loop 14 loop(NewInternalState); 15 ... 16 {From,Error} -> % Worker crash 17 % Get the last operational state before the crash 18 WorkerState = last_state(From,InternalState), 19 20 % Free all allocated resources 21 free_resources(From,InternalState), 22 23 % Create a new worker with the last operational 24 % state of the crashed worker 25 Pid = spawn(worker,proc,[WorkerState]), 26 27 % Add the new worker to the supervisor state 28 % storage 29 NewInternalState =replace(From,Pid,InternalState), 30 31 % Infinite loop 32 loop(NewInternalState); 33 end. Listing 2: Regeneration of a Worker Process in Case of Failure This represents the part of the main loop of the supervisor process dealing with the management of the failure of a worker. As soon as a worker error is received (line 19), the supervisor retrieves the last operational state of the worker that has stopped (line 22), releases all of its allocated resources (line 26) and recreates a new worker process with the operational state of the stopped process (line 31). The supervisor continually turns in loop while awaiting new messages (line 40). The loop function call (lines 17 and 40) is tail recursive, thereby guaranteeing that the supervision process will grow in a constant memory space. 3.2 DNS Resolution Before contacting a Web server, the worker process needs to convert the Domain Name Server (DNS) into a valid IP address. Whereas other systems (Mercator, Internet Archive) are forced to set up DNS resolvers each time a new link is identified, this is not necessary with Dominos. Indeed, in the framework of French Web legal deposit, the sites to be archived have been identified beforehand, thus requiring only one DNS resolution per domain name. This considerably increases crawl speed. The sites concerned include all online newspapers , such as LeMonde (http://www.lemonde.fr/ ), LeFigaro (http://www.lefigaro.fr/ ) . . . , online television/radio such as TF1(http://www.tf1.fr/ ), M6 (http://www.m6.fr/ ) . . . DETAILS OF IMPLEMENTATION The workers are the medium responsible for physically crawling on-line contents. They provide a specialized wrapper around the libCURL 5 library that represents the heart of the HTTP client. Each worker is interfaced to libCURL by a C driver (shared library). As the system seeks maximum network accessibility (communication protocol support), libCURL appeared to be the most judicious choice when compared with other available libraries. 6 . The protocols supported include: FTP, FTPS, HTTP, HTTPS, LDAP, Certifications, Proxies, Tunneling etc. Erlang's portability was a further factor favoring the choice of libCURL. Indeed, libCURL is available for various architectures: Solaris, BSD, Linux, HPUX, IRIX, AIX, Windows, Mac OS X, OpenVMS etc. Furthermore, it is fast, thread-safe and IPv6 compatible. This choice also opens up a wide variety of functions. Redirections are accounted for and powerful filtering is possible according to the type of content downloaded, headers, and size (partial storage on RAM or disk depending on the document's size). 4.2 Document Fingerprint For each download, the worker extracts the hypertext links included in the HTML documents and initiates a fingerprint (signature operation). A fast fingerprint (HAVAL on 256 bits) is calculated for the document's content itself so as to differentiate those with similar contents (e.g. mirror sites). This technique is not new and has already been used in Mercator[13]. It allows redundancies to be eliminated in the archive. 4.3 URL Extraction and Normalization Unlike other systems that use libraries of regular expressions such as PCRE 7 for URL extraction, we have opted 5 Available at http://curl.haxx.se/libcurl/ 6 See http://curl.haxx.se/libcurl/competitors.html 7 Available at http://www.pcre.org/ 303 for the Flex tool that definitely generates a faster parser. Flex was compiled using a 256Kb buffer in which all table compression options were activated during parsing "-8 -f Cf -Ca -Cr -i". Our current parser analyzes around 3,000 pages/second for a single worker for an average 49Kb per page. According to [20], a URL extraction speed of 300 pages/second may generate a list of more than 2,000 URLs on average. A naive representation of structures in the memory may soon saturate the system. Various solutions have been proposed to alleviate this problem. The Internet Archive [4] crawler uses Bloom filters in random access memory. This makes it possible to have a compact representation of links retrieved, but also generates errors (false-positive), i.e. certain pages are never downloaded as they create collisions with other pages in the Bloom filter. Compression without loss may reduce the size of URLs to below 10Kb [2, 21], but this remains insufficient in the case of large-scale crawls. A more ingenious approach is to use persistent structures on disk coupled with a cache as in Mercator [13]. 4.4 URL Caching In order to speed up processing, we have developed a scalable cache structure for the research and storage of URLs already archived. Figure 1describes how such a cache works: Links Local Cache - Worker Rejected Links 0 1 2 255 JudyL-Array URL CRC URL #URL key value JudySL-Array Figure 1: Scalable Cache The cache is available at the level of each worker. It acts as a filter on URLs found and blocks those already encountered. The cache needs to be scalable to be able to deal with increasing loads. Rapid implementation using a non-reversible hash function such as HAVAL, TIGER, SHA1 , GOST, MD5, RIPEMD . . . would be fatal to the system's scalability. Although these functions ensure some degree of uniqueness in fingerprint constructionthey are too slow to be acceptable in these constructions. We cannot allow latency as far as lookup or URL insertion in the cache is concerned, if the cache is apt to exceed a certain size (over 10 7 key-value on average). This is why we have focused on the construction of a generic cache that allows key-value insertion and lookup in a scalable manner. The Judy-Array API 8 enabled us to achieve this objective. Without going into detail about Judy-Array (see their site for more information), our cache is a coherent coupling between a JudyL-Array and N JudySL-Array. The JudyL-Array represents a hash table of N = 2 8 or N = 2 16 buckets able to fit into the internal cache of the CPU. It is used to store "key-numeric value" pairs where the key represents a CRC of the 8 Judy Array at the address: http://judy.sourceforge.net/ URL and whose value is a pointer to a JudySL-Array. The second, JudySL-Array, is a "key-compressed character string value" type of hash, in which the key represents the URL identifier and whose value is the number of times that the URL has been viewed. This cache construction is completely scalable and makes it possible to have sub-linear response rates, or linear in the worst-case scenario (see Judy-Array at for an in-depth analysis of their performance). In the section on experimentation (section 5) we will see the results of this type of construction. 4.5 Limiting Disk Access Our aim here is to eliminate random disk access completely . One simple idea used in [20] is periodically to switch structures requiring much memory over onto disk. For example, random access memory can be used to keep only those URLs found most recently or most frequently, in order to speed up comparisons. This requires no additional development and is what we have decided to use. The persistency of data on disk depends on the size of data in DS memory, and their DA age. The data in the memory are distributed transparently via Mnesia, specially designed for this kind of situation. Data may be duplicated ( {ram copies, [Nodes]}, {disc copies, [Nodes]}) or fragmented ( {frag properties, .....}) on the nodes in question. According to [20], there are on average 8 non-duplicated hypertext links per page downloaded. This means that the number of pages retrieved and not yet archived is considerably increased. After archiving 20 million pages, over 100 million URLs would still be waiting. This has various repercussions, as newly-discovered URLs will be crawled only several days, or even weeks, later. Given this speed, the base's data refresh ability is directly affected. 4.6 High Availability In order to apprehend the very notion of High Availability, we first need to tackle the differences that exist between a system's reliability and its availability. Reliability is an attribute that makes it possible to measure service continuity when no failure occurs. Manufacturers generally provide a statistical estimation of this value for this equipment: we may use the term MTBF (Mean Time Between Failure). A strong MTBF provides a valuable indication of a component's ability to avoid overly frequent failure. In the case of a complex system (that can be broken down into hardware or software parts), we talk about MTTF (Mean Time To Failure). This denotes the average time elapsed until service stops as the result of failure in a component or software. The attribute of availability is more difficult to calculate as it includes a system's ability to react correctly in case of failure in order to restart service as quickly as possible. It is therefore necessary to quantify the time interval during which service is unavailable before being re-established: the acronym MTTR (Mean Time To Repair) is used to represent this value. The formula used to calculate the rate of a system's availability is as follows: availability = M T T F M T T F + M T T R (4) 304 A system that looks to have a high level of availability should have either a strong MTTF, or a weak MTTR. Another more practical approach consists in measuring the time period during which service is down in order to evaluate the level of availability. This is the method most frequently adopted, even if it fails to take account of the frequency of failure, focusing rather on its duration. Calculation is usually based on a calendar year. The higher the percentage of service availability, the nearer it comes to High Availability. It is fairly easy to qualify the level of High Availability of a service from the cumulated downtime, by using the normalized principle of "9's" (below 3 nine, we are no longer talking about High Availability, but merely availability). In order to provide an estimation of Dominos' High Availability, we carried out performance tests by fault injection. It is clear that a more accurate way of measuring this criterion would be to let the system run for a whole year as explained above. However, time constraints led us to adopt this solution. Our injector consists in placing pieces of false code in each part of the system and then measuring the time required for the system to make the service available. Once again, Erlang has proved to be an excellent choice for the setting up of these regression tests. The table below shows the average time required by Dominos to respond to these cases of service unavailability. Table 1clearly shows Dominos' High Availability. We Service Error MTTR (microsec) GenServer 10 3 bad match 320 cRPC 10 3 bad match 70 DDB 10 7 tuples 9 10 6 Node 10 3 bad match 250 Supervisor 10 3 bad match 60 Worker 10 3 bad match 115 Table 1: MTTR Dominos see that for 10 3 matches of error, the system resumes service virtually instantaneously. The DB was tested on 10 7 tuples in random access memory and resumed service after approximately 9 seconds. This corresponds to an excellent MTTR, given that the injections were made on a PIII-966Mhz with 512Mb of RAM. From these results, we may label our system as being High Availability, as opposed to other architectures that consider High Availability only in the sense of failure not affecting other components of the system, but in which service restart of a component unfortunately requires manual intervention every time. EXPERIMENTATION This section describes Dominos' experimental results tested on 5 DELL machines: nico: Intel Pentium 4 - 1.6 Ghz, 256 Mb RAM. Crawl node (supervisor, workers). Activates a local cRPC. zico: Intel Pentium 4 - 1.6 Ghz, 256 Mb RAM. Crawl node (supervisor, workers). Activates a local cRPC. chopin: Intel Pentium 3 - 966 Mhz, 512 Mb RAM. Main node loaded on ServerGen and DB. Also handles crawling (supervisor, workers). Activates a local cRPC. gao: Intel Pentium 3 - 500 Mhz, 256 Mb RAM. Node for DB fragmentation. Activates a local cRPC. margo: Intel Pentium 2 - 333 Mhz, 256 Mb RAM. Node for DB fragmentation. Activates a local cRPC. Machines chopin, gao and margo are not dedicated solely to crawling and are used as everyday workstations. Disk size is not taken into account as no data were actually stored during these tests. Everything was therefore carried out using random access memory with a network of 100 Mb/second. Dominos performed 25,116,487 HTTP requests after 9 hours of crawling with an average of 816 documents/second for 49Kb per document. Three nodes (nico, zico and chopin) were used in crawling, each having 400 workers. We restricted ourselves to a total of 1,200 workers, due to problems generated by Dominos at intranet level. The firewall set up to filter access is considerably detrimental to performance because of its inability to keep up with the load imposed by Dominos. Third-party tests have shown that peaks of only 4,000 HTTP requests/second cause the immediate collapse of the firewall. The firewall is not the only limiting factor, as the same tests have shown the incapacity of Web servers such as Apache2, Caudium or Jigsaw to withstand such loads (see http://www.sics.se/ joe/apachevsyaws.html). Figure 2 (left part) shows the average URL extraction per document crawled using a single worker. The abscissa (x) axis represents the number of documents treated, and the ordered (y) axis gives the time in microseconds corresponding to extraction. In the right-hand figure, the abscissa axis represents the same quantity, though this time in terms of data volume (Mb). We can see a high level of parsing reaching an average of 3,000 pages/second at a speed of 70Mb/second. In Figure 3 we see that URL normalization 0 500000 1e+06 1.5e+06 2e+06 2.5e+06 3e+06 3.5e+06 0 2000 4000 6000 8000 10000 Time (microsec) Documents Average number of parsed documents PD 0 500000 1e+06 1.5e+06 2e+06 2.5e+06 3e+06 3.5e+06 0 20 40 60 80 100 120 140 160 Time (microsec) Document Size (Mb) Average size of parsed documents PDS Figure 2: Link Extraction is as efficient as extraction in terms of speed. The abscissa axis at the top (and respectively at the bottom) represents the number of documents processed per normalization phase (respectively the quantity of documents in terms of volume). Each worker normalizes on average 1,000 documents/second , which is equivalent to 37,000 URLs/second at a speed of 40Mb/second. Finally, the URL cache structure ensures a high degree of scalability (Figure 3). The abscissa axis in this figure represents the number of key-values inserted or retrieved. The cache is very close to a step function due to key compression in the Judy-Array. Following an increase in insertion/retrieval time in the cache, it appears to plateau by 100,000 key-value bands. We should however point out that URL extraction and normalization also makes use of this type of cache so as to avoid processing a URL already encountered. 305 0 10000 20000 30000 40000 50000 60000 0 2000 4000 6000 8000 10000 Time (microsec) Normalized documents Average number of normalized documents AD 0 10000 20000 30000 40000 50000 60000 0 2000 4000 6000 8000 10000 12000 14000 16000 Time (microsec) Urls Average number of normalized Url AU 0 10000 20000 30000 40000 50000 60000 0 20 40 60 80 100 120 140 160 Time (microsec) Document Size (Mb) Average size of normalized documents ADS 0 50000 100000 150000 200000 250000 300000 350000 0 20000 40000 60000 80000 100000 Time (microsec) Key-Value Scalable Cache : Insertion vs Retrieval Cache Insertion Cache Retrieval Figure 3: URL Normalization and Cache Performance CONCLUSION In the present paper, we have introduced a high availability system of crawling called Dominos. This system has been created in the framework of experimentation for French Web legal deposit carried out at the Institut National de l'Audiovisuel (INA). Dominos is a dynamic system, whereby the processes making up its kernel are mobile. 90% of this system was developed using Erlang programming language, which accounts for its highly flexible deployment, maintainability and enhanced fault tolerance. Despite having different objectives, we have been able to compare it with other documented Web crawling systems (Mercator, InternetArchive . . . ) and have shown it to be superior in terms of crawl speed, document parsing and process management without system restart. Dominos is more complex than its description here. We have not touched upon archival storage and indexation. We have preferred to concentrate rather on the detail of implementation of the Dominos kernel itself, a strategic component that is often overlooked by other systems (in particular those that are proprietary, others being inefficient). However, there is still room for the system's improvement. At present, crawled archives are managed by NFS, a file system that is moderately efficient for this type of problem. Switchover to Lustre 9 , a distributed file system with a radically higher level of performance, is underway. REFERENCES [1] Z. BarYossef, A. Berg, S. Chien, J. Fakcharoenphol, and D. Weitz. Approximating aggregate queries about web pages via random walks. In Proc. of 26th Int. Conf. on Very Large Data Bases, 2000. [2] K. Bharat, A. Broder, M. Henzinger, P. Kumar, and S. Venkatasubramanian. The connectivity server: Fast access to linkage. Information on the Web, 1998. [3] S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. In Proc. of the Seventh World-Wide Web Conference, 1998. [4] M. Burner. Crawling towards eternity: Building an archive of the world wide web. 9 http://www.lustre.org/ http://www.webtechniques.com/archives/1997/05/burner/, 1997. [5] S. Chakrabarti, M. V. D. Berg, and B. Dom. Distributed hypertext resource discovery through example. In Proc. of 25th Int. Conf. on Very Large Data Base, pages 375386, 1997. [6] S. Chakrabarti, M. V. D. Berg, and B. Dom. Focused crawling: A new approach to topic-specific web resource discovery. In Proc. of the 8th Int. World Wide Web Conference, 1999. [7] J. Cho and H. Garcia-Molina. The evolution of the web and implications for an incremental crawler. In Proc. of 26th Int. Conf. on Very Large Data Bases, pages 117128, 2000. [8] J. Cho and H. Garcia-Molina. Synchronizing a database to improve freshness. In Proc. of the ACM SIGMOD Int. Conf. on Management of Data, 2000. [9] J. Cho, H. Garcia-Molina, and L. Page. Efficient crawling through url ordering. In 7th Int. World Wide Web Conference, 1998. [10] M. Diligenti, F. Coetzee, S. Lawrence, C. Giles, and M. Gori. Focused crawling using context graphs. In Proc. of 26th Int. Conf. on Very Large Data Bases, 2000. [11] M. Henzinger, A. Heydon, M. Mitzenmacher, and M. Najork. Measuring index quality using random walks on the web. In Proc. of the 8th Int. World Wide Web Conference (WWW8), pages 213225, 1999. [12] M. Henzinger, A. Heydon, M. Mitzenmacher, and M. Najork. On near-uniform url sampling. In Proc. of the 9th Int. World Wide Web Conference, 2000. [13] A. Heydon and M. Najork. Mercator: A scalable, extensible web crawler. World Wide Web Conference, pages 219229, 1999. [14] J. Hirai, S. Raghavan, H. Garcia-Molina, and A. Paepcke. Webbase : : A repository of web pages. In Proc. of the 9th Int. World Wide Web Conference, 2000. [15] B. Kahle. Archiving the internet. Scientific American, 1997. [16] S. Lawrence and C. L. Giles. Searching the world wide web. Science 280, pages 98100, 1998. [17] M. Najork and J. Wiener. Breadth-first search crawling yields high-quality pages. In 10th Int. World Wide Web Conference, 2001. [18] J. Rennie and A. McCallum. Using reinforcement learning to spider the web efficiently. In Proc. of the Int. Conf. on Machine Learning, 1999. [19] S. Russel and P. Norvig. Artificial Intelligence: A modern Approach. Prentice Hall, 1995. [20] V. Shkapenyuk and T. Suel. Design and implementation of a high-performance distributed web crawler. Polytechnic University: Brooklyn, Mars 2001. [21] T. Suel and J. Yuan. Compressing the graph structure of the web. In Proc. of the IEEE Data Compression Conference, 2001. [22] J. Talim, Z. Liu, P. Nain, and E. Coffman. Controlling robots of web search engines. In SIGMETRICS Conference, 2001. 306

Breadth first crawling;Hierarchical Cooperation;limiting disk access;fault tolerance;Dominos nodes;dominos process;Dominos distributed database;breadth-first crawling;repetitive crawling;URL caching;Dominos Generic server;Document fingerprint;Deep web crawling;Dominos RPC concurrent;Random walks and sampling;Web Crawler;maintaiability and configurability;deep web crawling;High Availability System;real-time distributed system;crawling system;high performance crawling system;high availability;Erlang development kit;targeted crawling

101

Hiperlan/2 Public Access Interworking with 3G Cellular Systems

This paper presents a technical overview of the Hiperlan/2 3G interworking concept. It does not attempt to provide any business justification or plan for Public Access operation. After a brief resume of public access operation below, section 2 then introduces an overview of the technologies concerned. Section 3 describes the system approach and presents the current reference architecture used within the BRAN standardisation activity. Section 4 then goes on to cover in more detail the primary functions of the system such as authentication, mobility, quality of service (QoS) and subscription. It is worth noting that since the Japanese WLAN standard HiSWANa is very similar to Hiperlan/2, much of the technical information within this paper is directly applicable to this system, albeit with some minor changes to the authentication scheme. Additionally the high level 3G and external network interworking reference architecture is also applicable to IEEE 802.11. Finally, section 5 briefly introduces the standardisation relationships between ETSI BRAN, WIG, 3GPP, IETF, IEEE 802.11 and MMAC HSWA.

1.1. Public access operation Recently, mobile business professionals have been looking for a more efficient way to access corporate information systems and databases remotely through the Internet backbone. However, the high bandwidth demand of the typical office applications , such as large email attachment downloading, often calls for very fast transmission capacity. Indeed certain hot spots, like hotels, airports and railway stations are a natural place to use such services. However, in these places the time available for information download typically is fairly limited. In light of this, there clearly is a need for a public wireless access solution that could cover the demand for data intensive applications and enable smooth on-line access to corporate data services in hot spots and would allow a user to roam from a private, micro cell network (e.g., a Hiperlan/2 Network) to a wide area cellular network or more specifically a 3G network. Together with high data rate cellular access, Hiperlan/2 has the potential to fulfil end user demands in hot spot environments . Hiperlan/2 offers a possibility for cellular operators to offer additional capacity and higher bandwidths for end users without sacrificing the capacity of the cellular users, as Hiperlans operate on unlicensed or licensed exempt frequency bands. Also, Hiperlan/2 has the QoS mechanisms that are capable to meet the mechanisms that are available in the 3G systems. Furthermore, interworking solutions enable operators to utilise the existing cellular infrastructure investments and well established roaming agreements for Hiperlan/2 network subscriber management and billing. Technology overview This section briefly introduces the technologies that are addressed within this paper. 2.1. Hiperlan/2 summary Hiperlan/2 is intended to provide local wireless access to IP, Ethernet, IEEE 1394, ATM and 3G infrastructure by both stationary and moving terminals that interact with access points. The intention is that access points are connected to an IP, Ethernet , IEEE 1394, ATM or 3G backbone network. A number of these access points are required to service all but the small-44 MCCANN AND FLYGARE est networks of this kind, and therefore the wireless network as a whole supports handovers of connections between access points. 2.2. Similar WLAN interworking schemes It should be noted that the interworking model presented in this paper is also applicable to the other WLAN systems, i.e. IEEE 802.11a/b and MMAC HiSWANa (High Speed Wireless Access Network), albeit with some minor modifications to the authentications schemes. It has been the intention of BRAN to produce a model which not only fits the requirements of Hiperlan/23G interworking, but also to try and meet those of the sister WLAN systems operating in the same market. A working agreement has been underway between ETSI BRAN and MMAC HSWA for over 1 year, and with the recent creation of WIG (see section 5), IEEE 802.11 is also working on a similar model. 2.3. 3G summary Within the framework of International Mobile Telecommunications 2000 (IMT-2000), defined by the International Telecommunications Union (ITU), the 3rd Generation Partnership Project (3GPP) are developing the Universal Mobile Telecommunications System (UMTS) which is one of the major third generation mobile systems. Additionally the 3rd Generation Partnership Project 2 (3GPP2) is also developing another 3G system, Code Division Multiple Access 2000 (CDMA-2000). Most of the work within BRAN has concentrated on UMTS, although most of the architectural aspects are equally applicable to Hiperlan/2 interworking with CDMA-2000 and indeed pre-3G systems such as General Packet Radio Services (GPRS). The current working UMTS standard, Release 4, of UMTS was finalised in December 2000 with ongoing development work contributing to Release 5, due to be completed by the end of 2002. A future release 6 is currently planned for the autumn of 2003, with worldwide deployment expected by 2005. System approach This section describes the current interworking models being worked upon within BRAN at the current time. The BRAN Network Reference Architecture, shown in figure 1, identifies the functions and interfaces that are required within a Hiperlan/2 network in order to support inter-operation with 3G systems . The focus of current work is the interface between the Access Point (AP) and the Service provider network (SPN) which is encapsulated by the Lx interface. The aim of the Hiperlan/23G interworking work item is to standardise these interfaces, initially focusing on AAA (Authentication, Authorisation and Accounting) functionality. A secondary aim is to create a model suitable for all the 5 GHz WLAN systems (e.g., Hiperlan/2, HiSWANa, IEEE Figure 1. Reference architecture. 802.11a) and all 3G systems (e.g., CDMA-2000, UMTS), thus creating a world wide standard for interworking as mentioned in section 5. Other interfaces between the AP and external networks and interfaces within the AP are outside the scope of this current work. Figure 1 shows the reference architecture of the interworking model. It presents logical entities within which the following functions are supported: Authentication: supports both SIM-based and SIM-less authentication. The mobile terminal (MT) communicates via the Attendant with an authentication server in the visited network, for example a local AAA server, across the Ls interface. Authorisation and User Policy: the SPN retrieves authorisation and user subscription information from the home network when the user attaches to it. Authorisation information is stored within a policy decision function in the SPN. Interfaces used for this are Lp and Ls. Accounting: the resources used by a MT and the QoS provided to a user are both monitored by the Resource Monitor . Accounting records are sent to accounting functions in the visited network via the La interface. Network Management: the Management Agent provides basic network and performance monitoring, and allows the configuration of the AP via the Lm interface. Admission Control and QoS: a policy decision function in the SPN decides whether a new session with a requested QoS can be admitted based on network load and user subscription information. The decision is passed to the Policy Enforcement function via the Lp interface. Inter-AP Context Transfer: the Handover Support function allows the transfer of context information concerning a user/mobile node, e.g., QoS state, across the Lh interface from the old to the new AP between which the mobile is handing over. HIPERLAN/2 PUBLIC ACCESS INTERWORKING WITH 3G CELLULAR SYSTEMS 45 Mobility: mobility is a user plane function that performs re-routing of data across the network. The re-routing may simply be satisfied by layer 2 switching or may require support for a mobility protocol such as Mobile IP depending on the technology used within the SPN. Mobility is an attribute of the Lr interface. Location Services: the Location Server function provides positioning information to support location services. Information is passed to SPN location functions via the Ll interface. Primary functions This section describes the primary functions of this model (refer to figure 1) in further detail, specifically: authentication and accounting, mobility and QoS. 4.1. Authentication and authorisation A key element to the integration of disparate systems is the ability of the SPN to extract both authentication and subscription information from the mobile users' home networks when an initial association is requested. Many users want to make use of their existing data devices (e.g., Laptop, Palmtop) without additional hardware/software requirements. Conversely for both users and mobile operators it is beneficial to be able to base the user authentication and accounting on existing cellular accounts, as well as to be able to have Hiperlan/2-only operators and users; in any case, for reasons of commonality in MT and network (indeed SPN) development it is important to be able to have a single set of AAA protocols which supports all the cases. 4.1.1. Loose coupling The rest of this paper concentrates on loose coupling solutions . "Loose coupling", is generally defined as the utilisation of Hiperlan/2 as a packet based access network complementary to current 3G networks, utilising the 3G subscriber databases but without any user plane Iu type interface, as shown in figure 1. Within the UMTS context, this scheme avoids any impact on the SGSN and GGSN nodes. Security, mobility and QoS issues are addressed using Internet Engineering Task Force (IETF) schemes. Other schemes which essentially replace the User Terminal Radio Access Network (UTRAN) of UMTS with a HIRAN (Hiperlan Radio Access Network) are referred to as "Tight Coupling", but are not currently being considered within the work of BRAN. 4.1.2. Authentication flavours This section describes the principle functions of the loose coupling interworking system and explains the different authentication flavours that are under investigation. The focus of current work is the interface between the AP and the SPN. Other interfaces between the AP and external networks and interfaces within the AP are initially considered to be implementation or profile specific. The primary difference between these flavours is in the authentication server itself, and these are referred to as the "IETF flavour" and the "UMTS-HSS flavour", where the Home Subscriber Server (HSS) is a specific UMTS term for a combined AAA home server (AAAH)/Home Location Register (HLR) unit. The motivation for network operators to build up Hiperlan/2 networks based on each flavour may be different for each operator. However, both flavours offer a maximum of flexibility through the use of separate Interworking Units (IWU) and allow loose coupling to existing and future cellular mobile networks. These alternatives are presented in figure 2. IETF flavour. The IETF flavour outlined in figure 2 is driven by the requirement to add only minimal software functionality to the terminals (e.g., by downloading java applets), so that the use of a Hiperlan/2 mobile access network does not require a radical change in the functionality (hardware or software ) compared to that required by broadband wireless data access in the corporate or home scenarios. Within a multiprovider network, the WLAN operator (who also could be a normal ISP) does not necessary need to be the 3G operator as well, but there could still be an interworking between the networks. Within this approach Hiperlan/2 users may be either existing 3G subscribers or just Hiperlan/2 network subscribers. These users want to make use of their existing data devices (e.g., Laptop, Palmtop) without additional hardware/software requirements. For both users and mobile operators it is beneficial to be able to base the user authentication and accounting on existing cellular accounts, as well as to be able to have Hiperlan/2-only operators and users; in any case, for reasons of commonality in MT and AP development it is important to be able to have a single set of AAA protocols which supports all the cases. UMTS-HSS flavour. Alternatively the UMTS flavour (also described within figure 1) allows a mobile subscriber using a Hiperlan/2 mobile access network for broadband wireless data access to appear as a normal cellular user employing standard procedures and interfaces for authentication purposes . It is important to notice that for this scenario functionality normally provided through a user services identity module (USIM) is required in the user equipment. The USIM provides new and enhanced security features in addition to those provided by 2nd Generation (2G) SIM (e.g., mutual authentication ) as defined by 3GPP (3G Partnership Program). The UMTS-HSS definitely requires that a user is a native cellular subscriber while in addition and distinctly from the IETF flavoured approach standard cellular procedures and parameters for authentication are used (e.g., USIM quintets). In this way a mobile subscriber using a Hiperlan/2 mobile access network for broadband wireless data access will appear as a normal cellular user employing standard procedures and interfaces for authentication purposes. It is important to notice that for this scenario USIM functionality is required in the user equipment. 46 MCCANN AND FLYGARE Figure 2. Loose coupling authentication flavours. For the IETF flavoured approach there is no need to integrate the Hiperlan/2 security architecture with the UMTS security architecture [2]. It might not even be necessary to implement all of the Hiperlan/2 security features if security is applied at a higher level, such as using IPsec at the IP level. An additional situation that must be considered is the use of pre-paid SIM cards. This scenario will introduce additional requirements for hot billing and associated functions. 4.1.3. EAPOH For either flavour authentication is carried out using a mechanism based on EAP (Extensible Authentication Protocol) [3]. This mechanism is called EAPOH (EAP over Hiperlan/2) and is analogous to the EAPOL (EAP over LANs) mechanism as defined in IEEE 802.1X. On the network side, Diameter [4] is used to relay EAP packets between the AP and AAAH. Between the AP and MT, EAP packets and additional Hiperlan/2 specific control packets (termed pseudo-EAP packets) are transferred over the radio interface. This scheme directly supports IETF flavour authentication, and by use of the pro-posed EAP AKA (Authentication and Key Agreement) mechanism would also directly support the UMTS flavour authentication . Once an association has been established, authorisation information (based on authentication and subscription) stored within a Policy Decision Function within the SPN itself can be transmitted to the AP. This unit is then able to regulate services such as time-based billing and allocation of network and radio resources to the required user service. Mobile users with different levels of subscription (e.g., "bronze, silver, gold") can be supported via this mechanism, with different services being configured via the policy interface. A change in authentication credentials can also be managed at this point. 4.1.4. Key exchange Key agreement for confidentiality and integrity protection is an integral part of the UMTS authentication procedure, and hence the UTRAN confidentiality and integrity mechanisms should be reused within the Hiperlan/2 when interworking with a 3G SPN (i.e. core network). This will also increase the applied level of security. The Diffie-Hellman encryption key agreement procedure, as used by the Hiperlan/2 air interface, could be used to improve user identity confidentiality. By initiating encryption before UMTS AKA is performed, the user identity will not have to be transmitted in clear over the radio interface, as is the case in UMTS when the user enters a network for the first time. Thus, this constitutes an improvement compared to UMTS security. It is also important to have a secure connection between APs within the same network if session keys or other sensitive information are to be transferred between them. A secure connection can either be that they for some reason trust each other and that no one else can intercept the communication between them or that authentication is performed and integrity and confidentiality protection are present. 4.1.5. Subscriber data There are three basic ways in which the subscriber management for Hiperlan/2 and 3G users can be co-ordinated: Have the interworking between the Hiperlan/2 subscriber database and HLR/HSS. This is for the case where the in-HIPERLAN/2 PUBLIC ACCESS INTERWORKING WITH 3G CELLULAR SYSTEMS 47 terworking is managed through a partnership or roaming agreement. The administrative domains' AAA servers share security association or use an AAA broker. The Hiperlan/2 authentication could be done on the basis of a (U)SIM token. The 3G authentication and accounting capabilities could be extended to support access authentication based on IETF protocols. This means either integrating HLR and AAA functions within one unit (e.g., a HSS unit), or by merging native HLR functions of the 3G network with AAA functions required to support IP access. Based on these different ways for subscriber management, the user authentication identifier can be on three different formats : Network Address Identifier (NAI), International Mobile Subscriber Identity (IMSI) (requires a (U)SIM card), and IMSI encapsulated within a NAI (requires a (U)SIM card). 4.1.6. Pre-paid SIM cards As far as the HLR within the SPN is concerned, it cannot tell the difference between a customer who is pre-paid or not. Hence, this prevents a non-subscriber to this specific 3G network from using the system, if the operator wishes to impose this restriction. As an example, pre-paid calls within a 2G network are handled via an Intelligent Network (IN) probably co-located with the HLR. When a call is initiated, the switch can be pro-grammed with a time limit, or if credit runs out the IN can signal termination of the call. This then requires that the SPN knows the remaining time available for any given customer. Currently the only signals that originate from the IN are to terminate the call from the network side. This may be undesirable in a Hiperlan/23G network, so that a more graceful solution is required. A suitable solution is to add pre-paid SIM operation to our system together with hot billing (i.e. bill upon demand) or triggered session termination . This could be achieved either by the AAAL polling the SPN utilising RADIUS [5] to determine whether the customer is still in credit, or by using a more feature rich protocol such as Diameter [4] which allows network signalling directly to the MT. The benefit of the AAA approach is to allow the operator to present the mobile user with a web page (for example), as the pre-paid time period is about to expire, allowing them to purchase more airtime. All these solutions would require an increased integration effort with the SPN subscriber management system. Further additional services such as Customized Applications for Mobile Network Enhanced Logic (CAMEL) may also allow roaming with pre-paid SIM cards. 4.2. Accounting In the reference architecture of figure 2, the accounting function monitors the resource usage of a user in order to allow cost allocation, auditing and billing. The charging/accounting is carried out according to a series of accounting and resource monitoring metrics, which are derived from the policy function and network management information. The types of information needed in order to monitor each user's resource consumption could include parameters such as, for example, volume of traffic, bandwidth consumption, etc. Each of these metrics could have AP specific aspects concerning the resources consumed over the air interface and those consumed across the SPN, respectively. As well as providing data for billing and auditing purposes, this information is exchanged with the Policy Enforcement/Decision functions in order to provide better information on which to base policy decisions. The accounting function processes the usage related information including summarisation of results and the generation of session records. This information may then be forwarded to other accounting functions within and outside the network, for example a billing function. This information may also be passed to the Policy Decision function in order to improve the quality of policy decisions; vice versa the Policy Decision function can give information about the QoS, which may affect the session record. There are also a number of extensions and enhancements that can be made to the basic interworking functionality such as those for the provision of support for QoS and mobility. In a multiprovider network, different sorts of inter-relationships between the providers can be established. The inter-relationship will depend upon commercial conditions, which may change over time. Network Operators have exclusive agreements with their customers, including charging and billing, and also for services provided by other Network Op-erators/Service Providers. Consequently, it must be possible to form different charging and accounting models and this requires correspondent capabilities from the networks. Charging of user service access is a different issue from the issue of accounting between Network Operators and Service Providers. Although the issues are related, charging and accounting should be considered separately. For the accounting issue it is important for the individual Network Operator or Service Provider to monitor and register access use provided to his customers. Network operators and service providers that regularly provide services to the same customers could either charge and bill them individually or arrange a common activity. For joint provider charging/billing, the providers need revenue accounting in accordance with the service from each provider. For joint provider charging of users, it becomes necessary to transfer access/session related data from the providers to the charging entity. Mechanisms for revenue accounting are needed, such as technical configuration for revenue accounting . This leads to transfer of related data from the Network 48 MCCANN AND FLYGARE Operator and/or Service Providers to the revenue accounting entity. The following parameters may be used for charging and revenue accounting: basic access/session (pay by subscription), toll free (like a 0800 call), premium rate access/session, access/session duration, credit card access/session, pre-paid, calendar and time related charging, priority, Quality of Service, duration dependent charging, flat rate, volume of transferred packet traffic, rate of transferred packet traffic (Volume/sec), multiple rate charge. 4.3. Mobility Mobility can be handled by a number of different approaches. Indeed many mobility schemes have been developed in the IETF that could well be considered along with the work of the MIND (Mobile IP based Network Developments) project that has considered mobility in evolved IP networks with WLAN technologies. Mobility support is desirable as this functionality would be able to provide support for roaming with an active connection between the interworked networks, for example , to support roaming from UMTS to WLAN in a hotspot for the downloading of large data. In the loose coupling approach, the mobility within the Hiperlan/2 network is provided by native Hiperlan/2 (i.e. RLC layer) facilities, possibly extended by the Convergence Layer (CL) in use (e.g., the current Ethernet CL [6], or a future IP CL). This functionality should be taken unchanged in the loose coupling approach, i.e. handover between access points of the same Hiperlan/2 network does not need to be considered especially here as network handover capabilities of Hiperlan/2 RLC are supported by both MTs and APs. Given that Hiperlan/2 network handover is supported, further details for completing the mobility between access points are provided by CL dependent functionality. Completion of this functionality to cover interactions between the APs and other parts of the network (excluding the terminal and therefore independent of the air interface) are currently under development outside BRAN. In the special case where the infrastructure of a single Hiperlan/2 network spans more than one IP sub-network, some of the above approaches assume an additional level of mobility support that may involve the terminal. 4.3.1. Roaming between Hiperlan/2 and 3G For the case of mobility between Hiperlan/2 and 3G access networks, recall that we have the following basic scenario: A MT attaches to a Hiperlan/2 network, authenticates and acquires an IP address. At that stage, it can access IP services using that address while it remains within that Hiperlan/2 network . If the MT moves to a network of a different technology (i.e. UMTS), it can re-authenticate and acquire an IP address in the packet domain of that network, and continue to use IP services there. We have referred to this basic case as AAA roaming. Note that while it provides mobility for the user between networks, any active sessions (e.g., multimedia calls or TCP connections ) will be dropped on the handover between the networks because of the IP address change (e.g., use Dynamic Host Configuration Protocol DHCP). It is possible to provide enhanced mobility support, including handover between Hiperlan/2 access networks and 3G access networks in this scenario by using servers located outside the access network. Two such examples are: The MT can register the locally acquired IP address with a Mobile IP (MIP) home agent as a co-located care-of address , in which case handover between networks is handled by mobile IP. This applies to MIPv4 and MIPv6 (and is the only mode of operation allowed for MIPv6). The MT can register the locally acquired IP address with an application layer server such as a Session Initiation Protocol (SIP) proxy. Handover between two networks can then be handled using SIP (re-invite message). Note that in both these cases, the fact that upper layer mobility is in use is visible only to the terminal and SPN server, and in particular is invisible to the access network. Therefore, it is automatically possible, and can be implemented according to existing standards, without impact on the Hiperlan/2 network itself. We therefore consider this as the basic case for the loose coupling approach. Another alternative is the use of a Foreign Agent care-of address (MIPv4 only). This requires the integration of Foreign Agent functionality with the Hiperlan/2 network, but has the advantage of decreasing the number of IPv4 addresses that have to be allocated. On the other hand, for MTs that do not wish to invoke global mobility support in this case, a locally assigned IP address is still required, and the access network therefore has to be able to operate in two modes. Two options for further study are: The option to integrate access authentication (the purpose of this loose coupling standard) with Mobile IP home agent registration (If Diameter is used, it is already present). This would allow faster attach to the network in the case of a MT using MIP, since it only requires one set of authentication exchanges; however, it also requires integration on the control plane between the AAAH and the Mobile IP home agent itself. It is our current assumption that this integration should be carried out in a way that is independent of the particular access network being used, and is therefore out of scope of this activity. HIPERLAN/2 PUBLIC ACCESS INTERWORKING WITH 3G CELLULAR SYSTEMS 49 The implications of using services (e.g., SIP call control ) from the UMTS IMS (Internet Multimedia Subsys-tem ), which would provide some global mobility capability . This requires analysis of how the IMS would interface to the Hiperlan/2 access network (if at all). 4.3.2. Handover For handovers within the Hiperlan/2 network, the terminal must have enough information to be able to make a handover decision for itself, or be able to react to a network decision to handover. Indeed these decision driven events are referred to as triggers, resulting in Network centric triggers or Terminal centric triggers. Simple triggers include the following: Network Centric: Poor network resources or low bandwidth , resulting in poor or changing QoS. Change of policy based on charging (i.e. end of pre-paid time). Terminal Centric: Poor signal strength. Change of QoS. 4.4. QoS QoS support is available within the Hiperlan/2 specification but requires additional functionality in the interworking specifications for the provision of QoS through the CN rather than simply over the air. QoS is a key concept, within UMTS, and together with the additional QoS functionality in Hiperlan/2, a consistent QoS approach can therefore be provided. A number of approaches to QoS currently exist which still need to be considered at this stage. QoS within the Hiperlan/2 network must be supported between the MT and external networks, such as the Internet. In the loose coupling scenario, the data path is not constrained to travelling across the 3G SPN, e.g., via the SGSN/GGSNs. Therefore no interworking is required between QoS mechanisms used within the 3G and Hiperlan/2 network. There is a possible interaction regarding the interpretation and mapping of UMTS QoS parameters onto the QoS mechanisms used in the Hiperlan/2 network. The actual provisioning of QoS across the Hiperlan/2 network is dependent on the type of the infrastructure technology used, and therefore the capabilities of the CL. 4.4.1. HiperLAN2/Ethernet QoS mapping Within the Hiperlan/2 specification, radio bearers are referred to as DLC connections. A DLC connection is characterised by offering a specific support for QoS, for instance in terms of bandwidth, delay, jitter and bit error rate. The characteristics of supported QoS classes are implementation specific. A user might request for multiple DLC connections, each transferring a specific traffic type, which indicates that the traffic division is traffic type based and not application based. The DLC connection set-up does not necessarily result in immediate assignment of resources though. If the MT has not negotiated a fixed capacity agreement with the AP, it must request capacity by sending a resource request (RR) to the AP whenever it has data to transmit. The allocation of resources may thereby be very dynamic. The scheduling of the resources is vendor specific and is therefore not included in the Hiperlan/2 standard, which also means that QoS parameters from higher layers are not either. Hiperlan/2 specific QoS support for the DLC connection comprises centralised resource scheduling through the TDMA-based MAC structure, appropriate error control (acknowledged , unacknowledged or repetition) with associated protocol settings (ARQ window size, number of retransmis-sions and discarding), and the physical layer QoS support. Another QoS feature included in the Hiperlan/2 specification is a polling mechanism that enables the AP to regularly poll the MT for its traffic status, thus providing rapid access for real-time services. The CL acts as an integrator of Hiperlan/2 into different existing service provider networks, i.e. it connects the SPNs to the Hiperlan/2 data link control (DLC) layer. IEEE 802.1D specifies an architecture and protocol for MAC bridges interconnecting IEEE 802 LANs by relaying and filtering frames between the separate MACs of the Bridged LAN. The priority mechanism within IEEE 802.1D is handled by IEEE 802.1p, which is incorporated into IEEE 802.1D. All traffic types and their mappings presented in the tables of this section only corresponds to default values specified in the IEEE 802.1p standard, since these parameters are vendor specific. IEEE 802.1p defines eight different priority levels and describes the traffic expected to be carried within each priority level. Each IEEE 802 LAN frame is marked with a user priority (07) corresponding to the traffic type [8]. In order to support appropriate QoS in Hiperlan/2 the queues are mapped to the different QoS specific DLC connections (maximum of eight). The use of only one DLC connection between the AP and the MT results in best effort traffic only, while two to eight DLC connections indicates that the MT wants to apply IEEE 802.1p. A DLC connection ID is only MT unique, not cell unique. The AP may take the QoS parameters into account in the allocation of radio resources (which is out of the Hiperlan/2 scope). This means that each DLC connection, possibly operating in both directions, can be assigned a specific QoS, for instance in terms of bandwidth, delay, jitter and bit error rate, as well as being assigned a priority level relative to other DLC connections. In other words, parameters provided by the application, including UMTS QoS parameters if desired , are used to determine the most appropriate QoS level to be provided by the network, and the traffic flow is treated accordingly. The support for IEEE 802.1p is optional for both the MT and AP. 4.4.2. End-to-end based QoS Adding QoS, especially end-to-end QoS, to IP based connections raises significant alterations and concerns since it represents a digression from the "best-effort" model, which constitutes the foundation of the great success of Internet. However, the need for IP QoS is increasing and essential work is cur-50 MCCANN AND FLYGARE rently in progress. End-to-end IP QoS requires substantial consideration and further development. Since the Hiperlan/2 network supports the IEEE 802.1p priority mechanism and since Differentiated Services (DiffServ ) is priority based, the natural solution to the end-to-end QoS problem would be the end-to-end implementation of DiffServ. The QoS model would then appear as follows. QoS from the MT to the AP is supported by the Hiperlan/2 specific QoS mechanisms, where the required QoS for each connection is identified by a unique Data Link Control (DLC) connection ID. In the AP the DLC connection IDs may be mapped onto the IEEE 802.1p priority queues. Using the IEEE 802.1p priority mechanisms in the Ethernet, the transition to a DiffServ network is easily realised by mapping the IEEE 802.1p user priorities into DiffServ based priorities. Neither the DiffServ nor the IEEE 802.1p specification elaborates how a particular packet stream will be treated based on the Differentiated Services (DS) field and the layer 2 priority level. The mappings between the IEEE 802.1p priority classes and the DiffServ service classes are also unspec-ified . There is however an Integrated Services over Specific Link Layers (ISSLL) draft mapping for Guaranteed and Controlled Load services to IEEE 802.1p user priority, and a mapping for Guaranteed and Controlled Load services, to DiffServ which together would imply a DiffServ to IEEE 802.1p user priority mapping. DiffServ provides inferior support of QoS than IntServ, but the mobility of a Hiperlan/2 MT indicates a need to keep the QoS signalling low. IntServ as opposed to DiffServ involves significant QoS signalling. The DiffServ model provides less stringent support of QoS than the IntServ/RSVP model but it has the advantage over IntServ/RSVP of requiring less protocol signalling, which might be a crucial factor since the mobility of a Hiperlan/2 MT indicates a need to keep the QoS signalling low. Furthermore , the implementation of an end-to-end IntServ/RSVP based QoS architecture is much more complex than the implementation of a DiffServ based one. Discussions around end-to-end QoS support raise some critical questions that need to be considered and answered before a proper solution can be developed; which performance can we expect from the different end-to-end QoS models, what level of QoS support do we actually need, how much bandwidth and other resources are we willing to sacrifice on QoS, and how much effort do we want to spend on the process of developing well-supported QoS? Relationships with other standardisation bodies BRAN is continuing to have a close working relationship with the following bodies: WLAN Interworking Group (WIG) This group met for the first time in September 2002. Its broad aim is to provide a single point of contact for the three main WLAN standardisation bodies (ETSI BRAN, IEEE 802.11 and MMAC HSWA) and to produce a generic approach to both Cellular and external network interworking of WLAN technology. It has been also decided to work upon, complete and then share a common standard for WLAN Public Access and Cellular networks. 3rd Generation Partnership Project (3GPP) The System Architecture working group 1 (SA1) is currently developing a technical report detailed the requirements for a UMTSWLAN interworking system. They have defined 6 scenarios detailing aspects of differently coupled models, ranging from no coupling, through loose coupling to tight coupling. Group 2 (SA2) is currently investigating reference architecture models, concentrating on the network interfaces towards the WLAN. Group 3 (SA3) has now started work on security and authentication issues with regard to WLAN interworking . ETSI BRAN is currently liasing with the SA2 and SA3 groups. Internet Engineering Task Force (IETF) Within the recently created `eap' working group, extensions are being considered to EAP (mentioned in section 4), which will assist in system interworking. Institute of Electrical and Electronics Engineers (IEEE) USA The 802.11 WLAN technical groups are continuing to progress their family of standards. Many similarities exist between the current 802.11a standard and Hiperlan2/HiSWANa with regard to 3G interworking. ETSI BRAN is currently liasing with the Wireless Next Generation (WNG) group of the IEEE 802.11 project. Multimedia Mobile Access Communication (MMAC) Japan The High Speed Wireless Access (HSWA) group's HiSWANa (High Speed Wireless Access Network system A) is essentially identical to Hiperlan/2, except that it mandates the use of an Ethernet convergence layer within the access point. An agreement between ETSI BRAN and MMAC HSWA has now been in place for some time to share the output of the ETSI BRAN 3G interworking group. Conclusions This paper has addressed some of the current thinking within ETSI BRAN (and indeed WIG) regarding the interworking of the Hiperlan2 and HiSWANa wireless LAN systems into a 3G Cellular System. Much of this information is now appearing in the technical specification being jointly produced by ETSI and MMAC, expected to be published in the first half of 2003. Of the two initial solutions investigated (tight and loose coupling), current work has concentrated on the loose variant, producing viable solutions for security, mobility and QoS. The authentication schemes chosen will assume that EAP is carried over the air interface, thus being compatible, at the interworking level, with IEEE 802.11 and 3GPP. HIPERLAN/2 PUBLIC ACCESS INTERWORKING WITH 3G CELLULAR SYSTEMS 51 This standardisation activity thus hopes to ensure that all WLAN technologies can provide a value added service within hotspot environments for both customers and operators of 3G systems. Acknowledgements The authors wish to thank Maximilian Riegel (Siemens AG, Germany), Dr. Robert Hancock and Eleanor Hepworth (Roke Manor, UK) together with se Jevinger (Telia Research AB, Sweden) for their invaluable help and assistance with this work. References [1] ETSI TR 101 957 (V1.1.1): Broadband Radio Access Networks (BRAN); HIPERLAN Type 2; Requirements and Architectures for Interworking between Hiperlan/2 and 3rd Generation Cellular Systems (August 2001). [2] 3GPP TS 33.102: 3rd Generation Partnership Project; Technical Specification Group Services and System Aspects; 3G Security; Security Architecture . [3] L. Blunk, J. Vollbrecht and B. Aboba, PPP Extensible Authentication Protocol (EAP), RFC 2284bis, draft-ietf-pppext-rfc2284bis -04.txt (April 2002). [4] P. Calhoun et al., Diameter base protocol, draft-ietf-aaa-diameter -10 (April 2002). [5] C. Rigney et al., Remote Authentication Dial In User Service (RADIUS), RFC 2058 (January 1997). [6] HIPERLAN Type 2; Packet Based Convergence Layer; Part 2: Ethernet Service Specific Convergence Sublayer (SSCS), ETSI TS 101 493-2, BRAN. [7] HIPERLAN Type 2; System overview, ETSI TR 101 683, BRAN. [8] Information Technology Telecommunications and Information Exchange between Systems Local and Metropolitan Area Networks Common Specifications Part 3: Media Access Control (MAC) Bridges (Revision and redesignation of ISO/IEC 10038: 1993 [ANSI/IEEE Std 802.1D, 1993 Edition], incorporating IEEE supplements P802.1p, 802.1j-1996, 802.6k-1992, 802.11c-1998, and P802.12e)", ISO/IEC 15802-3: 1998. Stephen McCann holds a B.Sc. (Hons) degree from the University of Birmingham, England. He is currently editor of the ETSI BRAN "WLAN3G" interworking specification, having been involved in ETSI Hiperlan/2 standardisation for 3 years. He is also involved with both 802.11 work and that of the Japanese HiSWANa wireless LAN system. In the autumn of 2002, Stephen co-organised and attended the first WLAN Interworking Group (WIG) between ETSI BRAN, MMAC HSWA and IEEE 802.11. He is currently researching multimode WLAN/3G future terminals and WLAN systems for trains and ships, together with various satellite communications projects. In parallel to his Wireless LAN activities, Stephen has also been actively involved in the `rohc' working group of the IETF, looking at various Robust Header Compression schemes. Previously Stephen has been involved with avionics and was chief software integrator for the new Maastricht air traffic control system from 1995 to 1998. He is a chartered engineer and a member of the Institute of Electrical Engineers. E-mail: stephen.mccann@roke.co.uk Helena Flygare holds a M.Sc. degree in electrical engineering from Lund Institute of Technology, Sweden, where she also served as a teacher in Automatic Control for the Master Degree program. Before her present job she worked in various roles with system design for hardware and software development . In 1999 she joined Radio System Malm at Telia Research AB. She works with specification, design and integration between systems with different access technologies, e.g. WLANs, 2.5/3G, etc. from a technical, as well as from a business perspective. Since the year 2000, she has been active with WLAN interworking with 3G and other public access networks in HiperLAN/2 Global Forum, ETSI/BRAN, and 3GPP. E-mail: helena.flygare@telia.se

Hiperlan/2;interworking;3G;ETSI;BRAN;WIG;public access

102

2D Information Displays

Many exploration and manipulation tasks benefit from a coherent integration of multiple views onto complex information spaces. This paper proposes the concept of Illustrative Shadows for a tight integration of interactive 3D graphics and schematic depictions using the shadow metaphor. The shadow metaphor provides an intuitive visual link between 3D and 2D visualizations integrating the different displays into one combined information display. Users interactively explore spatial relations in realistic shaded virtual models while functional correlations and additional textual information are presented on additional projection layers using a semantic network approach. Manipulations of one visualization immediately influence the others, resulting in an in-formationally and perceptibly coherent presentation.

INTRODUCTION In many areas knowledge about structures and their meaning as well as their spatial and functional relations are required to comprehend possible effects of an intervention. For example, engineers must understand the construction of machines as a prerequisite for maintenance whereas the spatial composition of molecules and hence possible reactions are of importance for the discovering of new drugs in chemistry. Medical students need to imagine the wealth of spatial and functional correlations within the human body to master anatomy. To date, novices as well as domain experts are required to consult several, often voluminous documents in parallel to extract information for a certain intervention. Spatial relations , characteristics of structures inherently three-dimensional , such as the shape and location of structures, however , are difficult to convey on paper. Besides requiring a significant amount of images to illustrate spatial relations between only a few structures, the mental integration of multiple views to form a three-dimensional picture in mind is demanding. Spatial relations can be conveyed more ef-fectively by means of 3D models <A href="102.html#8">[18]. Using interactive 3D graphics, available to more and more people due to recent advances in consumer graphics hardware, the user may actively explore the spatial correlations of structures within a photorealistic virtual model (see upper left of <A href="102.html#1">Figure 1). Here, the visual realism of the model facilitates recognition on real encounters. Information about functional correlations, such as the interplay of muscles causing an upward motion of the human foot, has been traditionally provided by means of text and illustrations as found in textbooks. Simple, non-photorealistic drawings enriched with annotations and metagraphical symbols can be extremely powerful in conveying complex relationships and procedures (see upper right of <A href="102.html#1">Figure 1). Abstraction techniques reduce the complexity of the depicted structures to illustrate the important aspects thereby guiding the attention of the viewer to relevant details. In contrast to the visualization of spatial relations, 3D graphics add no significant value to the illustration of functional correlations. Figure 1: Illustrative Shadows provide an intuitive, visual link between spatial (3d) and non-spatial (2d) information displays integrating them into one combined information display. M. tibialis anterior M. extensor hallucis longus M. extensor digitorum longus M. tibialis posterior M. flexor digitorum longus M. flexor hallucis longus M. fibularis brevis M. fibularis longus interactive 3d-graphic 2d-information display shadow Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. IUI'03, January 1215, 2003, Miami, Florida, USA. Copyright 2003 ACM 1-58113-586-6/03/0001...$5.00. 166 The integration of both aspects in one visualization is difficult since each serves to fulfill a different goal with, partly mutually exclusive, visualization techniques. It becomes even more complicated if the 3D model is frequently manipulated , such as in construction or recent interactive learning environments. Here, occludings by annotations and metagraphical symbols are annoying and may even interrupt the current manipulation for the user. Additional views, whether as insets <A href="102.html#8">[20], separate objects like mirrors <A href="102.html#8">[8], or in form of lenses and volumetric cursors changing the rendition of embedded structures <A href="102.html#8">[7, 21, 23], are either not close enough to the manipulated structures to be fully recognized without dividing the users attention between different views <A href="102.html#8">[13] or require sometimes tedious manipulations to be placed or moved within the scene. Nonetheless , additional information as to restrictions or functional correlations pertaining the current manipulated structures is highly desired, even necessary. In this paper we present an approach called Illustrative Shadows that provides an intuitive, visual link between an actively manipulated 3D visualization and a supplemental 2D information display integrating them into one combined information display (see <A href="102.html#1">Figure 1). One of the main ideas behind Illustrative Shadows is the integration of secondary information, or in other words, background information, into an interactive 3D scene. By analyzing the users' manipulation of 3D structures and finding correlations, graphical and textual information about the current interaction context, such as graphical object-details and textual labels, are displayed in the ``background''--the shadow--to give guidance as well as to further enhance the users' understanding of the context. The paper is structured as follows: After reviewing related approaches to combine multiple visual and textual information displays, we present the design of Illustrative Shadows. Furthermore, an architecture realizing these concepts is discussed in this section. Thereafter, the major components of this architecture are described. Realization issues are subject of the subsequent section, whereas application examples and the summary conclude the paper. RELATED WORK Recently proposed tools for the exploration of virtual scenes extend possibilities to display covered structures or hidden details. A complementary view called Magic Mirror that mimics a hand mirror has been introduced in <A href="102.html#8">[8]. In addition to providing the optical effects of a real mirror, it also allows to explore the insight of objects by clipping against the mirror front-frustum. Magic lens filters as presented in <A href="102.html#8">[7, 21] go further by combining an arbitrarily-shaped region with an operator that changes the view of objects viewed through that region thereby displaying different aspects of the visualized information space. In <A href="102.html#8">[5, 23] the 2D lens approach is extended to 3D using volumetric lenses. All aforementioned techniques require the user to actively manipulate a tool within the scene. These techniques assume the user already knows which parts of the presented visualization offer additional information or is willing to explore the model. While this might be feasible in explorative environments where navigation is the main interaction task, it certainly hinders manipulation. Several approaches to combine 3D and 2D visualizations have been made using a corner cube environment ( ). The three orthogonal sides show image slices that provide a visual context for a 3D model or structures displayed in the center. In <A href="102.html#8">[11] the images have been integrated as back-planes to ground the 3D representation of anatomic structures visually. By outlining the 3D structures in the images, the spatial correspondence between the 3D renditions of activated foci in the context of human brain slices is emphasized in <A href="102.html#8">[16]. The images, however are precomputed and do not change according to the users' interaction nor is there any visualization of functional correlations. An interesting interactive approach has been proposed by <A href="102.html#8">[10]. The projection of the 3D model onto the sides of the corner cube can be manipulated by the user in order to change the position and orientation of the model. Fully rendered shadows of certain objects resemble real-world mirrors and may be used to stress importance. There is, however, no discussion on how to use this feature to provide, for instance, additional context information for the user. To establish hypotheses on the interaction context in order to be able to display additional context information and to provide meaningful descriptions of relationships knowledge modeling is required. Promising approaches to connect those knowledge with 3D graphics have been developed in the area of medical applications. The Digital Anatomist <A href="102.html#8">[4] incorporates a logic-based description comprising class and subclass relationships (is-a) as well as partitive and qualitative spatial relationships (has-parts, is-superior -to). The information is presented in tree-like textual form that can be explored by folding and unfolding. Corresponding structures are displayed in a 3D visualization aside. There is no visual integration of both information displays. The semantic network described in <A href="102.html#8">[14] is used to create various `views' in which correlating structures are displayed to communicate specific aspects with a voxel model of the human anatomy. The highly detailed visualization , however, cannot be interactively explored, nor is there any kind of abstraction to focus the users' attention. Interaction is only possible by tree-like menus. SYSTEM DESIGN USING ILLUSTRATIVE SHADOWS With the term Illustrative Shadows we refer to a coherent integration of photorealistic depictions of a virtual model with abstract illustrations and textual explanations. Both Figure 2: Architecture of a system incorporating the Illustrative Shadows approach. 3D model 3D visualization 2D visualization Visualization annotations Knowledge -based server Client interface System Event control 167 kinds of depictions serve to fulfill different and somehow contradicting goals: on the one hand to enable navigation and manipulation of complex spatial models and on the other hand to provide adjusted visualizations that guide the user's attention to additional information about the most relevant objects in the current interaction context. Both visualizations are achieved by applying photorealistic and specific non-photorealistic rendering techniques <A href="102.html#8">[22] to geometric models. Furthermore, textual information describing the most relevant structures and functional correlations between them must be integrated. The estimation of the relevance with respect to the current interaction context as well as the selection or generation of textual explanations heavily rely on non-geometric formal and informal representations and are therefore determined by external inference mechanisms. Moreover, co-referential relations between the entities within the geometric model and the formal and informal representations have to be established in order to link the different representations. Based on these requirements we designed a system architecture which comprises three basic components (see <A href="102.html#2">Figure 2): The visual component renders a photorealistic 3D model with a standard camera model as well as a non-photorealistic illustration that is projected onto a ground plane. A client interface enables external control of the non-photorealistic rendering techniques. Finally , the visual component also renders text and metagraphical annotations, such as labels, hypertext and arrows. The event control allows the user to modify the parameters of a virtual camera and to select and manipulate geometric objects within the scene. Interactions are tracked and ranked within an interaction history to communicate the current interaction context via a client interface to an external knowledge-based component. The knowledge-based server receives notification of user manipulations and establishes hypotheses on the degree of interest values (DOI) for geometric objects. These DOI values guide the selection of appropriate text fragments presented in text annotations as well as the modification of parameters of the non-photorealistic rendering techniques for emphasizing in the illustration . The following sections discuss important aspects of these system components. VISUALIZATION Besides displaying a photorealistic rendition of a 3D model that the user can manipulate, the illustration of functional correlations between structures of the model in the "background" has to be accomplished. To focus the user's attention on relevant structures and to facilitate perception, important objects must be emphasized and surroundings abstracted. Furthermore, both visualizations must be integrated in a coherent manner, so that a visual connection between the 3D and 2D renditions of the relevant objects is established by the user. Several crucial aspects have to be considered: How can objects be emphasized such that they attract the user's attention while still being in the background? Are additional graphical elements required to establish a visual correlation between the two model representations ? What illustration techniques can be applied to differentiate between important and less important objects? Is a continuous synchronization between the photorealistic and the schematic representation of the model necessary during user interactions? Integrating Different Model Representations The question coming up at this point is how a secondary, schematic model representation can be integrated such that the following requirements are fulfilled: The second representation must be placed near the original 3D representation in order to perceive structures in <A href="102.html#8">both representations [13]. The secondary representation may never occlude the central, realistic 3D visualization, which a user manipulates directly. The relevant information, however, must be visible in order to be recognized but should not distract from the interaction with the 3D model. An exact copy of the 3D model representation is not appropriate for the task of depicting functional correlations, because their illustration in the "background" requires abstraction . Without abstraction, a lot of the users' attention would be required to extract the relevant information. An exact copy, however, can be used to provide a mirrored view below the 3D model to visualize structures otherwise not visible for the user (see <A href="102.html#7">Figure 9 at the end). The integration of a secondary view as inset <A href="102.html#8">[20] has some disadvantages too. Firstly, the inset must not occlude the 3D model, thus it must be placed in distance which in turn means visual separation. Secondly, the inset framing complicates the visualization of object correlations, e.g. by lines. An ideal solution in many respects is to project the 3D model onto a plane below the model, just like casting a shadow. This 2D representation may then be modified in various ways to illustrate associated concepts and relations and therefore is called Illustrative Shadows. Besides, this approach satisfies the requirements specified above. Illustrative Shadows Cast shadows <A href="102.html#8">[3] have proven to be beneficial for perceiving spatial relationships and to facilitate object positioning Figure 3: Different types of model projections onto the illustration plane. Beside simple, monochrome projections (shadows), the color of individual structures can be preserved. The mirrored projection shows details otherwise hidden in the current view. ;; ;;; ;;; ;; ;; ;; ;; Monochrome projection Colored projection Mirrored projection 168 <A href="102.html#8">[24]. Thus, their use, if already present, occupies no additional space for the display of additional information, or in the case of prior absence, also add valuable depth cues to the 3D visualization. Furthermore, the shadows can be used to interact with the underlying information context or, as proposed in <A href="102.html#8">[10], with the shadowing 3D objects. To sum up: The shadow projection results in an abstraction which is very important for illustrations. Additional depth cues facilitate the perception of spatial relations and accelerate the relative positioning of objects while manipulating the 3D model. The projection establishes a link between a 3D object and its 2D shadow providing additional information. Displaying and Focusing in the Illustration Plane Besides simple, monochrome shadow projection further possibilities to project the 3D model representation onto the illustration plane come to mind (see <A href="102.html#3">Figure 3). Preserving the colors of the different structures of the model, for instance , enables distinct renditions and perception of the objects in the shadow. As an extension, the objects can be mirrored before projecting them. Hence, objects or hidden details of objects otherwise not visible become visible in the illustration plane. To illustrate correlations or to annotate structures, the relevant objects must be emphasized to be easily distinguished from the remaining 2D visualization. Also, the viewer must be able to differentiate between relevant and less relevant objects. For monochrome shadow projections, the object color must contrast with the shadow color. The selection of emphasizing colors should be based on a perception oriented color model, such as the HSV. Moreover, an outline can be used to attract the viewer's interest. A colored projection makes it somewhat more difficult since the variation of the color won't always result in a noticeable distinct representation . By using a conspicuous texture or shaded representations of the relevant object whereas the remaining objects are flat-shaded, the viewer's attention can be directed to those relevant objects. Significance of Accentuations In addition to objects being significant to the current interaction context supplementary objects have to be included in the illustration. These objects are not of primary importance within the concepts to be illustrated but guide the viewer and maintain context. It is important that such objects are recognized by the viewer as objects of minor significance . The relevance of objects that have been emphasized by outlining, for instance, can be judged by line width or line style (e.g. contrast, waviness). Also preserving the objects color as well as the use of texture indicates a higher importance than an interpolation of the object's color and the background color. Recognition of Correlations between both Representations An important aspect in using two different but coherent representations of the same model is the identification of correlations between those visualizations by the viewer. If an object is being emphasized in the illustration plane, it must be possible for the viewer to find its counterpart in the detailed photorealistic representation too. Often shape and color give enough hints. However, if the projection of the relevant objects results in uniform shapes, the viewer may have difficulties to recognize individual objects in the 3D visualization (see <A href="102.html#4">Figure 4). Besides accentuating those objects in the 3D representation, the integration of additional elements can be beneficial. Semi-transparent shadow volumes originally developed to facilitate object positioning in 3D <A href="102.html#8">[19] indicate direct correspondence (see Figure <A href="102.html#4">5). Integration of Annotations The conveyance of important related facts by means of graphical abstraction, accentuation, or modification of relevant structures alone is difficult. Therefore conventional book illustrations often contain annotations with short descriptions . Those annotations must be placed close to the described objects and should not occlude relevant parts of the presentation. The latter, however, cannot always be guaranteed. A simple but effective solution places the annotation on a semi-transparent background face that increases the contrast between text annotation and illustration and still does not block vision (see <A href="102.html#4">Figure 5). To further facilitate absorption of shown concepts, single words or groups of words can be emphasized and graphically linked to relating structures in the illustration. Hypertext functionality reduces the amount to which textual annotations must be displayed at once. The user can request more detailed information by activating links. Figure 4: Recognition of individual objects in different scenes. To the left, unequivocal correspondence of shape and color facilitates identification. To the right, no clear correspondence. Figure 5: A direct connection to an object's shadow is established by displaying its semi-transparent shadow volume. The integration of annotations gives meaning to unknown objects and relationships. 169 INTERACTION A human illustrator is required to identify important aspects and characteristic features of the subject or concept that is to be conveyed in order to draw a focussed visualization. One way to identify those features for the computer is to watch the user interacting with the information space. Since our goal has been to enhance the users' understanding by providing background information in the current interaction context, an Illustrative Shadow depicting correlations in that context must be generated. By navigating within the 3D model on the one hand, the user is free to explore spatial relations by changing the view. Here, single structures may be tracked by the computer regarding their visibility hence obtaining information about the users' current focus. On the other hand, the user may interact with the structures, thus expressing specific interest . As a result of the integrated 3D/2D visualization, interaction is possible within the 3D visualization as well as in the projection layers, a technique inspired by <A href="102.html#8">[10]. Using the shadow for interaction facilitates certain tasks, such as selection, since structures hidden in the 3D visualization may be visible in the projection. Furthermore, 2D input device coordinates can be mapped directly onto the plane thereby enabling the use of 2D interaction techniques. The provided manipulation tasks highly depend on the application that employs the concept of Illustrative Shadows. In our application, the user is able to compose and to decompose a given 3D model like a 3D jigsaw. Thus, translation and rotation of individual structures are main interaction tasks. User-interactions are tracked within an interaction history. By assigning relevance values to each interaction task, accumulations of these values show a distribution of interest within the model over time. Thus each single structure's degree of interest (DOI, a normalized value) is a measure for its importance to the user at a certain time. As shown in <A href="102.html#5">Table 1, touching a structure with the mouse pointer has a much lower relevance than actually selecting it. The degree of interest is communicated to the knowledge server which in turn may modify the 2D visualization of the shadow. To give an example, the user is interested in a certain structure of an anatomic 3D visualization, such as a ligament, that is part of a functional relation between a bone and a muscle. Only one of those objects, that is the bone or the muscle, should be highlighted and annotated, because of space restrictions in the shadow layer. At this point, the DOI is used to decide. If the interaction history shows more user-interest for muscles, information about the functional relation between the ligament and the muscle is displayed. KNOWLEDGE MODELING While segmentation of the 3D model into individual structures (objects) provides a spatial description, the presentation of correlations also requires a linked symbolic, textual description. Moreover, in order to establish hypotheses on the current interaction context, formal knowledge is required . Thus, the system presented in this paper comprises a knowledge base, i.e. a media-independent formal representation , media-specific realization statements of entities within the formal representation as well as a large multi-lingual text corpus. Realization statements establish co-reference relations between independent formal representations describing different aspects of the underlying information space. They also guide the generation or selection of texts used to annotate structures in the 2D visualization. The medical education application presented later in this paper is based on a knowledge base describing the objects and functional correlations of the musculo-skeletal system. It covers the area of the lower limb and the pelvic girdle. The knowledge base was created by manually analyzing several anatomy textbooks, anatomy atlases, medical dictionaries , and lexica. This analysis reveals important concepts , their hierarchical classification, and the instance attribute values forming a complex semantic network. Our system contains a hierarchical representation of basic anatomic concepts such as bones, muscles, articulations, tendons , as well as their parts and regions. The corpus contains fragments of several anatomic textbooks describing global concepts of the osteology, syndesmology, and myology as well as descriptions of all the entities of these anatomic systems within the lower leg and the pelvic girdle. In order to present appropriate system reactions the event control informs the knowledge server of user interactions. First, exploiting the visual annotations, the knowledge action parameter value relevance mouseOver time short 1 long 2 mouseButtonPressed location 3D object 4 2D object 4 annotation 6 Table 1. Relevance of certain user interactions Figure 6: Visualization of intermediate steps within a retrieval which discovers the association between the distal phalanx of the big toe and the extensor hallucis longus. has-Basis has-Area has-Basis Os-Longum Bone-Basis Musculus Bone-Area Phalanx distalis pedis Basis phalangis distalis pedis Area insertion of M. extensor hallucis longus M. extensor hallucis longus 1 2 3 1 2 3 Instance Relation Concept 170 server extracts co-referring formal entities and assigns relevance values according to <A href="102.html#5">Table 1. Subsequently, the knowledge server searches for associations between the most relevant entities. Our system pursues two alternative strategies: retrievals and suggestions. Retrievals discover relations between entities by tracking predefined paths within the knowledge base. <A href="102.html#5">Figure 6 illustrates the intermediate steps in order to extract relations between bones and muscles. From a functional point of view (i.e. the muscle mechanics), bones are insertions or origins of muscles. Its contraction produces force, which in turn changes the orientation of these bones. These retrievals also need to consider substructures (e.g. bone-volumes and bone-area). The following retrieval extracts those muscles, which originate in a given bone (first logic order): The has-Part* relation represents the transitive hull over several spatial part-of-relations <A href="102.html#7">[1] (e.g. the has-Basis relation between Bones and Bone-Volumes and the has-Area relation between Areas or Volumes and Areas). These retrievals rely on knowledge about the structure of the knowledge base. Moreover, they refer to a small number of relevant objects. In many situations, however, the event control comes up with a huge number of potential relevant objects, which cannot easily be mapped to a predefined query. Hence, we adopt a bottom-up search approach within a complex semantic network developed within cognitive psychology. In his model of human comprehension <A href="102.html#8">[15] Quillian assumed that spatially and temporally independent aspects of human long-term memory are organized in a semantic network . Furthermore, he assumed that cognitive processes that access a node of the semantic network activate all connected nodes in parallel. The term spreading activation refers to a recursive propagation of initial stimuli. Nowadays, this term subsumes breadth-first search algorithms for paths connecting the nodes of a start and a destination set in directed graphs satisfying an evaluation criterion. Collins and Loftus <A href="102.html#8">[6] modify the propagation algorithm to consider activation strength. In our system, the knowledge server uses the objects' DOI from the event control as an initial activation, which spreads through the semantic network. These initial activa-tions also take the content presented on textual labels and inspected by the user into account. The spreading activation approach generates a focus structure which contains information how dominant graphical objects must be presented in the schematic illustration <A href="102.html#8">[9]. <A href="102.html#6">Figure 7 illustrates how visual dominance values control the render parameters. REALIZATION DETAILS The visual component of the prototypical implementation extends the Open Inventor graphical library with powerful scenegraph nodes to display hypertext on overlay regions and to render semi-transparent shadow volumes. Other nodes encapsulate the mirror and shading projection onto the ground plane as well as the user interaction, and emphasize techniques (e.g. computation of silhouette lines). The layer management (see <A href="102.html#6">Figure 8) employs the OpenGL polygon offset feature to allow graphics to overlap specifically whereas visibility-tests of individual structures are accomplished by offscreen rendering and analyzing OpenGL p-buffers. Additional OSF Motif widgets enable the user to add personalized annotations, which are inserted into the knowledge base. The knowledge base encodes both the media-independent formal knowledge representation as well as media specific realization statements using XML topic maps <A href="102.html#7">[2]. To process this information, XML statements are transformed into LISP-code. The authoring system contains export filters for the NeoClassic and the LOOM <A href="102.html#8">[12] description logic inference machine. In the current version it covers about 50 basic anatomic concepts, 70 relations, and over 1500 instances, with linguistic realization statements in Latin, German and English. Furthermore, visual annotations refer to a small number of geometric models and 2D illustrations. The interface between the knowledge server and the visual component is described using CORBA's interface definition language (IDL). The CORBA-based interface implementation enables us to experiment with several knowledge serv-Figure 7: Application of different emphasize techniques to the 2D information representation according to decreasing dominance values. { muscle | bone: Bone(bone) Musculus(muscle) ( is-Origin-of (bone, muscle) part: has-Part*(bone, part) is-Origin-of (part, muscle)) } Figure 8: Multiple layers are used to place the different visual information in order. Thereby, occluded details may be visible in the shadow (3). Outline Highlight Ordinary Detail Annotation 1 3 2 171 ers implemented in Common-LISP (LOOM) and C++ (NeoClassic). APPLICATIONS We applied the concept of Illustrative Shadows to an application of medical education. This system had been previously designed to foster the understanding of spatial relationships by means of 3D models based on a virtual 3D jigsaw approach <A href="102.html#8">[19]. While composing anatomical structures has proven to help medical students to build an understanding of the spatial composition <A href="102.html#8">[17], most of the users expressed their desire for detailed information about functional relations between structures. Consequently, students would be able to playfully study human organs including their spatial and functional correlations. Figures <A href="102.html#7">9 and <A href="102.html#7">10 depict the screen of the prototype in typical learning sessions . Individual objects can be detached and moved within the scene to expose occluded structures. In <A href="102.html#7">Figure 9, one of the alternative visualization, the mirror, is shown to demonstrate the various employments of the plane. It enables the user to simultaneously look at two different views of the 3D model. Graphical accentuations are used to attract the viewer's attention (left atrium). Additional textual annotations are assigned to correlating structures in a hypertext representation which can be explored by following sepa-rately marked links. Another application that has not yet been realized is to support users of CAD systems by Illustrative Shadows. CAD systems are not only used to design individual components in 3D but also to assemble complex systems. Information about parameters of single components and relationships are of major importance. Being able to retrieve this information directly from the scene while interacting with the components is of great benefit for the design engineer. CONCLUSIONS For educational, engineering, or maintenance purposes a wealth of information about spatial and functional correlation as well as textual information is required. In this paper we developed a new metaphor-based approach to coher-ently integrate different views onto such a complex information space within an interactive system. Illustrative Shadows provide an intuitive visual link and a level of integration between interactive 3D graphics and supplemental 2D information displays that is hard to achieve with other concepts. Shadow projections have proven to be beneficial for perceiving spatial relationships and to facilitate object positioning . Thus, their use, if already present, occupies no additional space for the display of additional information, or in the case of prior absence, also add valuable depth cues to the 3D visualization. The shadow projection onto a flat plane enables schematic illustrations which are focused on specific information extraction tasks and facilitates the integration of generated textual information that leads to further meaning. Thus, Illustrative Shadows promote the comprehension of complex spatial relations and functional correlations. Furthermore, the secondary information display does not hinder manipulations of the 3D model. Our approach is well suited for compact 3D models, and has been successfully applied to an application of medical education REFERENCES 1. Bernauer, J. Analysis of Part-Whole Relation and Subsumption in the Medical Domain. Data & Knowledge Engineering, 20(3):405415, October 1996. 2. Biezunski, M., Bryan, M., and Newcomb, S., editors. ISO/IEC 13250:2000 Topic Maps: Information Technology Document Description and Markup Language . International Organization for Standarization Figure 9: Alternative visualization. The Mirror facilitates manipulation in 3D, by providing a better view of the structures. The central representation is never occluded by annotations. Figure 10: German annotation of objects previously selected. Realization statements of the semantic network provide alternative German, English, and Latin phrases referring to formal entities. 172 (ISO) and International Electrotechnical Commission (IEC), December 1999. 1. Draft. 3. Blinn, J.F. Me and my (fake) shadow. IEEE Computer Graphics & Applications, 8(1):8286, January/Febru-ary 1988. 4. Brinkley, J.F., Wong, B.A., Hinshaw, K.P., and Rosse, C. Design of an Anatomy Information System. IEEE Computer Graphics & Applications, 19(3):3848, May/June 1999. 5. Cignoni, P., Montani, C., and Scopigno, R. Magic-sphere : An insight tool for 3d data visualization. IEEE Computer Graphics Forum, 13(3):317328, 1994. 6. Collins, A. and Loftus, E. A Spreading-Activation Theory of Semantic Processing. Psychological Review, 82(6):407428, 1975. 7. Fishkin, K. and Stone, M.C. Enhanced dynamic queries via moveable filters. In Katz, I.R., Mack, R., Marks, L., Rosson, M.B., and Nielsen, J., editors, Proc. of ACM CHI Conference on Human Factors in Computing Systems (Denver, May 1995), pages 415420. ACM Press, New York, 1995. 8. Grosjean, J. and Coquillart, S. The magic mirror: A metaphor for assisting the exploration of virtual worlds. In Zara, J., editor, Proc. of Spring Conference on Computer Graphics (Budmerice, Slovakia, April 1999), pages 125129, 1999. 9. Hartmann, K., Schlechtweg, S., Helbing, R., and Strothotte, T. Knowledge-Supported Graphical Illustration of Texts. In De Marsico, M., Levialdi, S., and Panizzi, E., editors, Proc. of the Working Conference on Advanced Visual Interfaces (AVI 2002), pages 300307, Trento, Italy, May 2002. ACM Press, New York. 10. Herndon, K.P., Zeleznik, R.C., Robbins, D.C., Conner, D.B., Snibbe, S.S., and van Dam, A. Interactive shadows . In Proc. of ACM Symposium on User Interface and Software Technology (Monterey, November 1992), pages 16. ACM Press, New York, 1992. 11. Hhne, K.H., Pflesser, B., Pommert, A., Riemer, M., Schiemann, T., Schubert, R., and Tiede, U. A virtual body model for surgical education and rehearsal. IEEE Computer, 29(1):2531, January 1996. 12. MacGregor, R. A Description Classifier for the Predicate Calculus. In Hayes-Roth, B. and Korf, R., editors, Proc. of the Twelfth Annual National Conference on Artificial Intelligence (AAAI-94), pages 213220, Seattle , Washington, August 1994. AAAI Press, Menlo Park. 13. Moreno, R. and Mayer, R.E. Cognitive principles of multimedia learning. Journal of Educational Psychology , 91:358368, 1999. 14. Pommert, A., Hhne, K.H., Pflesser, B., Richter, E., Riemer, M., Schiemann, T., Schubert, R., Schumacher, U., and Tiede, U. Creating a high-resolution spatial/ symbolic model of the inner organs based on the visible human. Medical Image Analysis, 5(3):221228, 2001. 15. Quillian, M. Semantic Memory. In Minsky, M., editor, Semantic Information Processing, chapter 4, pages 227270. MIT Press, Cambridge., 1968. 16. Rehm, K., Lakshminaryan, K., Frutiger, S., Schaper, K.A., Sumners, D.W., Strother, S.C., Anderson, J.R., and Rottenberg, D.A. A symbolic environment for visualizing activated foci in functional neuroimaging datasets. Medical Image Analysis, 2(3):215226, ??? 1998. 17. Ritter, F., Berendt, B., Fischer, B., Richter, R., and Preim, B. Virtual 3d jigsaw puzzles: Studying the effect of exploring spatial relations with implicit guidance. In Herczeg, M., Prinz, W., and Oberquelle, H., editors, Proc. of Mensch & Computer (Hamburg, September 2002), pages 363372, Stuttgart Leipzig Wiesbaden, 2002. B.G.Teubner. 18. Ritter, F., Deussen, O., Preim, B., and Strothotte, T. Virtual 3d puzzles: A new method for exploring geometric models in vr. IEEE Computer Graphics & Applications, 21(5):1113, September/October 2001. 19. Ritter, F., Preim, B., Deussen, O., and Strothotte, T. Using a 3d puzzle as a metaphor for learning spatial relations. In Fels, S.S. and Poulin, P., editors, Proc. of Graphics Interface (Montral, May 2000), pages 171 178. Morgan Kaufmann Publishers, San Francisco, 2000. 20. Seligmann, D.D. and Feiner, S. Automated generation of intent-based 3d illustrations. In Proc. of ACM SIG-GRAPH Conference on Computer Graphics (Las Vegas, July 1991), pages 123132. ACM Press, New York, 1991. 21. Stone, M.C., Fishkin, K., and Bier, E.A. The moveable filter as a user interface tool. In Plaisant, C., editor, Proc. of ACM CHI Conference on Human Factors in Computing Systems (Boston, April 1994), pages 306 312. ACM Press, New York, 1994. 22. Strothotte, T. and Schlechtweg, S. Non-Photorealistic Computer Graphics: Modeling, Rendering, and Animation . Morgan Kaufmann Publishers, San Francisco, 2002. 23. Viega, J., Conway, M.J., Williams, G., and Pausch, R. 3d magic lenses. In Proc. of ACM Symposium on User Interface and Software Technology (Seattle, November 1996), pages 5158. ACM Press, New York, 1996. 24. Wanger, L.R. The effect of shadow quality on the perception of spatial relationships in computer generated imagery. In Proc. of Symposium on Interactive 3D Graphics (Cambridge, March 1992), pages 3942. ACM Press, New York, 1992. See also: http://isgwww.cs.uni-magdeburg.de/research/is/ 173

Information visualization;Spreading activation

103

Impedance Coupling in Content-targeted Advertising

The current boom of the Web is associated with the revenues originated from on-line advertising. While search-based advertising is dominant, the association of ads with a Web page (during user navigation) is becoming increasingly important . In this work, we study the problem of associating ads with a Web page, referred to as content-targeted advertising , from a computer science perspective. We assume that we have access to the text of the Web page, the keywords declared by an advertiser, and a text associated with the advertiser's business. Using no other information and operating in fully automatic fashion, we propose ten strategies for solving the problem and evaluate their effectiveness. Our methods indicate that a matching strategy that takes into account the semantics of the problem (referred to as AAK for "ads and keywords") can yield gains in average precision figures of 60% compared to a trivial vector-based strategy. Further, a more sophisticated impedance coupling strategy, which expands the text of the Web page to reduce vocabulary impedance with regard to an advertisement, can yield extra gains in average precision of 50%. These are first results . They suggest that great accuracy in content-targeted advertising can be attained with appropriate algorithms.

INTRODUCTION The emergence of the Internet has opened up new marketing opportunities. In fact, a company has now the possibility of showing its advertisements (ads) to millions of people at a low cost. During the 90's, many companies invested heavily on advertising in the Internet with apparently no concerns about their investment return [16]. This situation radically changed in the following decade when the failure of many Web companies led to a dropping in supply of cheap venture capital and a considerable reduction in on-line advertising investments [15, 16]. It was clear then that more effective strategies for on-line advertising were required. For that, it was necessary to take into account short-term and long-term interests of the users related to their information needs [9, 14]. As a consequence, many companies intensified the adoption of intrusive techniques for gathering information of users mostly without their consent [8]. This raised privacy issues which stimu-lated the research for less invasive measures [16]. More recently, Internet information gatekeepers as, for example , search engines, recommender systems, and comparison shopping services, have employed what is called paid placement strategies [3]. In such methods, an advertiser company is given prominent positioning in advertisement lists in return for a placement fee. Amongst these methods, the most popular one is a non-intrusive technique called keyword targeted marketing [16]. In this technique, keywords extracted from the user's search query are matched against keywords associated with ads provided by advertisers. A ranking of the ads, which also takes into consideration the amount that each advertiser is willing to pay, is computed. The top ranked ads are displayed in the search result page together with the answers for the user query. The success of keyword targeted marketing has motivated information gatekeepers to offer their advertisement services in different contexts. For example, as shown in Figure 1, relevant ads could be shown to users directly in the pages of information portals. The motivation is to take advantage of 496 the users immediate information interests at browsing time. The problem of matching ads to a Web page that is browsed, which we also refer to as content-targeted advertising [1], is different from that of keyword marketing. In this case, instead of dealing with users' keywords, we have to use the contents of a Web page to decide which ads to display. Figure 1: Example of content-based advertising in the page of a newspaper. The middle slice of the page shows the beginning of an article about the launch of a DVD movie. At the bottom slice, we can see advertisements picked for this page by Google's content-based advertising system, AdSense. It is important to notice that paid placement advertising strategies imply some risks to information gatekeepers. For instance, there is the possibility of a negative impact on their credibility which, at long term, can demise their market share [3]. This makes investments in the quality of ad recommendation systems even more important to minimize the possibility of exhibiting ads unrelated to the user's interests. By investing in their ad systems, information gatekeepers are investing in the maintenance of their credibility and in the reinforcement of a positive user attitude towards the advertisers and their ads [14]. Further, that can translate into higher clickthrough rates that lead to an increase in revenues for information gatekeepers and advertisers, with gains to all parts [3]. In this work, we focus on the problem of content-targeted advertising. We propose new strategies for associating ads with a Web page. Five of these strategies are referred to as matching strategies. They are based on the idea of matching the text of the Web page directly to the text of the ads and its associated keywords. Five other strategies, which we here introduce, are referred to as impedance coupling strategies. They are based on the idea of expanding the Web page with new terms to facilitate the task of matching ads and Web pages. This is motivated by the observation that there is frequently a mismatch between the vocabulary of a Web page and the vocabulary of an advertisement. We say that there is a vocabulary impedance problem and that our technique provides a positive effect of impedance coupling by reducing the vocabulary impedance. Further, all our strategies rely on information that is already available to information gatekeepers that operate keyword targeted advertising systems. Thus, no other data from the advertiser is required. Using a sample of a real case database with over 93,000 ads and 100 Web pages selected for testing, we evaluate our ad recommendation strategies. First, we evaluate the five matching strategies. They match ads to a Web page using a standard vector model and provide what we may call trivial solutions. Our results indicate that a strategy that matches the ad plus its keywords to a Web page, requiring the keywords to appear in the Web page, provides improvements in average precision figures of roughly 60% relative to a strategy that simply matches the ads to the Web page. Such strategy, which we call AAK (for "ads and keywords"), is then taken as our baseline. Following we evaluate the five impedance coupling strategies . They are based on the idea of expanding the ad and the Web page with new terms to reduce the vocabulary impedance between their texts. Our results indicate that it is possible to generate extra improvements in average precision figures of roughly 50% relative to the AAK strategy. The paper is organized as follows. In section 2, we introduce five matching strategies to solve content-targeted advertising. In section 3, we present our impedance coupling strategies. In section 4, we describe our experimental methodology and datasets and discuss our results. In section 5 we discuss related work. In section 6 we present our conclusions. MATCHING STRATEGIES Keyword advertising relies on matching search queries to ads and its associated keywords. Context-based advertising , which we address here, relies on matching ads and its associated keywords to the text of a Web page. Given a certain Web page p, which we call triggering page, our task is to select advertisements related to the contents of p. Without loss of generality, we consider that an advertisement a i is composed of a title, a textual description, and a hyperlink. To illustrate, for the first ad by Google shown in Figure 1, the title is "Star Wars Trilogy Full", the description is "Get this popular DVD free. Free w/ free shopping. Sign up now", and the hyperlink points to the site "www.freegiftworld.com". Advertisements can be grouped by advertisers in groups called campaigns, such that a campaign can have one or more advertisements. Given our triggering page p and a set A of ads, a simple way of ranking a i A with regard to p is by matching the contents of p to the contents of a i . For this, we use the vector space model [2], as discussed in the immediately following. In the vector space model, queries and documents are represented as weighted vectors in an n-dimensional space. Let w iq be the weight associated with term t i in the query q and w ij be the weight associated with term t i in the document d j . Then, q = (w 1q , w 2q , ..., w iq , ..., w nq ) and d j = (w 1j , w 2j , ..., w ij , ..., w nj ) are the weighted vectors used to represent the query q and the document d j . These weights can be computed using classic tf-idf schemes. In such schemes, weights are taken as the product between factors that quantify the importance of a term in a document (given by the term frequency, or tf, factor) and its rarity in the whole collection (given by the inverse document factor, or idf, factor), see [2] for details. The ranking of the query q with regard to the document d j is computed by the cosine similarity 497 formula, that is, the cosine of the angle between the two corresponding vectors: sim(q, d j ) = q d j |q| |d j | = P n i=1 w iq w ij qP n i=1 w 2 iq qP n i=1 w 2 ij (1) By considering p as the query and a i as the document, we can rank the ads with regard to the Web page p. This is our first matching strategy. It is represented by the function AD given by: AD(p, a i ) = sim(p, a i ) where AD stands for "direct match of the ad, composed by title and description" and sim(p, a i ) is computed according to Eq. (1). In our second method, we use other source of evidence provided by the advertisers: the keywords. With each advertisement a i an advertiser associates a keyword k i , which may be composed of one or more terms. We denote the association between an advertisement a i and a keyword k i as the pair (a i , k i ) K, where K is the set of associations made by the advertisers. In the case of keyword targeted advertising, such keywords are used to match the ads to the user queries. In here, we use them to match ads to the Web page p. This provides our second method for ad matching given by: KW(p, a i ) = sim(p, k i ) where (a i , k i ) K and KW stands for "match the ad keywords" . We notice that most of the keywords selected by advertisers are also present in the ads associated with those keywords . For instance, in our advertisement test collection, this is true for 90% of the ads. Thus, instead of using the keywords as matching devices, we can use them to emphasize the main concepts in an ad, in an attempt to improve our AD strategy. This leads to our third method of ad matching given by: AD KW(p, a i ) = sim(p, a i k i ) where (a i , k i ) K and AD KW stands for "match the ad and its keywords". Finally, it is important to notice that the keyword k i associated with a i could not appear at all in the triggering page p, even when a i is highly ranked. However, if we assume that k i summarizes the main topic of a i according to an advertiser viewpoint, it can be interesting to assure its presence in p. This reasoning suggests that requiring the occurrence of the keyword k i in the triggering page p as a condition to associate a i with p might lead to improved results. This leads to two extra matching strategies as follows: ANDKW(p, a i ) = sim(p, a i ) if k i p 0 if otherwise AD ANDKW(p, a i ) = AAK(p, a i ) = sim(p, a i k i ) if k i p 0 if otherwise where (a i , k i ) K, ANDKW stands for "match the ad keywords and force their appearance", and AD ANDKW (or AAK for "ads and keywords") stands for "match the ad, its keywords, and force their appearance". As we will see in our results, the best among these simple methods is AAK. Thus, it will be used as baseline for our impedance coupling strategies which we now discuss. IMPEDANCE COUPLING STRATEGIES Two key issues become clear as one plays with the content-targeted advertising problem. First, the triggering page normally belongs to a broader contextual scope than that of the advertisements. Second, the association between a good advertisement and the triggering page might depend on a topic that is not mentioned explicitly in the triggering page. The first issue is due to the fact that Web pages can be about any subject and that advertisements are concise in nature. That is, ads tend to be more topic restricted than Web pages. The second issue is related to the fact that, as we later discuss, most advertisers place a small number of advertisements. As a result, we have few terms describing their interest areas. Consequently, these terms tend to be of a more general nature. For instance, a car shop probably would prefer to use "car" instead of "super sport" to describe its core business topic. As a consequence, many specific terms that appear in the triggering page find no match in the advertisements. To make matters worst, a page might refer to an entity or subject of the world through a label that is distinct from the label selected by an advertiser to refer to the same entity. A consequence of these two issues is that vocabularies of pages and ads have low intersection, even when an ad is related to a page. We cite this problem from now on as the vocabulary impedance problem. In our experiments, we realized that this problem limits the final quality of direct matching strategies. Therefore, we studied alternatives to reduce the referred vocabulary impedance. For this, we propose to expand the triggering pages with new terms. Figure 2 illustrates our intuition. We already know that the addition of keywords (selected by the advertiser ) to the ads leads to improved results. We say that a keyword reduces the vocabulary impedance by providing an alternative matching path. Our idea is to add new terms (words) to the Web page p to also reduce the vocabulary impedance by providing a second alternative matching path. We refer to our expansion technique as impedance coupling. For this, we proceed as follows. expansion terms keyword vocabulary impedance triggering page p ad Figure 2: Addition of new terms to a Web page to reduce the vocabulary impedance. An advertiser trying to describe a certain topic in a concise way probably will choose general terms to characterize that topic. To facilitate the matching between this ad and our triggering page p, we need to associate new general terms with p. For this, we assume that Web documents similar to the triggering page p share common topics. Therefore, 498 by inspecting the vocabulary of these similar documents we might find good terms for better characterizing the main topics in the page p. We now describe this idea using a Bayesian network model [10, 11, 13] depicted in Figure 3. R D 0 D 1 D j D k T 1 T 2 T 3 T i T m ... ... ... ... Figure 3: Bayesian network model for our impedance coupling technique. In our model, which is based on the belief network in [11], the nodes represent pieces of information in the domain. With each node is associated a binary random variable, which takes the value 1 to mean that the corresponding entity (a page or terms) is observed and, thus, relevant in our computations. In this case, we say that the information was observed. Node R represents the page r, a new representation for the triggering page p. Let N be the set of the k most similar documents to the triggering page, including the triggering page p itself, in a large enough Web collection C. Root nodes D 0 through D k represent the documents in N , that is, the triggering page D 0 and its k nearest neighbors, D 1 through D k , among all pages in C. There is an edge from node D j to node R if document d j is in N . Nodes T 1 through T m represent the terms in the vocabulary of C. There is an edge from node D j to a node T i if term t i occurs in document d j . In our model, the observation of the pages in N leads to the observation of a new representation of the triggering page p and to a set of terms describing the main topics associated with p and its neighbors. Given these definitions, we can now use the network to determine the probability that a term t i is a good term for representing a topic of the triggering page p. In other words, we are interested in the probability of observing the final evidence regarding a term t i , given that the new representation of the page p has been observed, P (T i = 1|R = 1). This translates into the following equation 1 : P (T i |R) = 1 P (R) X d P (T i |d)P (R|d)P (d) (2) where d represents the set of states of the document nodes. Since we are interested just in the states in which only a single document d j is observed and P (d) can be regarded as a constant, we can rewrite Eq. (2) as: P (T i |R) = P (R) k X j=0 P (T i |d j )P (R|d j ) (3) where d j represents the state of the document nodes in which only document d j is observed and is a constant 1 To simplify our notation we represent the probabilities P (X = 1) as P (X) and P (X = 0) as P (X). associated with P (d j ). Eq. (3) is the general equation to compute the probability that a term t i is related to the triggering page. We now define the probabilities P (T i |d j ) and P (R|d j ) as follows: P (T i |d j ) = w ij (4) P (R|d j ) = (1 - ) j = 0 sim(r, d j ) 1 j k (5) where is a normalizing constant, w ij is the weight associated with term t i in the document d j , and sim(p, d j ) is given by Eq. (1), i.e., is the cosine similarity between p and d j . The weight w ij is computed using a classic tf-idf scheme and is zero if term t i does not occur in document d j . Notice that P (T i |d j ) = 1 - P (T i |d j ) and P (R|d j ) = 1 - P (R|d j ). By defining the constant , it is possible to determine how important should be the influence of the triggering page p to its new representation r. By substituting Eq. (4) and Eq. (5) into Eq. (3), we obtain: P (T i |R) = ((1 - ) w i0 + k X j=1 w ij sim(r, d j )) (6) where = is a normalizing constant. We use Eq. (6) to determine the set of terms that will compose r, as illustrated in Figure 2. Let t top be the top ranked term according to Eq. (6). The set r is composed of the terms t i such that P (T i |R) P (T top |R) , where is a given threshold. In our experiments, we have used = 0.05. Notice that the set r might contain terms that already occur in p. That is, while we will refer to the set r as expansion terms, it should be clear that p r = . By using = 0, we simply consider the terms originally in page p. By increasing , we relax the context of the page p, adding terms from neighbor pages, turning page p into its new representation r. This is important because, sometimes, a topic apparently not important in the triggering page offers a good opportunity for advertising. For example, consider a triggering page that describes a congress in London about digital photography. Although London is probably not an important topic in this page, advertisements about hotels in London would be appropriate. Thus, adding "hotels" to page p is important. This suggests using &gt; 0, that is, preserving the contents of p and using the terms in r to expand p. In this paper, we examine both approaches. Thus, in our sixth method we match r, the set of new expansion terms, directly to the ads, as follows: AAK T(p, a i ) = AAK(r, a i ) where AAK T stands for "match the ad and keywords to the set r of expansion terms". In our seventh method, we match an expanded page p to the ads as follows: AAK EXP(p, a i ) = AAK(p r, a i ) where AAK EXP stands for "match the ad and keywords to the expanded triggering page". 499 To improve our ad placement methods, other external source that we can use is the content of the page h pointed to by the advertisement's hyperlink, that is, its landing page. After all, this page comprises the real target of the ad and perhaps could present a more detailed description of the product or service being advertised. Given that the advertisement a i points to the landing page h i , we denote this association as the pair (a i , h i ) H, where H is the set of associations between the ads and the pages they point to. Our eighth method consists of matching the triggering page p to the landing pages pointed to by the advertisements, as follows: H(p, a i ) = sim(p, h i ) where (a i , h i ) H and H stands for "match the hyperlink pointed to by the ad". We can also combine this information with the more promising methods previously described, AAK and AAK EXP as follows . Given that (a i , h i ) H and (a i , k i ) K, we have our last two methods: AAK H(p, a i ) = sim(p, a i h i k i ) if k i p 0 if otherwise AAK EXP H(p, a i ) = sim(p r, a i h i k i ) if k i (p r) 0 if otherwise where AAK H stands for "match ads and keywords also considering the page pointed by the ad" and AAH EXP H stands for "match ads and keywords with expanded triggering page, also considering the page pointed by the ad". Notice that other combinations were not considered in this study due to space restrictions. These other combinations led to poor results in our experimentation and for this reason were discarded. EXPERIMENTS To evaluate our ad placement strategies, we performed a series of experiments using a sample of a real case ad collection with 93,972 advertisements, 1,744 advertisers, and 68,238 keywords 2 . The advertisements are grouped in 2,029 campaigns with an average of 1.16 campaigns per advertiser. For the strategies AAK T and AAK EXP, we had to generate a set of expansion terms. For that, we used a database of Web pages crawled by the TodoBR search engine [12] (http://www.todobr.com.br/). This database is composed of 5,939,061 pages of the Brazilian Web, under the domain ".br". For the strategies H, AAK H, and AAK EXP H, we also crawled the pages pointed to by the advertisers. No other filtering method was applied to these pages besides the removal of HTML tags. Since we are initially interested in the placement of advertisements in the pages of information portals, our test collection was composed of 100 pages extracted from a Brazilian newspaper. These are our triggering pages. They were crawled in such a way that only the contents of their articles was preserved. As we have no preferences for particular 2 Data in portuguese provided by an on-line advertisement company that operates in Brazil. topics, the crawled pages cover topics as diverse as politics, economy, sports, and culture. For each of our 100 triggering pages, we selected the top three ranked ads provided by each of our 10 ad placement strategies. Thus, for each triggering page we select no more than 30 ads. These top ads were then inserted in a pool for that triggering page. Each pool contained an average of 15.81 advertisements. All advertisements in each pool were submitted to a manual evaluation by a group of 15 users. The average number of relevant advertisements per page pool was 5.15. Notice that we adopted the same pooling method used to evaluate the TREC Web-based collection [6]. To quantify the precision of our results, we used 11-point average figures [2]. Since we are not able to evaluate the entire ad collection, recall values are relative to the set of evaluated advertisements. 4.2 Tuning Idf factors We start by analyzing the impact of different idf factors in our advertisement collection. Idf factors are important because they quantify how discriminative is a term in the collection. In our ad collection, idf factors can be computed by taking ads, advertisers or campaigns as documents. To exemplify, consider the computation of "ad idf" for a term t i that occurs 9 times in a collection of 100 ads. Then, the inverse document frequency of t i is given by: idf i = log 100 9 Hence, we can compute ad, advertiser or campaign idf factors . As we observe in Figure 4, for the AD strategy, the best ranking is obtained by the use of campaign idf, that is, by calculating our idf factor so that it discriminates campaigns. Similar results were obtained for all the other methods. 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0 0.2 0.4 0.6 0.8 1 precision recall Campaign idf Advertiser idf Ad idf Figure 4: Precision-recall curves obtained for the AD strategy using ad, advertiser, and campaign idf factors. This reflects the fact that terms might be better discriminators for a business topic than for an specific ad. This effect can be accomplished by calculating the factor relative to idf advertisers or campaigns instead of ads. In fact, campaign idf factors yielded the best results. Thus, they will be used in all the experiments reported from now on. 500 4.3 Results Matching Strategies Figure 5 displays the results for the matching strategies presented in Section 2. As shown, directly matching the contents of the ad to the triggering page (AD strategy) is not so effective. The reason is that the ad contents are very noisy. It may contain messages that do not properly describe the ad topics such as requisitions for user actions (e.g, "visit our site") and general sentences that could be applied to any product or service (e.g, "we delivery for the whole country" ). On the other hand, an advertiser provided keyword summarizes well the topic of the ad. As a consequence, the KW strategy is superior to the AD and AD KW strategies. This situation changes when we require the keywords to appear in the target Web page. By filtering out ads whose keywords do not occur in the triggering page, much noise is discarded. This makes ANDKW a better alternative than KW. Further, in this new situation, the contents of the ad becomes useful to rank the most relevant ads making AD ANDKW (or AAK for "ads and keywords") the best among all described methods. For this reason, we adopt AAK as our baseline in the next set of experiments. 0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.2 0.4 0.6 0.8 1 precision recall AAK ANDKW KW AD_KW AD Figure 5: Comparison among our five matching strategies. AAK ("ads and keywords") is superior. Table 1 illustrates average precision figures for Figure 5. We also present actual hits per advertisement slot. We call "hit" an assignment of an ad (to the triggering page) that was considered relevant by the evaluators. We notice that our AAK strategy provides a gain in average precision of 60% relative to the trivial AD strategy. This shows that careful consideration of the evidence related to the problem does pay off. Impedance Coupling Strategies Table 2 shows top ranked terms that occur in a page covering Argentinean wines produced using grapes derived from the Bordeaux region of France. The p column includes the top terms for this page ranked according to our tf-idf weighting scheme. The r column includes the top ranked expansion terms generated according to Eq. (6). Notice that the expansion terms not only emphasize important terms of the target page (by increasing their weights) such as "wines" and Methods Hits 11-pt average #1 #2 #3 total score gain(%) AD 41 32 13 86 0.104 AD KW 51 28 17 96 0.106 +1.9 KW 46 34 28 108 0.125 +20.2 ANDKW 49 37 35 121 0.153 +47.1 AD ANDKW (AAK) 51 48 39 138 0.168 +61.5 Table 1: Average precision figures, corresponding to Figure 5, for our five matching strategies. Columns labelled #1, #2, and #3 indicate total of hits in first, second, and third advertisement slots, respectively . The AAK strategy provides improvements of 60% relative to the AD strategy. Rank p r term score term score 1 argentina 0.090 wines 0.251 2 obtained* 0.047 wine* 0.140 3 class* 0.036 whites 0.091 4 whites 0.035 red* 0.057 5 french* 0.031 grape 0.051 6 origin* 0.029 bordeaux 0.045 7 france* 0.029 acideness* 0.038 8 grape 0.017 argentina 0.037 9 sweet* 0.016 aroma* 0.037 10 country* 0.013 blanc* 0.036 ... 35 wines 0.010 ... Table 2: Top ranked terms for the triggering page p according to our tf-idf weighting scheme and top ranked terms for r, the expansion terms for p, generated according to Eq. (6). Ranking scores were normalized in order to sum up to 1. Terms marked with `*' are not shared by the sets p and r. "whites", but also reveal new terms related to the main topic of the page such as "aroma" and "red". Further, they avoid some uninteresting terms such as "obtained" and "country". Figure 6 illustrates our results when the set r of expansion terms is used. They show that matching the ads to the terms in the set r instead of to the triggering page p (AAK T strategy) leads to a considerable improvement over our baseline, AAK. The gain is even larger when we use the terms in r to expand the triggering page (AAK EXP method). This confirms our hypothesis that the triggering page could have some interesting terms that should not be completely discarded. Finally, we analyze the impact on the ranking of using the contents of pages pointed by the ads. Figure 7 displays our results. It is clear that using only the contents of the pages pointed by the ads (H strategy) yields very poor results. However, combining evidence from the pages pointed by the ads with our baseline yields improved results. Most important , combining our best strategy so far (AAK EXP) with pages pointed by ads (AAK EXP H strategy) leads to superior results. This happens because the two additional sources of evidence, expansion terms and pages pointed by the ads, are distinct and complementary, providing extra and valuable information for matching ads to a Web page. 501 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.2 0.4 0.6 0.8 1 precision recall AAK_EXP AAK_T AAK Figure 6: Impact of using a new representation for the triggering page, one that includes expansion terms. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.2 0.4 0.6 0.8 1 precision recall AAK_EXP_H AAK_H AAK H Figure 7: Impact of using the contents of the page pointed by the ad (the hyperlink). Figure 8 and Table 3 summarize all results described in this section. In Figure 8 we show precision-recall curves and in Table 3 we show 11-point average figures. We also present actual hits per advertisement slot and gains in average precision relative to our baseline, AAK. We notice that the highest number of hits in the first slot was generated by the method AAK EXP. However, the method with best overall retrieval performance was AAK EXP H, yielding a gain in average precision figures of roughly 50% over the baseline (AAK). 4.4 Performance Issues In a keyword targeted advertising system, ads are assigned at query time, thus the performance of the system is a very important issue. In content-targeted advertising systems, we can associate ads with a page at publishing (or updating ) time. Also, if a new ad comes in we might consider assigning this ad to already published pages in offline mode. That is, we might design the system such that its performance depends fundamentally on the rate that new pages 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.2 0.4 0.6 0.8 1 precision recall AAK_EXP_H AAK_EXP AAK_T AAK_H AAK H Figure 8: Comparison among our ad placement strategies. Methods Hits 11-pt average #1 #2 #3 total score gain(%) H 28 5 6 39 0.026 -84.3 AAK 51 48 39 138 0.168 AAK H 52 50 46 148 0.191 +13.5 AAK T 65 49 43 157 0.226 +34.6 AAK EXP 70 52 53 175 0.242 +43.8 AAK EXP H 64 61 51 176 0.253 +50.3 Table 3: Results for our impedance coupling strategies . are published and the rate that ads are added or modified. Further, the data needed by our strategies (page crawling, page expansion, and ad link crawling) can be gathered and processed offline, not affecting the user experience. Thus, from this point of view, the performance is not critical and will not be addressed in this work. RELATED WORK Several works have stressed the importance of relevance in advertising. For example, in [14] it was shown that advertisements that are presented to users when they are not interested on them are viewed just as annoyance. Thus, in order to be effective, the authors conclude that advertisements should be relevant to consumer concerns at the time of exposure. The results in [9] enforce this conclusion by pointing out that the more targeted the advertising, the more effective it is. Therefore it is not surprising that other works have addressed the relevance issue. For instance, in [8] it is proposed a system called ADWIZ that is able to adapt online advertisement to a user's short-term interests in a non-intrusive way. Contrary to our work, ADWIZ does not directly use the content of the page viewed by the user. It relies on search keywords supplied by the user to search engines and on the URL of the page requested by the user. On the other hand, in [7] the authors presented an intrusive approach in which an agent sits between advertisers and the user's browser allowing a banner to be placed into the currently viewed page. In spite of having the opportunity to use the page's content, 502 the agent infers relevance based on category information and user's private information collected along the time. In [5] the authors provide a comparison between the ranking strategies used by Google and Overture for their keyword advertising systems. Both systems select advertisements by matching them to the keywords provided by the user in a search query and rank the resulting advertisement list according to the advertisers' willingness to pay. In particular , Google approach also considers the clickthrough rate of each advertisement as an additional evidence for its relevance . The authors conclude that Google's strategy is better than that used by Overture. As mentioned before, the ranking problem in keyword advertising is different from that of content-targeted advertising. Instead of dealing with keywords provided by users in search queries, we have to deal with the contents of a page which can be very diffuse. Finally, the work in [4] focuses on improving search engine results in a TREC collection by means of an automatic query expansion method based on kNN [17]. Such method resembles our expansion approach presented in section 3. Our method is different from that presented by [4]. They expand user queries applied to a document collection with terms extracted from the top k documents returned as answer to the query in the same collection. In our case, we use two collections: an advertisement and a Web collection. We expand triggering pages with terms extracted from the Web collection and then we match these expanded pages to the ads from the advertisement collection. By doing this, we emphasize the main topics of the triggering pages, increasing the possibility of associating relevant ads with them. CONCLUSIONS In this work we investigated ten distinct strategies for associating ads with a Web page that is browsed (content-targeted advertising). Five of our strategies attempt to match the ads directly to the Web page. Because of that, they are called matching strategies. The other five strategies recognize that there is a vocabulary impedance problem among ads and Web pages and attempt to solve the problem by expanding the Web pages and the ads with new terms. Because of that they are called impedance coupling strategies . Using a sample of a real case database with over 93 thousand ads, we evaluated our strategies. For the five matching strategies, our results indicated that planned consideration of additional evidence (such as the keywords provided by the advertisers) yielded gains in average precision figures (for our test collection) of 60%. This was obtained by a strategy called AAK (for "ads and keywords"), which is taken as the baseline for evaluating our more advanced impedance coupling strategies. For our five impedance coupling strategies, the results indicate that additional gains in average precision of 50% (now relative to the AAK strategy) are possible. These were generated by expanding the Web page with new terms (obtained using a sample Web collection containing over five million pages) and the ads with the contents of the page they point to (a hyperlink provided by the advertisers). These are first time results that indicate that high quality content-targeted advertising is feasible and practical. ACKNOWLEDGEMENTS This work was supported in part by the GERINDO project , grant MCT/CNPq/CT-INFO 552.087/02-5, by CNPq grant 300.188/95-1 (Berthier Ribeiro-Neto), and by CNPq grant 303.576/04-9 (Edleno Silva de Moura). Marco Cristo is supported by Fucapi, Manaus, AM, Brazil. REFERENCES [1] The Google adwords. Google content-targeted advertising. http://adwords.google.com/select/ct_faq.html, November 2004. [2] R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. Addison-Wesley-Longman, 1st edition, 1999. [3] H. K. Bhargava and J. Feng. Paid placement strategies for internet search engines. In Proceedings of the eleventh international conference on World Wide Web, pages 117123. ACM Press, 2002. [4] E. P. Chan, S. Garcia, and S. Roukos. Trec-5 ad hoc retrieval using k nearest-neighbors re-scoring. In The Fifth Text REtrieval Conference (TREC-5). National Institute of Standards and Technology (NIST), November 1996. [5] J. Feng, H. K. Bhargava, and D. Pennock. Comparison of allocation rules for paid placement advertising in search engines. In Proceedings of the 5th international conference on Electronic commerce, pages 294299. ACM Press, 2003. [6] D. Hawking, N. Craswell, and P. B. Thistlewaite. Overview of TREC-7 very large collection track. In The Seventh Text REtrieval Conference (TREC-7), pages 91104, Gaithersburg, Maryland, USA, November 1998. [7] Y. Kohda and S. Endo. Ubiquitous advertising on the www: merging advertisement on the browser. Comput. Netw. ISDN Syst., 28(7-11):14931499, 1996. [8] M. Langheinrich, A. Nakamura, N. Abe, T. Kamba, and Y. Koseki. Unintrusive customization techniques for web advertising. Comput. Networks, 31(11-16):12591272, 1999. [9] T. P. Novak and D. L. Hoffman. New metrics for new media: toward the development of web measurement standards. World Wide Web J., 2(1):213246, 1997. [10] J. Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of plausible inference. Morgan Kaufmann Publishers, 2nd edition, 1988. [11] B. Ribeiro-Neto and R. Muntz. A belief network model for IR. In Proceedings of the 19th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 253260, Zurich, Switzerland, August 1996. [12] A. Silva, E. Veloso, P. Golgher, B. Ribeiro-Neto, A. Laender, and N. Ziviani. CobWeb - a crawler for the brazilian web. In Proceedings of the String Processing and Information Retrieval Symposium (SPIRE'99), pages 184191, Cancun, Mexico, September 1999. [13] H. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187222, July 1991. [14] C. Wang, P. Zhang, R. Choi, and M. Daeredita. Understanding consumers attitude toward advertising. In Eighth Americas Conference on Information Systems, pages 11431148, August 2002. [15] M. Weideman. Ethical issues on content distribution to digital consumers via paid placement as opposed to website visibility in search engine results. In The Seventh ETHICOMP International Conference on the Social and Ethical Impacts of Information and Communication Technologies, pages 904915. Troubador Publishing Ltd, April 2004. [16] M. Weideman and T. Haig-Smith. An investigation into search engines as a form of targeted advert delivery. In Proceedings of the 2002 annual research conference of the South African institute of computer scientists and information technologists on Enablement through technology, pages 258258. South African Institute for Computer Scientists and Information Technologists, 2002. [17] Y. Yang. Expert network: Effective and efficient learning from human decisions in text categorization and retrieval. In W. B. Croft and e. C. J. van Rijsbergen, editors, Proceedings of the 17rd annual international ACM SIGIR conference on Research and development in information retrieval, pages 1322. Springer-Verlag, 1994. 503

;advertisements;triggering page;Bayesian networks;Advertising;matching;kNN;Web;content-targeted advertising;impedance coupling

104

Implementing the IT Fundamentals Knowledge Area

The recently promulgated IT model curriculum contains IT fundamentals as one of its knowledge areas. It is intended to give students a broad understanding of (1) the IT profession and the skills that students must develop to become successful IT professionals and (2) the academic discipline of IT and its relationship to other disciplines. As currently defined, the IT fundamentals knowledge area requires 33 lecture hours to complete. The model curriculum recommends that the material relevant to the IT fundamentals knowledge area be offered early in the curriculum, for example in an introduction to IT course; however, many institutions will have to include additional material in an introductory IT course. For example, the Introduction of IT course at Georgia Southern University is used to introduce students to the available second disciplines (an important part of the Georgia Southern IT curriculum aimed at providing students with in-depth knowledge of an IT application domain), some productivity tools, and SQL. For many programs there may be too much material in an introductory IT course. This paper describes how Georgia Southern University resolved this dilemma.

INTRODUCTION The recently promulgated IT Model Curriculum, available at http://sigite.acm.org/activities/curriculum/, consists of 12 knowledge areas including IT fundamentals (ITF). ITF is intended to provide students with a set of foundation skills and provide an overview of the discipline of IT and its relationship to other computing disciplines. It is also intended to help students understand the diverse contexts in which IT is used and the challenges inherent in the diffusion of innovative technology. Given its foundational nature, it will not come as a surprise that the model curriculum recommends that ITF is covered early in a student's program of study, and it seems most logical that this knowledge area be covered in an introductory course in a baccalaureate program in IT. The IT Model curriculum recommends a minimum coverage of 33 lecture hours for the ITF knowledge area; however, a typical 3-credit semester course gives an instructor, at most, 45 lecture hours, and many programs will have to include additional material in an introductory course. For example, an important element of the IT program at Georgia Southern University is the inclusion of second disciplines, coherent sets of 7 courses in an IT application area, such as electronic broadcasting, law enforcement, music technology, and supply chain management ([5], [6]). Since students must begin introductory courses in their second discipline relatively early in their academic program, it is important that they be exposed to the range of second disciplines available to them early, and the most appropriate place to do this is in the introductory IT course. Also, students enrolling in the introductory IT course at Georgia Southern are not expected to have taken a computer literacy course beforehand, and it has become clear that many are weak in the use of spreadsheets. Since the program strongly believes that IT graduates can be expected to be conversant with basic productivity tools, including spreadsheets, the course must cover the basics of spreadsheet application. Finally, the introductory IT course must also provide a basic coverage of SQL, because the web design course, which covers n-tier architectures and requires a basic knowledge of SQL, is taught before the data management course in which SQL is normally presented. While the additional material that has to be covered in an introductory IT course is likely to differ between institutions, it is likely that many, if not all, IT programs will have to cover some additional material. Given that ITF already requires 33 lecture hours, considerable pressure is placed upon instructors in introductory IT courses to cover both the ITF material and whatever additional material needs to be included. The intent of this paper is to describe how this particular dilemma was resolved at Georgia Southern University. Section 2 provides more details about the IT fundamentals knowledge area, while section 3 discusses the introduction to IT course offered at Georgia Southern University. Section 4 concludes. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SIGITE 05, October 2022, 2005, Newark, NJ, USA. Copyright 2005 ACM 1-59593-252-6/05/0010...$5.00. 1 THE IT FUNDAMENTALS KNOWLEDGE AREA The IT Model Curriculum follows the example set by the Computer Science model curriculum (http://www.acm.org/ education/curricula.html) and distinguishes between a number of knowledge areas, each consisting of a number of knowledge units. Knowledge units are themselves composed of topics and learning outcomes. For reasons explained in ([4]), the IT model curriculum differs from the computer science model curriculum in that it distinguishes between core learning outcomes, which every graduate from an IT program is expected to achieve, and elective learning outcomes, which only graduates specializing in this area are expected to achieve. Given the foundational nature of ITF, it should come as no surprise that ITF only has core learning outcomes associated with it. Below are listed the knowledge units and the core learning outcomes associated with each. The number behind each knowledge unit is the minimum recommended coverage expressed in lecture hours. ITF1. Pervasive themes in IT (17) 1. Describe the components of IT systems and their interrelationships. 2. Describe how complexity occurs in IT. 3. Recognize that an IT professional must know how to manage complexity. 4. List examples of tools and methods used in IT for managing complexity. 5. Describe the role of the IT professional as the user advocate. 6. Explain why life-long learning and continued professional development is critical for an IT professional. 7. Explain why adaptability and interpersonal skills are important to an IT professional. 8. Distinguish between data and information, and describe the interrelationship. 9. Describe the importance of data and information in IT. 10. Explain why the mastery of information and communication technologies is important to an IT professional. 11. Explain why the IAS perspective needs to pervade all aspects of IT. ITF2. Organizational Issues (6) 1. Describe the elements of a feasible IT application. 2. Identify the extent and activities involved in an IT application. 3. Understand the requirements of the business processes. 4. Outline the project management processes. 5. List the integration processes. ITF3. History of IT (3) 1. Outline the history of computing technology. 2. Describe significant impacts of computing on society. 3. Describe significant changes in human-computer interaction. 4. 4. Outline the history of the Internet. ITF4. IT and its related and informing disciplines (3) 1. Define "Information Technology." 2. Describe the relationship between IT and other computing disciplines. 3. Describe the relationship between IT and non-computing disciplines. 4. Explain why mathematics and statistics are important in IT. ITF5. Application domains (2) 1. Describe the application of IT in non-computing disciplines. 2. Describe how IT has impacted almost all aspects of modern living. 3. Describe ways and extents in which IT has changed the interaction and communication in our society. 4. Describe how IT has impacted the globalization of world economy, culture, political systems, health, security, warfare, etc . ITF6. Application of math and statistics to IT (2) 1. Recognize the foundation of IT is built upon the various aspects of mathematics. 2. Understand the number systems used in computation. 3. Explain data representation and encoding systems. 4. Describe the current encryption methods and their limitations. 5. Describe the pervasive usage of mathematical concepts, such as functions, relations, sets as well as basic logic used in programming. 6. Recognize the value of probability and statistics. 7. Describe the basic data analysis concepts and methods used in IT applications. The total minimum recommended coverage thus is 33 lecture hours. THE INTRODUCTION TO IT COURSE AT GEORGIA SOUTHERN UNIVERSITY The introduction to IT course (IT 1130) offered in the Department of IT at Georgia Southern University is designed to introduce students to IT as a discipline and cover some productivity tools, namely Excel and Access. In line with all other IT courses at Georgia Southern University, IT 1130 was formulated through a set of explicit learning outcomes. The learning outcomes for IT 1130 are 1. Demonstrate a basic understanding of the field of IT, including the ability to i. Define the term "Information Technology"; ii. Recognize the disciplines that have contributed to the emergence of IT, namely computer science, information systems, and computer engineering; iii. Identify areas in which IT has significantly impacted individuals, organizations and/or societies. 2. Demonstrate an understanding of basic information technology software applications, including the ability to i. Using a given specification, create a simple database; ii. Use SQL for simple queries; iii. Use an office productivity suite. The overlap between Objective 1 and the ITF Knowledge Area is significant; however, due to Objective 2, the introductory IT course at Georgia Southern must cover significant additional material not specified in the IT fundamentals knowledge area. 2 3.2 Course Outline and its Mapping to the IT Fundamentals Knowledge Area The Introduction to IT course at Georgia Southern consists of 45 lecture hours. Teaching productivity tools, Learning Outcome 2 listed in Section 3.1, accounts for roughly 9 hours of instruction. Exams conducted during the semester account for 3 hours of instruction. This leaves 33 lecture hours to cover the remaining topics for IT 1130 relating to Learning Outcome 1 listed in Section 3.1. Table 1 provides a breakdown of the topics covered in the remaining 33 hours of instruction, the number of lecture hours spent on that topic, as well as the learning outcome in the IT fundamentals knowledge area of the model curriculum to which the topic corresponds. TABLE 1: IT 1130 Topics and ITF Learning Outcomes IT 1130 Topic Objective # Hours 1 Define IT ITF4.1 1 2 Data and Information ITF1.8 ITF1.9 1 3 Components of IT Systems Hardware Software Networks User ITF1.1 8.5 4 Core Technologies Data Management Networking Web Systems SAD Programming HCI Specializations in BSIT ITF1.10 ITF2.1 ITF2.2 ITF2.3 ITF2.4 ITF2.5 8.5 5 Related Disciplines ITF4.2 ITF4.3 ITF4.4 2 6 Application Domains (Second Disciplines in BSIT) ITF5.1 ITF5.2 ITF5.3 ITF 5.4 ITF 3.2 7 7 History of IT ITF3.1 ITF3.4 1 8 Viruses, Crime, Law, Ethics, Privacy & Security ITF1.11 ITF 3.2 3 9 IT as a Profession ITF1.5 ITF1.6 ITF1.7 ITF1.10 1 TOTAL 33 Table 2 compares the number of hours of instruction in the IT 1130 course for each of the knowledge units in the IT fundamentals area to the minimum recommended number of lecture hours listed in the model curriculum. The next section, Section 3.3, discusses the discrepancies between the recommended number of hours and the actual number of hours taught. TABLE 2: Comparison of IT 1130 to ITF Knowledge Area ITF Knowledge Units ITF Recommended IT 1130 Knowledge Units Not Covered ITF1 17 14 1.2, 1.3, 1.4 ITF2 6 7.5 ITF3 3 2 3.3 ITF4 3 3 ITF5 2 6.5 ITF6 2 Not Covered 6.1 6.7 TOTAL 33 33 3.3 Some Observations Table 2 illustrates several noteworthy differences between the IT 1130 course at Georgia Southern University and the knowledge units in the ITF knowledge area. 1. A discrepancy exists between the minimum number of hours recommended for ITF1 (pervasive themes in IT) and the number of hours taught in IT 1130. The 3 hour discrepancy can be attributed to the lack of coverage in IT 1130 of outcomes ITF1.2 4. Thus, IT 1130 provides no explicit coverage of the reasons for the emergence of complexity in IT, the need for IT professionals to handle complexity, and the tools and techniques available to an IT professional in IT1130. Instead, the IT program at Georgia Southern covers complexity-related issues in a number of courses throughout the curriculum. For example, some complexity-related issues are discussed in a two-course sequence of Java programming courses. Standards are discussed in a number of courses throughout the curriculum, including a data communication course and a web design course in which students learns how to implement n-tier architectures. Finally, complexity related issues are also covered in a capstone course on IT issues and management. Since the need to manage complexity is identified in the IT model curriculum as a pervasive theme, this is a reasonable alternative to cover this issue. 2. The IT 1130 course devotes more lectures hours than the minimum recommendation to ITF2 (organizational issues) and ITF5 (application domains). As the recommendation is a minimum, this is not problematic; however, it is worth noting that the explanation for these discrepancies relates directly to the structure of the IT major at Georgia Southern University. IT majors are expected to take a number of core courses, including courses in programming; web design; software acquisition, implementation and integration; networking; 3 systems analysis and design; data management; and project management. In addition, IT majors specialize in either knowledge management and it integration, systems development and support, telecommunications and network administration, or web and multimedia foundations. It is useful to students starting out on their academic program in IT to receive information on the structure of the core of the program, the courses that it consists of and how they relate to each other, and on the different specializations available to them. Since, for most IT majors, IT 1130 is the first course in the program, it is the logical place to meet this aim. Clearly, a full discussion of the structure of the program covers more than just data management (ITF1.10), a broad overview of IT applications (ITF2.1) and their development (ITF2.2), systems analysis (ITF2.3), project management (ITF2.4), and IT integration (ITF2.5). This explains why IT 1130 devotes 1.5 more hours than the recommended minimum 6. Another important element of the IT program at Georgia Southern is the inclusion of second disciplines. One of the explicit program outcomes of the BS in IT program at Georgia Southern is that, on graduation, graduates will be able "to demonstrate sufficient understanding of an application domain to be able to develop IT applications suitable for that application domain." This outcome was included at the recommendation of industry representatives who were consulted when the IT program was designed ([5]). For students to develop this ability, they must be exposed to an IT application domain, and the BS IT program at Georgia Southern therefore contains so-called second disciplines. Second disciplines are coherent sets of 7 3-credit courses in potential IT application domains, such as electronic broadcasting, law enforcement, music technology, or supply chain management. Students typically start taking courses in their second discipline early in their program of study (the standard program of study suggests that students take their first second discipline course in the first semester of their sophomore year). It is therefore important that students be exposed to the different second disciplines available to them early, and IT 1130 is the logical place to do so. One fortunate side effect of the need to introduce a second discipline is that it gives the program an excellent opportunity to make students aware of the broad range of areas in which IT can be applied and, hence, cover ITF5 (application domains); however, since the number of second disciplines is large (currently, 26), adequate coverage requires 4.5 hours more than the minimum recommend coverage for ITF 5 (application domains) 3. One lecture hour is missing in ITF3 (history of IT) due to lack of coverage in the IT 1130 course of significant changes in HCI (ITF3.3). Some material relevant to this topic is introduced in other courses that students tend to take early in their program of study, such as the Introductory Java course and the introductory web design course. For example, the introductory web design course includes among its course objectives that students develop the ability to design Web pages in accordance with good design principles using appropriate styles and formats and the ability to design Web pages that are ADA compliant. Material relevant to both objectives allows us to expand on HCI design principles and place these in a historical context. Moreover, students are advised to take the introductory web design course in the semester following the one in which they take IT 1130, and they are therefore likely to be exposed to material relevant to ITF3.3 early in their program of study. 4. The final discrepancy lies in the coverage of the learning outcomes corresponding to the ITF6 (application of math and statistics to IT) in the IT 1130 course; however, the material related to this knowledge unit is covered in two courses that students are again advised to take early in their program of study. One course is a course in discrete mathematics, designed specifically for IT majors. It includes among its course objectives the ability to explain the importance of discrete mathematics in computer science and information technology and provides in-depth coverage of functions, sets, basic propositional logic, and algorithm design. Finally, all students enrolled in the IT major take a statistics course, which covers probability. 3.4 Support Material Since the ITF knowledge area is relatively new, no single textbook covers all relevant material. We therefore use a variety of sources to support the course. First, we use Excel 2003 ([8]) and Access 2003 ([7]) to support the teaching of spreadsheets and SQL (IT 1130 course outcomes 2i-2iii identified in section 3.1). Second, to support the teaching of Topics 3 (components of IT systems), 4 (core technologies) and 7 (history of IT), we use Discovering Computers 2005 ([9]). While the textbook provides a reasonable coverage of some of the subtopics discussed, it does not sufficiently stress the importance of the users and the importance of HCI in systems development, and we, therefore, emphasize this issue throughout the course. We discussed the way in which we cover these topics in Points 2 and 3 in section 3.3. Third, for topics 6 (Application Domains), 8 (Viruses, Crime, Law, Ethics, Privacy and Security) and 9 (IT as a profession), we use Computers in Our World ([3]); however, we do not rely solely on the textbook for our coverage of topic 6. Again, we discussed this in Point 2 in section 3.3. Finally, to support Topics 1 (define IT), 2 (data and information), and 5 (IT and its related disciplines), students are given material written specifically for the course. Also, we invite representatives from computer science and information systems to lecture on their specific disciplines and follow this up with a lecture on computer engineering and a discussion on the relationship between all four disciplines. Table 3 lists the core learning outcomes for each of the ITF knowledge units and maps them to the material in the IT 1130 course used to achieve that outcome. The material comes either from Discovering Computers 2005 ([9]) (DC), Computers in Our World ([3]) (CIOW), or material written specifically for the course (supplemental material) and/or lectures/discussions led by faculty members from other related departments. 4 TABLE 3: Course Materials Used in IT 1130 to Achieve ITF Learning Outcomes ITF Knowledge Units Learning Outcomes Material 1 DC Chapters 3-9 2-4 Not covered 5-7 DC Chapters 12 & 15, CIOW Chapters 8 & 9, Supplemental Materials 8.9 DC Chapter 10, Supplemental Materials 10 DC Chapters 2, 9, 10,12, 13, Supplemental Materials ITF 1 11 CIOW Chapters 7-9 ITF2 1-5 DC Chapters 2, 9, 10, 12, 13. Supplemental Materials 1 DC Timeline between Chapters 1 and 2, Chapter 2 2 CIOW Chapters 1 - 9 3 Not covered ITF3 4 DC Timeline between Chapters 1 and 2, Chapter 2 ITF 4 1-4 Supplemental Materials, Lecture and Class Discussion led by CS, IS and IT representatives ITF5 1-4 CIOW Chapters 1 9 ITF6 1-7 Not covered *Discovering Computers = DC, Computers in Our World = CIOW CONCLUSIONS The IT Fundamentals knowledge area in the IT model curriculum is of central importance to the design of an introductory IT course; however, since institutions will have to include additional materials in their introductory IT courses, depending on the nature of their program, the minimum requirement of 33 lecture hours to cover this material is likely to lead to problems. This paper presents the experience with an introductory IT course at Georgia Southern University, IT1130. In general, we believe that, despite the need to include additional material in IT1130, we are able to cover most of the knowledge units in the IT fundamentals knowledge area. We are confident that those knowledge units not covered in IT1130 are covered in other courses that students are advised to take early in their programs of study. Finally, despite the fact that the IT fundamentals knowledge area is new and that no textbooks cover all the knowledge units within the area, we have been able to identify a set of textbooks that, jointly, cover most of the material; however, we provide a relatively small amount of additional material, and the textbooks we identified do not always cover the material at the appropriate level. Therefore, support materials specifically for the IT fundamentals knowledge area need to be developed. Whether this is best provided in the form of a textbook, or, more dynamically, as a set of online learning objects ([1], [2]) is a question open to debate. REFERENCES [1] Abernethy, K., Treu, K, Piegari, G, Reichgelt. H. "An implementation model for a learning object repository", October 2005, E-learn 2005 World Conference on E-learning in corporate, government, healthcare and higher education. Vancouver, Canada. [2] Abernethy, K., Treu, K, Piegari, G, Reichgelt. H. "A learning object repository in support of introduction to information technology", August 2005, 6 th Annual Conference for the Higher Education Academy Subject Network for Information and Computer Science, York, England. [3] Jedlicka, L. Computers in Our World. Thompson Course Technologies, 2003. [4] Lawson, E, Reichgelt, H, Lunt, B. Ekstrom, J, Kamali, R. Miller, J and Gorka, S, The Information Technology Model Curriculum. Paper submitted to ISECON 2005. [5] Reichgelt, H., Price, B. and Zhang, A., "Designing an Information Technology curriculum: The Georgia Southern experience", Journal of Information Technology Education 2002, Vol. 1, No. 4, 213-221 [6] Reichgelt, H., Price, B. and Zhang, A., The Inclusion of Application Areas in IT Curricula, SIGITE--3, Rochester, NY, ACM-SIGITE (formerly SITE), September 2002 [7] Shelley, G., Cashman, T., Pratt, P. and Last, M. Microsoft Office Access 2003. Thompson Course Technologies, 2004. [8] Shelley, G., Cashman, T., Quasney, J. Microsoft Office Excel 2003. Thompson Course Technologies, 2004. [9] Shelley, G., Vermaat, M. and Cashman, T. Discovering Computers 2005: A Gateway to Information. Thompson Course Technologies, 2005. 5

IT Fundamentals Knowledge Area;IT Model Curriculum

105

Implicit User Modeling for Personalized Search

Information retrieval systems (e.g., web search engines) are critical for overcoming information overload. A major deficiency of existing retrieval systems is that they generally lack user modeling and are not adaptive to individual users, resulting in inherently non-optimal retrieval performance. For example, a tourist and a programmer may use the same word "java" to search for different information, but the current search systems would return the same results. In this paper, we study how to infer a user's interest from the user's search context and use the inferred implicit user model for personalized search . We present a decision theoretic framework and develop techniques for implicit user modeling in information retrieval. We develop an intelligent client-side web search agent (UCAIR) that can perform eager implicit feedback, e.g., query expansion based on previous queries and immediate result reranking based on clickthrough information. Experiments on web search show that our search agent can improve search accuracy over the popular Google search engine.

INTRODUCTION Although many information retrieval systems (e.g., web search engines and digital library systems) have been successfully deployed, the current retrieval systems are far from optimal. A major deficiency of existing retrieval systems is that they generally lack user modeling and are not adaptive to individual users [17]. This inherent non-optimality is seen clearly in the following two cases: Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. CIKM'05, October 31November 5, 2005, Bremen, Germany. Copyright 2005 ACM 1-59593-140-6/05/0010 ... $ 5.00. (1) Different users may use exactly the same query (e.g., "Java") to search for different information (e.g., the Java island in Indonesia or the Java programming language), but existing IR systems return the same results for these users. Without considering the actual user, it is impossible to know which sense "Java" refers to in a query. (2) A user's information needs may change over time. The same user may use "Java" sometimes to mean the Java island in Indonesia and some other times to mean the programming language. Without recognizing the search context, it would be again impossible to recognize the correct sense. In order to optimize retrieval accuracy, we clearly need to model the user appropriately and personalize search according to each individual user. The major goal of user modeling for information retrieval is to accurately model a user's information need, which is, unfortunately, a very difficult task. Indeed, it is even hard for a user to precisely describe what his/her information need is. What information is available for a system to infer a user's information need? Obviously, the user's query provides the most direct evidence. Indeed, most existing retrieval systems rely solely on the query to model a user's information need. However, since a query is often extremely short, the user model constructed based on a keyword query is inevitably impoverished . An effective way to improve user modeling in information retrieval is to ask the user to explicitly specify which documents are relevant (i.e., useful for satisfying his/her information need), and then to improve user modeling based on such examples of relevant documents. This is called relevance feedback, which has been proved to be quite effective for improving retrieval accuracy [19, 20]. Unfortunately, in real world applications, users are usually reluctant to make the extra effort to provide relevant examples for feedback [11]. It is thus very interesting to study how to infer a user's information need based on any implicit feedback information, which naturally exists through user interactions and thus does not require any extra user effort. Indeed, several previous studies have shown that implicit user modeling can improve retrieval accuracy. In [3], a web browser (Curious Browser) is developed to record a user's explicit relevance ratings of web pages (relevance feedback) and browsing behavior when viewing a page, such as dwelling time, mouse click, mouse movement and scrolling (implicit feedback). It is shown that the dwelling time on a page, amount of scrolling on a page and the combination of time and scrolling have a strong correlation with explicit relevance ratings, which suggests that implicit feedback may be helpful for inferring user information need. In [10], user clickthrough data is collected as training data to learn a retrieval function, which is used to produce a customized ranking of search results that suits a group of users' preferences. In [25], the clickthrough data collected over a long time period is exploited through query expansion to improve retrieval accuracy. 824 While a user may have general long term interests and preferences for information, often he/she is searching for documents to satisfy an "ad hoc" information need, which only lasts for a short period of time; once the information need is satisfied, the user would generally no longer be interested in such information. For example, a user may be looking for information about used cars in order to buy one, but once the user has bought a car, he/she is generally no longer interested in such information. In such cases, implicit feedback information collected over a long period of time is unlikely to be very useful, but the immediate search context and feedback information, such as which of the search results for the current information need are viewed, can be expected to be much more useful. Consider the query "Java" again. Any of the following immediate feedback information about the user could potentially help determine the intended meaning of "Java" in the query: (1) The previous query submitted by the user is "hashtable" (as opposed to, e.g., "travel Indonesia"). (2) In the search results, the user viewed a page where words such as "programming", "software", and "applet" occur many times. To the best of our knowledge, how to exploit such immediate and short-term search context to improve search has so far not been well addressed in the previous work. In this paper, we study how to construct and update a user model based on the immediate search context and implicit feedback information and use the model to improve the accuracy of ad hoc retrieval. In order to maximally benefit the user of a retrieval system through implicit user modeling, we propose to perform "eager implicit feedback". That is, as soon as we observe any new piece of evidence from the user, we would update the system's belief about the user's information need and respond with improved retrieval results based on the updated user model. We present a decision-theoretic framework for optimizing interactive information retrieval based on eager user model updating , in which the system responds to every action of the user by choosing a system action to optimize a utility function. In a traditional retrieval paradigm, the retrieval problem is to match a query with documents and rank documents according to their relevance values. As a result, the retrieval process is a simple independent cycle of "query" and "result display". In the proposed new retrieval paradigm, the user's search context plays an important role and the inferred implicit user model is exploited immediately to benefit the user. The new retrieval paradigm is thus fundamentally different from the traditional paradigm, and is inherently more general. We further propose specific techniques to capture and exploit two types of implicit feedback information: (1) identifying related immediately preceding query and using the query and the corresponding search results to select appropriate terms to expand the current query, and (2) exploiting the viewed document summaries to immediately rerank any documents that have not yet been seen by the user. Using these techniques, we develop a client-side web search agent UCAIR (User-Centered Adaptive Information Retrieval) on top of a popular search engine (Google). Experiments on web search show that our search agent can improve search accuracy over Google. Since the implicit information we exploit already naturally exists through user interactions, the user does not need to make any extra effort. Thus the developed search agent can improve existing web search performance without additional effort from the user. The remaining sections are organized as follows. In Section 2, we discuss the related work. In Section 3, we present a decision-theoretic interactive retrieval framework for implicit user modeling. In Section 4, we present the design and implementation of an intelligent client-side web search agent (UCAIR) that performs eager implicit feedback. In Section 5, we report our experiment results using the search agent. Section 6 concludes our work. RELATED WORK Implicit user modeling for personalized search has been studied in previous work, but our work differs from all previous work in several aspects: (1) We emphasize the exploitation of immediate search context such as the related immediately preceding query and the viewed documents in the same session, while most previous work relies on long-term collection of implicit feedback information [25]. (2) We perform eager feedback and bring the benefit of implicit user modeling as soon as any new implicit feedback information is available, while the previous work mostly exploits long-term implicit feedback [10]. (3) We propose a retrieval framework to integrate implicit user modeling with the interactive retrieval process , while the previous work either studies implicit user modeling separately from retrieval [3] or only studies specific retrieval models for exploiting implicit feedback to better match a query with documents [23, 27, 22]. (4) We develop and evaluate a personalized Web search agent with online user studies, while most existing work evaluates algorithms offline without real user interactions. Currently some search engines provide rudimentary personalization , such as Google Personalized web search [6], which allows users to explicitly describe their interests by selecting from predefined topics, so that those results that match their interests are brought to the top, and My Yahoo! search [16], which gives users the option to save web sites they like and block those they dislike . In contrast, UCAIR personalizes web search through implicit user modeling without any additional user efforts. Furthermore, the personalization of UCAIR is provided on the client side. There are two remarkable advantages on this. First, the user does not need to worry about the privacy infringement, which is a big concern for personalized search [26]. Second, both the computation of personalization and the storage of the user profile are done at the client side so that the server load is reduced dramatically [9]. There have been many works studying user query logs [1] or query dynamics [13]. UCAIR makes direct use of a user's query history to benefit the same user immediately in the same search session. UCAIR first judges whether two neighboring queries belong to the same information session and if so, it selects terms from the previous query to perform query expansion. Our query expansion approach is similar to automatic query expansion [28, 15, 5], but instead of using pseudo feedback to expand the query, we use user's implicit feedback information to expand the current query. These two techniques may be combined. OPTIMIZATION IN INTERACTIVE IR In interactive IR, a user interacts with the retrieval system through an "action dialogue", in which the system responds to each user action with some system action. For example, the user's action may be submitting a query and the system's response may be returning a list of 10 document summaries. In general, the space of user actions and system responses and their granularities would depend on the interface of a particular retrieval system. In principle, every action of the user can potentially provide new evidence to help the system better infer the user's information need. Thus in order to respond optimally, the system should use all the evidence collected so far about the user when choosing a response. When viewed in this way, most existing search engines are clearly non-optimal. For example, if a user has viewed some documents on the first page of search results, when the user clicks on the "Next" link to fetch more results, an existing retrieval system would still return the next page of results retrieved based on the original query without considering the new evidence that a particular result has been viewed by the user. 825 We propose to optimize retrieval performance by adapting system responses based on every action that a user has taken, and cast the optimization problem as a decision task. Specifically, at any time, the system would attempt to do two tasks: (1) User model updating: Monitor any useful evidence from the user regarding his/her information need and update the user model as soon as such evidence is available; (2) Improving search results: Rerank immediately all the documents that the user has not yet seen, as soon as the user model is updated. We emphasize eager updating and reranking, which makes our work quite different from any existing work. Below we present a formal decision theoretic framework for optimizing retrieval performance through implicit user modeling in interactive information retrieval. 3.1 A decision-theoretic framework Let A be the set of all user actions and R(a) be the set of all possible system responses to a user action a A. At any time, let A t = (a 1 , ..., a t ) be the observed sequence of user actions so far (up to time point t) and R t-1 = (r 1 , ..., r t-1 ) be the responses that the system has made responding to the user actions. The system's goal is to choose an optimal response r t R(a t ) for the current user action a t . Let M be the space of all possible user models. We further define a loss function L(a, r, m) , where a A is a user action, r R(a) is a system response, and m M is a user model. L(a, r, m) encodes our decision preferences and assesses the optimality of responding with r when the current user model is m and the current user action is a. According to Bayesian decision theory, the optimal decision at time t is to choose a response that minimizes the Bayes risk, i.e., r t = argmin rR(a t ) M L(a t , r, m t )P (m t |U, D, A t , R t-1 )dm t (1) where P (m t |U, D, A t , R t-1 ) is the posterior probability of the user model m t given all the observations about the user U we have made up to time t. To simplify the computation of Equation 1, let us assume that the posterior probability mass P (m t |U, D, A t , R t-1 ) is mostly concentrated on the mode m t = argmax m t P (m t |U, D, A t , R t-1 ). We can then approximate the integral with the value of the loss function at m t . That is, r t argmin rR(a t ) L(a t , r, m t ) (2) where m t = argmax m t P (m t |U, D, A t , R t-1 ). Leaving aside how to define and estimate these probabilistic models and the loss function, we can see that such a decision-theoretic formulation suggests that, in order to choose the optimal response to a t , the system should perform two tasks: (1) compute the current user model and obtain m t based on all the useful information . (2) choose a response r t to minimize the loss function value L(a t , r t , m t ). When a t does not affect our belief about m t , the first step can be omitted and we may reuse m t-1 for m t . Note that our framework is quite general since we can potentially model any kind of user actions and system responses. In most cases, as we may expect, the system's response is some ranking of documents, i.e., for most actions a, R(a) consists of all the possible rankings of the unseen documents, and the decision problem boils down to choosing the best ranking of unseen documents based on the most current user model. When a is the action of submitting a keyword query, such a response is exactly what a current retrieval system would do. However, we can easily imagine that a more intelligent web search engine would respond to a user's clicking of the "Next" link (to fetch more unseen results) with a more opti-mized ranking of documents based on any viewed documents in the current page of results. In fact, according to our eager updating strategy, we may even allow a system to respond to a user's clicking of browser's "Back" button after viewing a document in the same way, so that the user can maximally benefit from implicit feedback. These are precisely what our UCAIR system does. 3.2 User models A user model m M represents what we know about the user U , so in principle, it can contain any information about the user that we wish to model. We now discuss two important components in a user model. The first component is a component model of the user's information need. Presumably, the most important factor affecting the optimality of the system's response is how well the response addresses the user's information need. Indeed, at any time, we may assume that the system has some "belief" about what the user is interested in, which we model through a term vector x = (x 1 , ..., x |V | ), where V = {w 1 , ..., w |V | } is the set of all terms (i.e., vocabulary) and x i is the weight of term w i . Such a term vector is commonly used in information retrieval to represent both queries and documents . For example, the vector-space model, assumes that both the query and the documents are represented as term vectors and the score of a document with respect to a query is computed based on the similarity between the query vector and the document vector [21]. In a language modeling approach, we may also regard the query unigram language model [12, 29] or the relevance model [14] as a term vector representation of the user's information need. Intuitively, x would assign high weights to terms that characterize the topics which the user is interested in. The second component we may include in our user model is the documents that the user has already viewed. Obviously, even if a document is relevant, if the user has already seen the document, it would not be useful to present the same document again. We thus introduce another variable S D (D is the whole set of documents in the collection) to denote the subset of documents in the search results that the user has already seen/viewed. In general, at time t, we may represent a user model as m t = (S, x, A t , R t-1 ), where S is the seen documents, x is the system's "understanding" of the user's information need, and (A t , R t-1 ) represents the user's interaction history. Note that an even more general user model may also include other factors such as the user's reading level and occupation. If we assume that the uncertainty of a user model m t is solely due to the uncertainty of x, the computation of our current estimate of user model m t will mainly involve computing our best estimate of x. That is, the system would choose a response according to r t = argmin rR(a t ) L(a t , r, S, x , A t , R t-1 ) (3) where x = argmax x P (x|U, D, A t , R t-1 ). This is the decision mechanism implemented in the UCAIR system to be described later. In this system, we avoided specifying the probabilistic model P (x|U, D, A t , R t-1 ) by computing x directly with some existing feedback method. 3.3 Loss functions The exact definition of loss function L depends on the responses, thus it is inevitably application-specific. We now briefly discuss some possibilities when the response is to rank all the unseen documents and present the top k of them. Let r = (d 1 , ..., d k ) be the top k documents, S be the set of seen documents by the user, and x be the system's best guess of the user's information need. We 826 may simply define the loss associated with r as the negative sum of the probability that each of the d i is relevant, i.e., L(a, r, m) = k i=1 P (relevant|d i , m). Clearly, in order to minimize this loss function, the optimal response r would contain the k documents with the highest probability of relevance, which is intuitively reasonable. One deficiency of this "top-k loss function" is that it is not sensitive to the internal order of the selected top k documents, so switching the ranking order of a non-relevant document and a relevant one would not affect the loss, which is unreasonable. To model ranking , we can introduce a factor of the user model the probability of each of the k documents being viewed by the user, P (view|d i ), and define the following "ranking loss function": L(a, r, m) = k i=1 P (view|d i )P (relevant|d i , m) Since in general, if d i is ranked above d j (i.e., i &lt; j), P (view|d i ) &gt; P (view|d j ), this loss function would favor a decision to rank relevant documents above non-relevant ones, as otherwise, we could always switch d i with d j to reduce the loss value. Thus the system should simply perform a regular retrieval and rank documents according to the probability of relevance [18]. Depending on the user's retrieval preferences, there can be many other possibilities. For example, if the user does not want to see redundant documents, the loss function should include some redundancy measure on r based on the already seen documents S. Of course, when the response is not to choose a ranked list of documents, we would need a different loss function. We discuss one such example that is relevant to the search agent that we implement . When a user enters a query q t (current action), our search agent relies on some existing search engine to actually carry out search. In such a case, even though the search agent does not have control of the retrieval algorithm, it can still attempt to optimize the search results through refining the query sent to the search engine and/or reranking the results obtained from the search engine. The loss functions for reranking are already discussed above; we now take a look at the loss functions for query refinement. Let f be the retrieval function of the search engine that our agent uses so that f (q) would give us the search results using query q. Given that the current action of the user is entering a query q t (i.e., a t = q t ), our response would be f (q) for some q. Since we have no choice of f , our decision is to choose a good q. Formally, r t = argmin r t L(a, r t , m) = argmin f (q) L(a, f (q), m) = f (argmin q L(q t , f (q), m)) which shows that our goal is to find q = argmin q L(q t , f (q), m), i.e., an optimal query that would give us the best f (q). A different choice of loss function L(q t , f (q), m) would lead to a different query refinement strategy. In UCAIR, we heuristically compute q by expanding q t with terms extracted from r t-1 whenever q t-1 and q t have high similarity. Note that r t-1 and q t-1 are contained in m as part of the user's interaction history. 3.4 Implicit user modeling Implicit user modeling is captured in our framework through the computation of x = argmax x P (x|U, D, A t , R t-1 ), i.e., the system's current belief of what the user's information need is. Here again there may be many possibilities, leading to different algorithms for implicit user modeling. We now discuss a few of them. First, when two consecutive queries are related, the previous query can be exploited to enrich the current query and provide more search context to help disambiguation. For this purpose, instead of performing query expansion as we did in the previous section, we could also compute an updated x based on the previous query and retrieval results. The computed new user model can then be used to rank the documents with a standard information retrieval model. Second, we can also infer a user's interest based on the summaries of the viewed documents. When a user is presented with a list of summaries of top ranked documents, if the user chooses to skip the first n documents and to view the (n + 1)-th document, we may infer that the user is not interested in the displayed summaries for the first n documents, but is attracted by the displayed summary of the (n + 1)-th document. We can thus use these summaries as negative and positive examples to learn a more accurate user model x . Here many standard relevance feedback techniques can be exploited [19, 20]. Note that we should use the displayed summaries, as opposed to the actual contents of those documents, since it is possible that the displayed summary of the viewed document is relevant, but the document content is actually not. Similarly, a displayed summary may mislead a user to skip a relevant document. Inferring user models based on such displayed information, rather than the actual content of a document is an important difference between UCAIR and some other similar systems. In UCAIR, both of these strategies for inferring an implicit user model are implemented. UCAIR A PERSONALIZED SEARCH AGENT In this section, we present a client-side web search agent called UCAIR, in which we implement some of the methods discussed in the previous section for performing personalized search through implicit user modeling. UCAIR is a web browser plug-in 1 that acts as a proxy for web search engines. Currently, it is only implemented for Internet Explorer and Google, but it is a matter of engineering to make it run on other web browsers and interact with other search engines. The issue of privacy is a primary obstacle for deploying any real world applications involving serious user modeling, such as personalized search. For this reason, UCAIR is strictly running as a client-side search agent, as opposed to a server-side application. This way, the captured user information always resides on the computer that the user is using, thus the user does not need to release any information to the outside. Client-side personalization also allows the system to easily observe a lot of user information that may not be easily available to a server. Furthermore, performing personalized search on the client-side is more scalable than on the serverside , since the overhead of computation and storage is distributed among clients. As shown in Figure 1, the UCAIR toolbar has 3 major components : (1) The (implicit) user modeling module captures a user's search context and history information, including the submitted queries and any clicked search results and infers search session boundaries. (2) The query modification module selectively improves the query formulation according to the current user model. (3) The result re-ranking module immediately re-ranks any unseen search results whenever the user model is updated. In UCAIR, we consider four basic user actions: (1) submitting a keyword query; (2) viewing a document; (3) clicking the "Back" button; (4) clicking the "Next" link on a result page. For each of these four actions, the system responds with, respectively, (1) 1 UCAIR is available at: http://sifaka.cs.uiuc.edu/ir/ucair/download.html 827 Search Engine (e.g., Google) Search History Log (e.g.,past queries, clicked results) Query Modification Result Re-Ranking User Modeling Result Buffer UCAIR User query results clickthrough... Figure 1: UCAIR architecture generating a ranked list of results by sending a possibly expanded query to a search engine; (2) updating the information need model x; (3) reranking the unseen results on the current result page based on the current model x; and (4) reranking the unseen pages and generating the next page of results based on the current model x. Behind these responses, there are three basic tasks: (1) Decide whether the previous query is related to the current query and if so expand the current query with useful terms from the previous query or the results of the previous query. (2) Update the information need model x based on a newly clicked document summary. (3) Rerank a set of unseen documents based on the current model x. Below we describe our algorithms for each of them. 4.2 Session boundary detection and query expansion To effectively exploit previous queries and their corresponding clickthrough information, UCAIR needs to judge whether two adjacent queries belong to the same search session (i.e., detect session boundaries). Existing work on session boundary detection is mostly in the context of web log analysis (e.g., [8]), and uses statistical information rather than textual features. Since our client-side agent does not have access to server query logs, we make session boundary decisions based on textual similarity between two queries. Because related queries do not necessarily share the same words (e.g., "java island" and "travel Indonesia"), it is insufficient to use only query text. Therefore we use the search results of the two queries to help decide whether they are topically related. For example, for the above queries "java island" and "travel Indone-sia"' , the words "java", "bali", "island", "indonesia" and "travel" may occur frequently in both queries' search results, yielding a high similarity score. We only use the titles and summaries of the search results to calculate the similarity since they are available in the retrieved search result page and fetching the full text of every result page would sig-nificantly slow down the process. To compensate for the terseness of titles and summaries, we retrieve more results than a user would normally view for the purpose of detecting session boundaries (typ-ically 50 results). The similarity between the previous query q and the current query q is computed as follows. Let {s 1 , s 2 , . . . , s n } and {s 1 , s 2 , . . . , s n } be the result sets for the two queries. We use the pivoted normalization TF-IDF weighting formula [24] to compute a term weight vector s i for each result s i . We define the average result s avg to be the centroid of all the result vectors, i.e., (s 1 + s 2 + . . . + s n )/n. The cosine similarity between the two average results is calculated as s avg s avg / s 2 avg s 2 avg If the similarity value exceeds a predefined threshold, the two queries will be considered to be in the same information session. If the previous query and the current query are found to belong to the same search session, UCAIR would attempt to expand the current query with terms from the previous query and its search results. Specifically, for each term in the previous query or the corresponding search results, if its frequency in the results of the current query is greater than a preset threshold (e.g. 5 results out of 50), the term would be added to the current query to form an expanded query. In this case, UCAIR would send this expanded query rather than the original one to the search engine and return the results corresponding to the expanded query. Currently, UCAIR only uses the immediate preceding query for query expansion; in principle, we could exploit all related past queries. 4.3 Information need model updating Suppose at time t, we have observed that the user has viewed k documents whose summaries are s 1 , ..., s k . We update our user model by computing a new information need vector with a standard feedback method in information retrieval (i.e., Rocchio [19]). According to the vector space retrieval model, each clicked summary s i can be represented by a term weight vector s i with each term weighted by a TF-IDF weighting formula [21]. Rocchio computes the centroid vector of all the summaries and interpolates it with the original query vector to obtain an updated term vector. That is, x = q + (1 - ) 1 k k i=1 s i where q is the query vector, k is the number of summaries the user clicks immediately following the current query and is a parameter that controls the influence of the clicked summaries on the inferred information need model. In our experiments, is set to 0.5. Note that we update the information need model whenever the user views a document. 4.4 Result reranking In general, we want to rerank all the unseen results as soon as the user model is updated. Currently, UCAIR implements reranking in two cases, corresponding to the user clicking the "Back" button and "Next" link in the Internet Explorer. In both cases, the current (updated) user model would be used to rerank the unseen results so that the user would see improved search results immediately. To rerank any unseen document summaries, UCAIR uses the standard vector space retrieval model and scores each summary based on the similarity of the result and the current user information need vector x [21]. Since implicit feedback is not completely reliable , we bring up only a small number (e.g. 5) of highest reranked results to be followed by any originally high ranked results. 828 Google result (user query = "java map") UCAIR result (user query ="java map") previous query = "travel Indonesia" previous query = "hashtable" expanded user query = "java map Indonesia" expanded user query = "java map class" 1 Java map projections of the world ... Lonely Planet - Indonesia Map Map (Java 2 Platform SE v1.4.2) www.btinternet.com/ se16/js/mapproj.htm www.lonelyplanet.com/mapshells/... java.sun.com/j2se/1.4.2/docs/... 2 Java map projections of the world ... INDONESIA TOURISM : CENTRAL JAVA - MAP Java 2 Platform SE v1.3.1: Interface Map www.btinternet.com/ se16/js/oldmapproj.htm www.indonesia-tourism.com/... java.sun.com/j2se/1.3/docs/api/java/... 3 Java Map INDONESIA TOURISM : WEST JAVA - MAP An Introduction to Java Map Collection Classes java.sun.com/developer/... www.indonesia-tourism.com/ ... www.oracle.com/technology/... 4 Java Technology Concept Map IndoStreets - Java Map An Introduction to Java Map Collection Classes java.sun.com/developer/onlineTraining/... www.indostreets.com/maps/java/ www.theserverside.com/news/... 5 Science@NASA Home Indonesia Regions and Islands Maps, Bali, Java, ... Koders - Mappings.java science.nasa.gov/Realtime/... www.maps2anywhere.com/Maps/... www.koders.com/java/ 6 An Introduction to Java Map Collection Classes Indonesia City Street Map,... Hibernate simplifies inheritance mapping www.oracle.com/technology/... www.maps2anywhere.com/Maps/... www.ibm.com/developerworks/java/... 7 Lonely Planet - Java Map Maps Of Indonesia tmap 30.map Class Hierarchy www.lonelyplanet.com/mapshells/ www.embassyworld.com/maps/... tmap.pmel.noaa.gov/... 8 ONJava.com: Java API Map Maps of Indonesia by Peter Loud Class Scope www.onjava.com/pub/a/onjava/api map/ users.powernet.co.uk/... jalbum.net/api/se/datadosen/util/Scope.html 9 GTA San Andreas : Sam Maps of Indonesia by Peter Loud Class PrintSafeHashMap www.gtasanandreas.net/sam/ users.powernet.co.uk/mkmarina/indonesia/ jalbum.net/api/se/datadosen/... 10 INDONESIA TOURISM : WEST JAVA - MAP indonesiaphoto.com Java Pro - Union and Vertical Mapping of Classes www.indonesia-tourism.com/... www.indonesiaphoto.com/... www.fawcette.com/javapro/... Table 1: Sample results of query expansion EVALUATION OF UCAIR We now present some results on evaluating the two major UCAIR functions: selective query expansion and result reranking based on user clickthrough data. 5.1 Sample results The query expansion strategy implemented in UCAIR is inten-tionally conservative to avoid misinterpretation of implicit user models . In practice, whenever it chooses to expand the query, the expansion usually makes sense. In Table 1, we show how UCAIR can successfully distinguish two different search contexts for the query "java map", corresponding to two different previous queries (i.e., "travel Indonesia" vs. "hashtable"). Due to implicit user modeling, UCAIR intelligently figures out to add "Indonesia" and "class", respectively, to the user's query "java map", which would otherwise be ambiguous as shown in the original results from Google on March 21, 2005. UCAIR's results are much more accurate than Google's results and reflect personalization in search. The eager implicit feedback component is designed to immediately respond to a user's activity such as viewing a document. In Figure 2, we show how UCAIR can successfully disambiguate an ambiguous query "jaguar" by exploiting a viewed document summary . In this case, the initial retrieval results using "jaguar" (shown on the left side) contain two results about the Jaguar cars followed by two results about the Jaguar software. However, after the user views the web page content of the second result (about "Jaguar car") and returns to the search result page by clicking "Back" button , UCAIR automatically nominates two new search results about Jaguar cars (shown on the right side), while the original two results about Jaguar software are pushed down on the list (unseen from the picture). 5.2 Quantitative evaluation To further evaluate UCAIR quantitatively, we conduct a user study on the effectiveness of the eager implicit feedback component . It is a challenge to quantitatively evaluate the potential performance improvement of our proposed model and UCAIR over Google in an unbiased way [7]. Here, we design a user study, in which participants would do normal web search and judge a randomly and anonymously mixed set of results from Google and UCAIR at the end of the search session; participants do not know whether a result comes from Google or UCAIR. We recruited 6 graduate students for this user study, who have different backgrounds (3 computer science, 2 biology, and 1 chem-&lt ;top&gt; &lt;num&gt; Number: 716 &lt;title&gt; Spammer arrest sue &lt;desc&gt; Description: Have any spammers been arrested or sued for sending unsolicited e-mail? &lt;narr&gt; Narrative: Instances of arrests, prosecutions, convictions, and punishments of spammers, and lawsuits against them are relevant. Documents which describe laws to limit spam without giving details of lawsuits or criminal trials are not relevant. &lt;/top&gt; Figure 3: An example of TREC query topic, expressed in a form which might be given to a human assistant or librarian istry). We use query topics from TREC 2 2004 Terabyte track [2] and TREC 2003 Web track [4] topic distillation task in the way to be described below. An example topic from TREC 2004 Terabyte track appears in Figure 3. The title is a short phrase and may be used as a query to the retrieval system. The description field provides a slightly longer statement of the topic requirement, usually expressed as a single complete sentence or question. Finally the narrative supplies additional information necessary to fully specify the requirement, expressed in the form of a short paragraph. Initially, each participant would browse 50 topics either from Terabyte track or Web track and pick 5 or 7 most interesting topics. For each picked topic, the participant would essentially do the normal web search using UCAIR to find many relevant web pages by using the title of the query topic as the initial keyword query. During this process, the participant may view the search results and possibly click on some interesting ones to view the web pages, just as in a normal web search. There is no requirement or restriction on how many queries the participant must submit or when the participant should stop the search for one topic. When the participant plans to change the search topic, he/she will simply press a button 2 Text REtrieval Conference: http://trec.nist.gov/ 829 Figure 2: Screen shots for result reranking to evaluate the search results before actually switching to the next topic. At the time of evaluation, 30 top ranked results from Google and UCAIR (some are overlapping) are randomly mixed together so that the participant would not know whether a result comes from Google or UCAIR. The participant would then judge the relevance of these results. We measure precision at top n (n = 5, 10, 20, 30) documents of Google and UCAIR. We also evaluate precisions at different recall levels. Altogether, 368 documents judged as relevant from Google search results and 429 documents judged as relevant from UCAIR by participants . Scatter plots of precision at top 10 and top 20 documents are shown in Figure 4 and Figure 5 respectively (The scatter plot of precision at top 30 documents is very similar to precision at top 20 documents). Each point of the scatter plots represents the precisions of Google and UCAIR on one query topic. Table 2 shows the average precision at top n documents among 32 topics. From Figure 4, Figure 5 and Table 2, we see that the search results from UCAIR are consistently better than those from Google by all the measures. Moreover, the performance improvement is more dramatic for precision at top 20 documents than that at precision at top 10 documents. One explanation for this is that the more interaction the user has with the system, the more clickthrough data UCAIR can be expected to collect. Thus the retrieval system can build more precise implicit user models, which lead to better retrieval accuracy. Ranking Method prec@5 prec@10 prec@20 prec@30 Google 0.538 0.472 0.377 0.308 UCAIR 0.581 0.556 0.453 0.375 Improvement 8.0% 17.8% 20.2% 21.8% Table 2: Table of average precision at top n documents for 32 query topics The plot in Figure 6 shows the precision-recall curves for UCAIR and Google, where it is clearly seen that the performance of UCAIR 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 UCAIR prec@10 Google prec@10 Scatterplot of Precision at Top 10 Documents Figure 4: Precision at top 10 documents of UCAIR and Google is consistently and considerably better than that of Google at all levels of recall. CONCLUSIONS In this paper, we studied how to exploit implicit user modeling to intelligently personalize information retrieval and improve search accuracy. Unlike most previous work, we emphasize the use of immediate search context and implicit feedback information as well as eager updating of search results to maximally benefit a user. We presented a decision-theoretic framework for optimizing interactive information retrieval based on eager user model updating, in which the system responds to every action of the user by choosing a system action to optimize a utility function. We further propose specific techniques to capture and exploit two types of implicit feedback information: (1) identifying related immediately preceding query and using the query and the corresponding search results to select appropriate terms to expand the current query, and (2) exploiting the viewed document summaries to immediately rerank any documents that have not yet been seen by the user. Using these techniques, we develop a client-side web search agent (UCAIR) on top of a popular search engine (Google). Experiments on web search show that our search agent can improve search accuracy over 830 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 UCAIR prec@20 Google prec@20 Scatterplot of Precision at Top 20 documents Figure 5: Precision at top 20 documents of UCAIR and Google 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 recall precision Precision-Recall curves Google Result UCAIR Result Figure 6: Precision at top 20 result of UCAIR and Google Google. Since the implicit information we exploit already naturally exists through user interactions, the user does not need to make any extra effort. The developed search agent thus can improve existing web search performance without any additional effort from the user. ACKNOWLEDGEMENT We thank the six participants of our evaluation experiments. This work was supported in part by the National Science Foundation grants IIS-0347933 and IIS-0428472. REFERENCES [1] S. M. Beitzel, E. C. Jensen, A. Chowdhury, D. Grossman, and O. Frieder. Hourly analysis of a very large topically categorized web query log. In Proceedings of SIGIR 2004, pages 321328, 2004. [2] C. Clarke, N. Craswell, and I. Soboroff. Overview of the TREC 2004 terabyte track. In Proceedings of TREC 2004, 2004. [3] M. Claypool, P. Le, M. Waseda, and D. Brown. Implicit interest indicators. In Proceedings of Intelligent User Interfaces 2001, pages 3340, 2001. [4] N. Craswell, D. Hawking, R. Wilkinson, and M. Wu. Overview of the TREC 2003 web track. In Proceedings of TREC 2003, 2003. [5] W. B. Croft, S. Cronen-Townsend, and V. Larvrenko. Relevance feedback and personalization: A language modeling perspective. In Proeedings of Second DELOS Workshop: Personalisation and Recommender Systems in Digital Libraries, 2001. [6] Google Personalized. http://labs.google.com/personalized. [7] D. Hawking, N. Craswell, P. B. Thistlewaite, and D. Harman. Results and challenges in web search evaluation. Computer Networks, 31(11-16):13211330, 1999. [8] X. Huang, F. Peng, A. An, and D. Schuurmans. Dynamic web log session identification with statistical language models. Journal of the American Society for Information Science and Technology, 55(14):12901303, 2004. [9] G. Jeh and J. Widom. Scaling personalized web search. In Proceedings of WWW 2003, pages 271279, 2003. [10] T. Joachims. Optimizing search engines using clickthrough data. In Proceedings of SIGKDD 2002, pages 133142, 2002. [11] D. Kelly and J. Teevan. Implicit feedback for inferring user preference: A bibliography. SIGIR Forum, 37(2):1828, 2003. [12] J. Lafferty and C. Zhai. Document language models, query models, and risk minimization for information retrieval. In Proceedings of SIGIR'01, pages 111119, 2001. [13] T. Lau and E. Horvitz. Patterns of search: Analyzing and modeling web query refinement. In Proceedings of the Seventh International Conference on User Modeling (UM), pages 145 152, 1999. [14] V. Lavrenko and B. Croft. Relevance-based language models. In Proceedings of SIGIR'01, pages 120127, 2001. [15] M. Mitra, A. Singhal, and C. Buckley. Improving automatic query expansion. In Proceedings of SIGIR 1998, pages 206214, 1998. [16] My Yahoo! http://mysearch.yahoo.com. [17] G. Nunberg. As google goes, so goes the nation. New York Times, May 2003. [18] S. E. Robertson. The probability ranking principle in i. Journal of Documentation, 33(4):294304, 1977. [19] J. J. Rocchio. Relevance feedback in information retrieval. In The SMART Retrieval System: Experiments in Automatic Document Processing, pages 313323. Prentice-Hall Inc., 1971. [20] G. Salton and C. Buckley. Improving retrieval performance by retrieval feedback. Journal of the American Society for Information Science, 41(4):288297, 1990. [21] G. Salton and M. J. McGill. Introduction to Modern Information Retrieval. McGraw-Hill, 1983. [22] X. Shen, B. Tan, and C. Zhai. Context-sensitive information retrieval using implicit feedback. In Proceedings of SIGIR 2005, pages 4350, 2005. [23] X. Shen and C. Zhai. Exploiting query history for document ranking in interactive information retrieval (Poster). In Proceedings of SIGIR 2003, pages 377378, 2003. [24] A. Singhal. Modern information retrieval: A brief overview. Bulletin of the IEEE Computer Society Technical Committee on Data Engineering, 24(4):3543, 2001. [25] K. Sugiyama, K. Hatano, and M. Yoshikawa. Adaptive web search based on user profile constructed without any effort from users. In Proceedings of WWW 2004, pages 675684, 2004. [26] E. Volokh. Personalization and privacy. Communications of the ACM, 43(8):8488, 2000. [27] R. W. White, J. M. Jose, C. J. van Rijsbergen, and I. Ruthven. A simulated study of implicit feedback models. In Proceedings of ECIR 2004, pages 311326, 2004. [28] J. Xu and W. B. Croft. Query expansion using local and global document analysis. In Proceedings of SIGIR 1996, pages 411, 1996. [29] C. Zhai and J. Lafferty. Model-based feedback in KL divergence retrieval model. In Proceedings of the CIKM 2001, pages 403410, 2001. 831

user model;interactive retrieval;personalized search;information retrieval systems;user modelling;implicit feedback;retrieval accuracy;clickthrough information;UCAIR

106

Improvements of TLAESA Nearest Neighbour Search Algorithm and Extension to Approximation Search

Nearest neighbour (NN) searches and k nearest neighbour (k-NN) searches are widely used in pattern recognition and image retrieval. An NN (k-NN) search finds the closest object (closest k objects) to a query object. Although the definition of the distance between objects depends on applications, its computation is generally complicated and time-consuming. It is therefore important to reduce the number of distance computations. TLAESA (Tree Linear Approximating and Eliminating Search Algorithm) is one of the fastest algorithms for NN searches. This method reduces distance computations by using a branch and bound algorithm. In this paper we improve both the data structure and the search algorithm of TLAESA. The proposed method greatly reduces the number of distance computations. Moreover, we extend the improved method to an approximation search algorithm which ensures the quality of solutions. Experimental results show that the proposed method is efficient and finds an approximate solution with a very low error rate.

Introduction NN and k-NN searches are techniques which find the closest object (closest k objects) to a query object from a database. These are widely used in pattern recognition and image retrieval. We can see examples of their applications to handwritten character recognition in (Rico-Juan & Mico 2003) and (Mico & Oncina 1998), and so on. In this paper we consider NN (k-NN) algorithms that can work in any metric space. For any x, y, z in a metric space, the distance function d(, ) satisfies the following properties: d(x, y) = 0 x = y, d(x, y) = d(y, x), d(x, z) d(x, y) + d(y, z). Although the definition of the distance depends on applications, its calculation is generally complicated and time-consuming. We particularly call the calculation of d(, ) a distance computation. Copyright c 2006, Australian Computer Society, Inc. This paper appeared at Twenty-Ninth Australasian Computer Science Conference (ACSC2006), Hobart, Tasmania, Australia, January 2006. Conferences in Research and Practice in Information Technology, Vol. 48. Vladimir Estivill-Castro and Gill Dobbie, Ed. Reproduction for academic, not-for profit purposes permitted provided this text is included. For the NN and k-NN searches in metric spaces, some methods that can manage a large set of objects efficiently have been introduced(Hjaltason & Samet 2003). They are categorized into two groups. The methods in the first group manage objects with a tree structure such as vp-tree(Yianilos 1993), M-tree (Ciaccia, Patella & Zezula 1997), sa-tree (Navarro 2002) and so forth. The methods in the second group manage objects with a distance matrix, which stores the distances between objects. The difference between two groups is caused by their approaches to fast searching. The former aims at reducing the com-putational tasks in the search process by managing objects effectively. The latter works toward reducing the number of distance computations because generally their costs are higher than the costs of other calculations. In this paper we consider the latter approach . AESA (Approximating and Eliminating Search Algorithm)(Vidal 1986) is one of the fastest algorithms for NN searches in the distance matrix group. The number of distance computations is bounded by a constant, but the space complexity is quadratic. LAESA (Linear AESA)(Mico, Oncina & Vidal 1994) was introduced in order to reduce this large space complexity. Its space complexity is linear and its search performance is almost the same as that of AESA. Although LAESA is more practical than AESA, it is impractical for a large database because calculations other than distance computations increase. TLAESA (Tree LAESA)(Mico, Oncina & Carrasco 1996) is an improvement of LAESA and reduces the time complexity to sublinear. It uses two kinds of data structures: a distance matrix and a binary tree, called a search tree. In this paper, we propose some improvements of the search algorithm and the data structures of TLAESA in order to reduce the number of distance computations. The search algorithm follows the best first algorithm. The search tree is transformed to a multiway tree from a binary tree. We also improve the selection method of the root object in the search tree. These improvements are simple but very effective . We then introduce the way to perform a k-NN search in the improved TLAESA. Moreover, we propose an extension to an approximation search algorithm that can ensure the quality of solutions. This paper is organized as follows. In section 2, we describe the details of the search algorithm and the data structures of TLAESA. In section 3, we propose some improvements of TLAESA. In section 4, we present an extension to an approximation search algorithm. In section 5, we show some experimental results. Finally, in section 6, we conclude this paper. Figure 1: An example of the data structures in TLAESA. TLAESA TLAESA uses two kinds of data structures: the distance matrix and the search tree. The distance matrix stores the distances from each object to some selected objects. The search tree manages hierarchically all objects. During the execution of the search algorithm, the search tree is traversed and the distance matrix is used to avoid exploring some branches. 2.1 Data Structures We explain the data structures in TLAESA. Let P be the set of all objects and B be a subset consisting of selected objects called base prototypes. The distance matrix M is a two-dimensional array that stores the distances between all objects and base prototypes. The search tree T is a binary tree such that each node t corresponds to a subset S t P . Each node t has a pointer to the representative object p t S t which is called a pivot, a pointer to a left child node l, a pointer to a right child node r and a covering radius r t . The covering radius is defined as r t = max pS t d(p, p t ). (1) The pivot p r of r is defined as p r = p t . On the other hand, the pivot p l of l is determined so that p l = argmax pS t d(p, p t ). (2) Hence, we have the following equality: r t = d(p t , p l ). (3) S t is partitioned into two disjoint subsets S r and S l as follows: S r = {p S t |d(p, p r ) &lt; d(p, p l )}, S l = S t - S r . (4) Note that if t is a leaf node, S t = {p t } and r t = 0. Fig. 1 shows an example of the data structures. 2.2 Construction of the Data Structures We first explain the construction process of the search tree T . The pivot p t of the root node t is randomly selected and S t is set to P . The pivot p l of the left child node and the covering radius r t are defined by Eqs. (2) and (3). The pivot p r of the right child node is set to p t . S t is partitioned into S r and S l by Eq. (4). These operations are recursively repeated until |S t | = 1. The distance matrix M is constructed by selecting base prototypes. This selection is important because Figure 2: Lower bound. base prototypes are representative objects which are used to avoid some explorations of the tree. The ideal selection of them is that each object is as far away as possible from other objects. In (Mico et al. 1994), a greedy algorithm is proposed for this selection. This algorithm chooses an object that maximizes the sum of distances from the other base prototypes which have already been selected. In (Mico & Oncina 1998), another algorithm is proposed, which chooses an object that maximizes the minimum distance to the preselected base prototypes. (Mico & Oncina 1998) shows that the latter algorithm is more effective than the former one. Thus, we use the later algorithm for the selection of base prototypes. The search efficiency depends not only on the selection of base prototypes but also on the number of them. There is a trade-off between the search efficiency and the size of distance matrix, i.e. the memory capacity. The experimental results in (Mico et al. 1994) show that the optimal number of base prototypes depends on the dimensionality dm of the space. For example, the optimal numbers are 3, 16 and 24 if dm = 2, 4 and 8, respectively. The experimental results also show that the optimal number does not depend on the number of objects. 2.3 Search Algorithm The search algorithm follows the branch and bound strategy. It traverses the search tree T in the depth first order. The distance matrix M is referred whenever each node is visited in order to avoid unnecessary traverse of the tree T . The distance are computed only when a leaf node is reached. Given a query object q, the distance between q and the base prototypes are computed. These results are stored in an array D. The object which is the closest to q in B is selected as the nearest neighbour candidate p min , and the distance d(q, p min ) is recorded as d min . Then, the traversal of the search tree T starts at the root node. The lower bound for the left child node l is calculated whenever each node t is reached if it is not a leaf node. The lower bound of the distance between q and an object x is defined as g x = max bB |d(q, b) - d(b, x)|. (5) See Fig. 2. Recall that d(q, b) was precomputed before the traversals and was stored in D. In addition, the value d(b, x) was also computed during the construction process and stored in the distance matrix M . Therefore, g x is calculated without any actual distance computations. The lower bound g x is not actual distance d(q, x). Thus, it does not ensure that the number of visited nodes in the search becomes minimum . Though, this evaluation hardly costs, hence it is possible to search fast. The search process accesses the left child node l if g p l g p r , or the right child node r if g p l &gt; g p r . When a leaf node is reached, the distance is computed and both p min and d min are updated if the distance is less than d min . q p min p t r t S t Figure 3: Pruning Process. procedure NN search(q) 1: t root of T 2: d min = , g p t = 0 3: for b B do 4: D[b] = d(q, b) 5: if D[b] &lt; d min then 6: p min = b, d min = D[b] 7: end if 8: end for 9: g p t = max bB |(D[b] - M [b, p t ])| 10: search(t, g p t , q, p min , d min ) 11: return p min Figure 4: Algorithm for an NN search in TLAESA. We explain the pruning process. Fig. 3 shows the pruning situation. Let t be the current node. If the inequality d min + r t &lt; d(q, p t ) (6) is satisfied, we can see that no object exists in S t which is closer to q than p min and the traversal to node t is not necessary. Since g p t d(q, p t ), Eq. (6) can be replaced with d min + r t &lt; g p t . (7) Figs. 4 and 5 show the details of the search algorithm(Mico et al. 1996). Improvements of TLAESA In this section, we propose some improvements of TLAESA in order to reduce the number of distance computations. 3.1 Tree Structure and Search Algorithm If we can evaluate the lower bounds g in the ascending order of their values, the search algorithm runs very fast. However, this is not guaranteed in TLAESA since the evaluation order is decided according to the tree structure. We show such an example in Fig. 6. In this figure, u, v and w are nodes. If g p v &lt; g p w , it is desirable that v is evaluated before w. But, if g p v &gt; g p u , w might be evaluated before v. We propose the use of a multiway tree and the best first order search instead of a binary tree and the depth first search. During the best first search process, we can traverse preferentially a node whose subset may contain the closest object. Moreover, we can evaluate more nodes at one time by using of the multiway tree. The search tree in TLAESA has many nodes which have a pointer to the same object. In the proposed structure, we treat such nodes as one node. Each node t corresponds to a subset S t P and has a pivot p t , a covering radius r t = max pS t d(p, p t ) and pointers to its children nodes. procedure search(t, g p t , q, p min , d min ) 1: if t is a leaf then 2: if g p t &lt; d min then 3: d = d(q, p t ) {distance computation} 4: if d &lt; d min then 5: p min = p t , d min = d 6: end if 7: end if 8: else 9: r is a right child of t 10: l is a left child of t 11: g p r = g p t 12: g p l = max bB |(D[b] - M [b, p t ])| 13: if g p l &lt; g p r then 14: if d min + r l &gt; g p l then 15: search(l, g p l , p min , d min ) 16: end if 17: if d min + r r &gt; g p r then 18: search(r, g p r , p min , d min ) 19: end if 20: else 21: if d min + r r &gt; g p r then 22: search(r, g p r , p min , d min ) 23: end if 24: if d min + r l &gt; g p l then 25: search(l, g p l , p min , d min ) 26: end if 27: end if 28: end if Figure 5: A recursive procedure for an NN search in TLAESA. Figure 6: A case in which the search algorithm in TLAESA does not work well. We show a method to construct the tree structure in Fig. 7. We first select randomly the pivot p t of the root node t and set S t to P . Then we execute the procedure makeTree(t, p t , S t ) in Fig. 7. We explain the search process in the proposed structure. The proposed method maintains a priority queue Q that stores triples (node t, lower bound g p t , covering radius r t ) in the increasing order of g p t - r t . Given a query object q, we calculate the distances between q and base prototypes and store their values in D. Then the search process starts at the root of T . The following steps are recursively repeated until Q becomes empty. When t is a leaf node, the distance d(q, p t ) is computed if g p t &lt; d min . If t is not a leaf node and its each child node t satisfies the inequality g p t &lt; r t + d min , (8) the lower bound g p t is calculated and a triple (t , g p t , r t ) is added to Q. Figs. 8 and 9 show the details of the algorithm. procedure makeTree(t, p t , S t ) 1: t new child node of t 2: if |S t | = 1 then 3: p t = p t and S t = {p t } 4: else 5: p t = argmax pS t d(p, p t ) 6: S t = {p S t |d(p, p t ) &lt; d(p, p t )} 7: S t = S t - S t 8: makeTree(t , p t , S t ) 9: makeTree(t, p t , S t ) 10: end if Figure 7: Method to construct the proposed tree structure. procedure NN search(q) 1: t root of T 2: d min = , g p t = 0 3: for b B do 4: D[b] = d(q, b) 5: if D[b] &lt; d min then 6: p min = b, d min = D[b] 7: end if 8: end for 9: g t = max bB |(D[b] - M [b, p t ])| 10: Q {(t, g p t , r t )} 11: while Q is not empty do do 12: (t, g p t , r t ) element in Q 13: search(t, g p t , q, p min , d min ) 14: end while 15: return p min Figure 8: Proposed algorithm for an NN search. 3.2 Selection of Root Object We focus on base prototypes in order to reduce node accesses. The lower bound of the distance between a query q and a base prototype b is g b = max bB |d(q, b) - d(b, b)| = d(q, b). This value is not an estimated distance but an actual distance. If we can use an actual distance in the search process , we can evaluate more effectively which nodes are close to q. This fact means that the search is efficiently performed if many base prototypes are visited in the early stage. In other words, it is desirable that more base prototypes are arranged in the upper part of the search tree. Thus, in the proposed algorithm, we choose the first base prototype b 1 as the root object . 3.3 Extension to a k-NN Search LAESA was developed to perform NN searches and (Moreno-Seco, Mico & Oncina 2002) extended it so that k-NN searches can be executed. In this section, we extend the improved TLAESA to a k-NN search algorithm. The extension is simple modifications of the algorithm described above. We use a priority queue V for storing k nearest neighbour candidates and modify the definition of d min . V stores pairs (object p, distance d(q, p)) in the increasing order of procedure search(t, g p t , q, p min , d min ) 1: if t is a leaf then 2: if g p t &lt; d min then 3: d = d(q, p t ) {distance computation} 4: if d &lt; d min then 5: p min = p t , d min = d 6: end if 7: end if 8: else 9: for each child t of t do 10: if g p t &lt; r t + d min then 11: g p t = max bB |(D[b] - M [b, p t ])| 12: Q Q {(t , g p t , r t )} 13: end if 14: end for 15: end if Figure 9: A procedure used in the proposed algorithm for an NN search. procedure k-NN search(q, k) 1: t root of T 2: d min = , g p t = 0 3: for b B do 4: D[b] = d(q, b) 5: if D[b] &lt; d min then 6: V V {(b, D[b])} 7: if |V | = k + 1 then 8: remove (k + 1)th pair from V 9: end if 10: if |V | = k then 11: (c, d(q, c)) kth pair of V 12: d min = d(q, c) 13: end if 14: end if 15: end for 16: g p t = max bB |(D[b] - M [b, p t ])| 17: Q {(t, g p t , r t )} 18: while Q is not empty do 19: (t, g p t , r t ) element in Q 20: search(t, g p t , q, V, d min , k) 21: end while 22: return k objects V Figure 10: Proposed algorithm for a k-NN search. d(q, p). d min is defined as d min = (|V | &lt; k) d(q, c) (|V | = k) (9) where c is the object of the kth pair in V . We show in Figs. 10 and 11 the details of the k-NN search algorithm. The search strategy essentially follows the algorithm in Figs. 8 and 9, but the k-NN search algorithm uses V instead of p min . (Moreno-Seco et al. 2002) shows that the optimal number of base prototypes depends on not only the dimensionality of the space but also the value of k and that the number of distance computations increases as k increases. Extension to an Approximation Search In this section, we propose an extension to an approximation k-NN search algorithm which ensures the procedure search(t, g p t , q, V, d min , k) 1: if t is a leaf then 2: if g p t &lt; d min then 3: d = d(q, p t ) {distance computation} 4: if d &lt; d min then 5: V V {(p t , d(q, p t ))} 6: if |V | = k + 1 then 7: remove (k + 1)th pair from V 8: end if 9: if |V | = k then 10: (c, d(q, c)) kth pair of V 11: d min = d(q, c) 12: end if 13: end if 14: end if 15: else 16: for each child t of t do 17: if g p t &lt; r t + d min then 18: g p t = max bB |(D[b] - M [b, p t ])| 19: Q Q {(t , g p t , r t )} 20: end if 21: end for 22: end if Figure 11: A procedure used in the proposed algorithm for a k-NN search. quality of solutions. Consider the procedure in Fig. 11. We replace the 4th line with if d &lt; d min then and the 17th line with if g t &lt; r t + d min then where is real number such that 0 &lt; 1. The pruning process gets more efficient as these conditions become tighter. The proposed method ensures the quality of solutions . We can show the approximation ratio to an optimal solution using . Let a be the nearest neighbour object and a be the nearest neighbour candidate object. If our method misses a and give a as the answer, the equation g(q, a) d(q, a ) (10) is satisfied. Then a will be eliminated from targeted objects. Since g(q, a) d(q, a), we can obtain the following equation: d(q, a ) 1 d(q, a). (11) Thus, the approximate solution are suppressed by 1 times of the optimal solution. Experiments In this section we show some experimental results and discuss them. We tested on an artificial set of random points in the 8-dimensional euclidean space. We also used the euclidean distance as the distance function. We evaluated the number of distance computations and the number of accesses to the distance matrix in 1-NN and 10-NN searches. 0 50 100 150 200 250 300 0 10 20 30 40 50 60 70 80 90 100 110 120 Number of Distance Computations Number of Base Prototypes TLAESA(1-NN) TLAESA(10-NN) Proposed(1-NN) Proposed(10-NN) Figure 12: Relation of the number of distance computations to the number of base prototypes. 1-NN 10-NN TLAESA 40 80 Proposed 25 60 Table 1: The optimal number of base prototypes. 5.1 The Optimal Number of Base Prototypes We first determined experimentally the optimal number of base prototypes. The number of objects was fixed to 10000. We executed 1-NN and 10-NN searches for various numbers of base prototypes, and counted the number of distance computations. Fig. 12 shows the results. From this figure, we chose the number of base prototypes as shown in Table. 1. We can see that the values in the proposed method are fewer than those in TLAESA. This means that the proposed method can achieve better performance with smaller size of distance matrix. We used the values in Table. 1 in the following experiments. 5.2 Evaluation of Improvements We tested the effects of our improvements described in 3.1 and 3.2. We counted the numbers of distance computations in 1-NN and 10-NN searches for various numbers of objects. The results are shown in Figs. 13 and 14. Similar to TLAESA, the number of the distance computations in the proposed method does not depend on the number of objects. In both of 1-NN and 10-NN searches, it is about 60% of the number of distance computations in TLAESA. Thus we can see that our improvements are very effective. In the search algorithms of TLAESA and the proposed methods, various calculations are performed other than distance computations. The costs of the major part of such calculations are proportional to the number of accesses to the distance matrices. We therefore counted the numbers of accesses to the distance matrices. We examined the following two cases: (i) TLAESA vs. TLAESA with the improvement of selection of the root object. (ii) Proposed method only with improvement of tree structure and search algorithm vs. proposed method only with the improvement of selection of the root object. In the case (i), the number of accesses to the distance matrix is reduced by 12% in 1-NN searches and 4.5% in 10-NN searches. In the case (ii), it is reduced by 6.8% in 1-NN searches and 2.7% in 10-NN searches. 0 10 20 30 40 50 60 70 80 90 100 0 2000 4000 6000 8000 10000 Number of Distance Computations Number of Objects TLAESA Proposed Figure 13: The number of distance computations in 1-NN searches. 0 30 60 90 120 150 180 210 240 270 300 0 2000 4000 6000 8000 10000 Number of Distance Computations Number of Objects TLAESA Proposed Figure 14: The number of distance computations in 10-NN searches. Thus we can see that this improvement about selection of the root object is effective. 5.3 Evaluation of Approximation Search We tested the performance of the approximation search algorithm. We compared the proposed method to Ak-LAESA, which is the approximation search algorithm proposed in (Moreno-Seco, Mico & Oncina 2003). Each time a distance is computed in Ak-LAESA , the nearest neighbour candidate is updated and its value is stored. When the nearest neighbour object is found, the best k objects are chosen from the stored values. In Ak-LAESA, the number of distance computations of the k-NN search is exactly the same as that of the NN search. To compare the proposed method with Ak-LAESA , we examined how many objects in the approximate solutions exist in the optimal solutions. Thus, we define the error rate E as follows: E[%] = |{x i |x i / Opt, i = 1, 2, , k}| k 100 (12) where {x 1 , x 2 , , x k } is a set of k objects which are obtained by an approximation algorithm and Opt is a set of k closest objects to the query object. Fig. 15 shows the error rate when the value of is changed in 10-NN searches. Fig. 16 also shows the relation of the number of distance computations to the value of in 10-NN searches. In the range 0.5, the proposed method shows the lower error rate than 0 10 20 30 40 50 60 70 80 90 100 0 0.2 0.4 0.6 0.8 1 Error Rate [%] Ak-LAESA Proposed Figure 15: Error rate in 10-NN searches. 0 20 40 60 80 100 120 140 160 0 0.2 0.4 0.6 0.8 1 Number of distance computations Ak-LAESA Proposed Figure 16: Relation of the number of distance computations to the value of in 10-NN searches. Ak-LAESA. In particular, the error rate of the proposed method is almost 0 in range 0.9. From two figures, we can control the error rate and the number of distance computations by changing the value of . For example, the proposed method with = 0.9 reduces abount 28.6% of distance computations and its error rate is almost 0. Then we examined the accuracy of the approximate solutions. We used = 0.5 for the proposed method because the error rate of the proposed method with = 0.5 is equal to the one of Ak-LAESA . We performed 10-NN searches 10000 times for each method and examined the distribution of kth approximate solution to kth optimal solution. We show the results in Figs. 17 and 18. In each figure, x axis represents the distance between a query object q and the kth object in the optimal solution. y axis shows the distance between q and the kth object in the approximate solution. The point near the line y = x represents that kth approximate solution is very close to kth optimal solution. In Fig. 17, many points are widely distributed. In the worst case, some appriximate solutions reach about 3 times of the optimal solution. From these figures, we can see that the accuracy of solution by the proposed method is superior to the one by Ak-LAESA. We also show the result with = 0.9 in Fig. 19. Most points exist near the line y = x. Though Ak-LAESA can reduce drastically the number of distance computations, its approximate solutions are often far from the optimal solutions. On the other hand, the proposed method can reduce the number of distance computations to some extent with 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0.2 0.4 0.6 0.8 Distance to the k th Approximate Solution Distance to the k th Optimal Solution Figure 17: The distribution of the approximate solution by Ak-LAESA to the optimal solution. 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0.2 0.4 0.6 0.8 Distance to the k th Approximate Solution Distance to the k th Optimal Solution Figure 18: The distribution the approximate solution by the proposed method with = 0.5 to the optimal solution. very low error rate. Moreover, the accuracy of its approximate solutions is superior to that of Ak-LAESA. Conclusions In this paper, we proposed some improvements of TLAESA. In order to reduce the number of distance computations in TLAESA, we improved the search algorithm to best first order from depth first order and the tree structure to a multiway tree from a binary tree. In the 1-NN searches and 10-NN searches in a 8-dimensional space, the proposed method reduced about 40% of distance computations. We then proposed the selection method of root object in the search tree. This improvement is very simple but is effective to reduce the number of accesses to the distance matrix. Finally, we extended our method to an approximation k-NN search algorithm that can ensure the quality of solutions. The approximate solutions of the proposed method are suppressed by 1 times of the optimal solutions. Experimental results show that the proposed method can reduce the number of distance computations with very low error rate by selecting the appropriate value of , and that the accuracy of the solutions is superior to Ak-LAESA. From these viewpoints, the method presented in this paper is very effective when the distance computations are time-consuming. 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0.2 0.4 0.6 0.8 Distance to the k th Approximate Solution Distance to the k th Optimal Solution Figure 19: The distribution the approximate solution by the proposed method with = 0.9 to the optimal solution. References Ciaccia, P., Patella, M. & Zezula, P. (1997), M-tree: An efficient access method for similarity search in metric spaces, in `Proceedings of the 23rd International Conference on Very Large Data Bases (VLDB'97)', pp. 426435. Hjaltason, G. R. & Samet, H. (2003), `Index-driven similarity search in metric spaces', ACM Transactions on Database Systems 28(4), 517580. Mico, L. & Oncina, J. (1998), `Comparison of fast nearest neighbour classifiers for handwritten character recognition', Pattern Recognition Letters 19(3-4), 351356. Mico, L., Oncina, J. & Carrasco, R. C. (1996), `A fast branch & bound nearest neighbour classifier in metric spaces', Pattern Recognition Letters 17(7), 731739. Mico, M. L., Oncina, J. & Vidal, E. (1994), `A new version of the nearest-neighbour approximating and eliminating search algorithm (AESA) with linear preprocessing time and memory require-ments' , Pattern Recognition Letters 15(1), 917. Moreno-Seco, F., Mico, L. & Oncina, J. (2002), `Extending LAESA fast nearest neighbour algorithm to find the k-nearest neighbours', Lecture Notes in Computer Science - Lecture Notes in Artificial Intelligence 2396, 691699. Moreno-Seco, F., Mico, L. & Oncina, J. (2003), `A modification of the LAESA algorithm for ap-proximated k-NN classification', Pattern Recognition Letters 24(1-3), 4753. Navarro, G. (2002), `Searching in metric spaces by spatial approximation', The VLDB Journal 11(1), 2846. Rico-Juan, J. R. & Mico, L. (2003), `Comparison of AESA and LAESA search algorithms using string and tree-edit-distances', Pattern Recognition Letters 24(9-10), 14171426. Vidal, E. (1986), `An algorithm for finding nearest neighbours in (approximately) constant average time', Pattern Recognition Letters 4(3), 145157. Yianilos, P. N. (1993), Data structures and algorithms for nearest neighbor search in general metric spaces, in `SODA '93: Proceedings of the fourth annual ACM-SIAM Symposium on Discrete algorithms', pp. 311321.

Approximation Search;TLAESA;Distance Computaion;k Nearest Neighbour Search;Nearest Neighbour Search

107

Improving the Static Analysis of Embedded Languages via Partial Evaluation

Programs in embedded languages contain invariants that are not automatically detected or enforced by their host language. We show how to use macros to easily implement partial evaluation of embedded interpreters in order to capture invariants encoded in embedded programs and render them explicit in the terms of their host language . We demonstrate the effectiveness of this technique in improving the results of a value flow analysis.

1. One Language, Many Languages Every practical programming language contains small programming languages. For example, C's printf [18] supports a string-based output formatting language, and Java [3] supports a declarative sub-language for laying out GUI elements in a window. PLT Scheme [9] offers at least five such languages: one for formatting console output; two for regular expression matching; one for sending queries to a SQL server; and one for laying out HTML pages. In many cases, though not always, programs in these embedded special-purpose programming languages are encoded as strings. Library functions consume these strings and interpret them. Often the interpreters consume additional arguments, which they use as inputs to the little programs. Take a look at this expression in PLT Scheme: (regexp-match &quot;http://([a-z.]*)/([a-z]*)/&quot; line) The function regexp-match is an interpreter for the regular expression language. It consumes two arguments: a string in the regular expression language, which we consider a program, and another string, which is that program's input. A typical use looks like the example above. The first string is actually specified at the call site, while the second string is often given by a variable or an expression that reads from an input port. The interpreter attempts to match the regular expression and the second string. In PLT Scheme, the regular expression language allows programmers to specify subpatterns via parentheses. Our running example contains two such subexpressions: ([a-z.]*) and ([a-z]*) . If the regular expression interpreter fails to match the regular expression and the string, it produces false ( #f ); otherwise it produces a list with n + 1 elements: the first one for the overall match plus one per subexpression. Say line stands for &quot;http://aaa.bbb.edu/zzz/&quot; In this case, the regular expression matches the string, and regexp-match produces the list (list &quot;http://aaa.bbb.edu/zzz/&quot; &quot;aaa.bbb.edu&quot; &quot;zzz&quot;) The rest of the Scheme program extracts the pieces from this list and computes with them. The regexp-match expression above is a simplified excerpt from the PLT Web Server [12]. Here is a slightly larger fragment: (let ([r (regexp-match &quot;http://([a-z.]*)/([a-z]*)/&quot; line)]) (if r (process-url (third r) (dispatch (second r))) (log-error line))) Notice how the then-clause of the if -expression extracts the second 16 and third elements from r without any checks to confirm the length of the list. After all, the programmer knows that if r is not false, then it is a list of three elements. The embedded program says so; it is a regular expression and contains two subexpressions. Unfortunately, the static analysis tools for PLT Scheme cannot reason on both levels. MrFlow [20], a static debugger, uses a constraint-based analysis [22], a version of set-based analysis [2, 13, 10], to analyze the program and discover potential errors . If it finds one it can draw a flow graph from the source of the bad value to the faulty primitive operation. For the let -expression above, MrFlow finds that both (second r) and (third r) may raise runtime errors because r may not contain enough elements. In this paper, we show how using Scheme macros to partially evaluate calls to embedded interpreters such as regexp-match greatly increases the precision of the static analysis. Since we use macros, library designers can easily implement the partial evaluation, rather than relying on the host language implementor as they must for ad-hoc solutions. In Section 2 we give a brief overview of set-based analysis and MrFlow . In the next section we explain three examples of embedded languages and the problems they cause for MrFlow's static analysis . We then present in Section 4 our general approach to solving those problems, based on macros. An overview of the macro system we use is given in Section 5. Section 6 then presents a general technique for translating embedded interpreters into macros. In Section 7, we explain the properties of the static analysis that enable it to find more results in partially evaluated code. Finally, in Section 8, we show how partially evaluating Scheme programs that contain embedded programs helps MrFlow in our three examples. Section 9 presents related work and we conclude in Section 10. Set-Based Analysis To explain how the results of a static analysis can be improved by using partial evaluation of embedded languages, we first need to describe such an analysis. MrFlow, a static analyzer for DrScheme, uses a set-based value flow analysis to compute an approximation of the values that each subexpression of a program might evaluate to at runtime [22]. The approximation computed for each expression is a set of abstract values that can be displayed on demand. The debugger can also draw arrows showing the flow of values through the program. Figure 1 displays an example of analyzing a simple program. In the box next to the term 3 is the abstract value for that term, meaning that at runtime the term 3 might evaluate to the value 3. The arrow starting from the term 3 shows that at runtime the value 3 might flow into the argument x of the function f and from there flow into the reference to the variable x in the body of f . There is a second reference to x in f --the corresponding arrow is not shown in this example. In the box next to the call to the Scheme primitive gcd is the abstract value for the result of that call. Since the analysis never tries to evaluate expressions, it uses the abstract value integer to represent the result of the primitive call, if any, which is a conservative approximation of the actual value that that call might compute at runtime. The biggest box displays the type of the adjacent if -expression, which is the union of the integer abstract value computed by the gcd primitive and of the string "hello". Arrows show that the result of the if -expression can come from both the then- and else-branches: the analysis does not attempt to apply the number? predicate to the variable x , so it conservatively assumes that both branches of the if -expression may be evaluated at runtime. Three Embedded Languages We now turn to embedded languages, which are a useful technique for establishing abstraction layers for a particular design space. Functional languages are well-suited to writing interpreters for embedded languages, in which the higher-level embedded language is implemented as a set of functions in the general purpose host language and has access to all of its features [15, 16, 24]. But these abstractions come at a cost for program analysis. In particular, tools built to examine programs of the host language cannot derive information for the programs in the embedded languages because they do not understand the semantics of those languages. In this section we demonstrate three examples of practical embedded languages for Scheme and show their negative effects on static analysis. In the first example, properties of the embedded language create the possibility of errors that can go undetected by the analysis . In the next two examples, undetected properties lead to analyses that are too conservative, resulting in many false positives; that is, the analysis reports errors that can never actually occur. 3.1 Format Strings The PLT Scheme library provides a format function, similar to C's sprintf , which generates a string given a format specifier and a variable number of additional arguments. The format specifier is a string containing some combination of literal text and formatting tags. These tags are interpreted along with the remaining arguments to construct a formatted string. The format function is thus an interpreter for the format specifier language. The format specifier is a program in this language and the additional arguments are its inputs. To construct its output, the format function requires the number of extra arguments to match the number of format tags, and these arguments must be of the appropriate type. Consider the example of displaying an ASCII character and its encoding in hexadecimal: (format &quot;~c = 0x~x&quot; c n) In this example, the format specifier, which contains the format tags &quot;~c&quot; and &quot;~x&quot; and some literal text, expects to consume exactly two arguments. These arguments must be a character and an integer , respectively. An incorrect number of arguments or a type mismatch results in a runtime error. Unfortunately analysis tools for Scheme such as MrFlow have no a priori knowledge of the semantics of embedded languages. The analysis cannot infer any information about the dependencies between the contents of the format string and the rest of the arguments without knowledge of the syntax and semantics of the format language . As a result the analysis cannot predict certain categories of runtime errors, as shown in Figure 2. The application of format is not underlined as an error, even though its arguments appear in the wrong order and the analysis correctly computes the types of both c and n . 17 Figure 1. Analyzing a simple program with MrFlow. 3.2 Regular Expressions Regular expressions are used in all kinds of Scheme programs. The language of regular expression patterns is embedded in Scheme as strings. A library of functions interpret these strings as programs that consume additional arguments as input strings and return either a list of matched subpatterns or #f to indicate failure. Consider again the excerpt from the PLT Web Server from Section 1. Programmers know that if the match succeeds, then the result list contains exactly three elements: the result of the entire match, and the results of the two subpattern matches. Again the analysis is unable to discover this invariant on its own. Figure 3 shows the results of analyzing the sample code with MrFlow. The list accessors second and third are underlined in red because the analysis cannot prove that their arguments are sufficiently long lists. Programmers then must either go through each of these false positives and prove for themselves that the errors can never occur, or else learn to ignore some results of MrFlow. Neither option is desirable . The former creates more work for the programmer, rather than less; the latter is unsafe and easily leads to overlooked errors. 3.3 SchemeQL SchemeQL [28] is an embedded language for manipulating relational databases in Scheme. Unlike the string-based format language , SchemeQL programs consist of special forms directly embedded inside Scheme. The SchemeQL implementation provides a set of macros that recognize these forms and expand them into Scheme code. A typical database query in SchemeQL might look like this: (direct-query (name age phone) directory) corresponding to the SQL statement SELECT name, age, phone FROM directory The result of executing a query is a lazy stream representing a cursor over the result set from the database server. Each element in the stream is a list of values representing a single row of the result set. The cursor computes the rows by need when a program selects the next sub-stream. Programmers know that the number of elements in each row of a cursor is equal to the number of columns in the original request. Our analysis, however, cannot discover this fact automatically. Figure 4 shows the results of an analysis of a SchemeQL query in the context of a trivial Scheme program. The example query consists of exactly three columns, and the code references the third element of the first row. This operation can never fail, but the analysis is unable to prove this. Instead, it conservatively computes that row is a list of unknown length: rec-type describes a recursive abstract value, which in the present case is the union of null and a pair consisting of any value (top) and the abstract value itself, creating a loop in the abstract value that simulates all possible list lengths. MrFlow therefore mistakenly reports an error by underlining the primitive third in red, since, according to the analysis, row might have fewer than three elements at runtime. Macros for Partial Evaluation All the embedded languages presented in the previous section have one thing in common: they can encode invariants that are not visible to any analysis of the general purpose language in which they are embedded. These invariants can be exposed to analyses in two ways: by extending the analyses in an ad-hoc manner for each embedded language so that they understand its semantics, or by partially evaluating the embedded interpreters with regard to the embedded programs to make the invariants in the embedded programs explicit as invariants in the host language, whenever possible. The first solution requires modifying each analysis to support each embedded language. The second solution can simply be implemented from within the host language through the old Lisp trick of using "compiler macros" [25] as a light-weight partial evaluation mechanism. In the present case, instead of using partial evaluation to optimize programs for speed, we use it to increase the precision of program analyses. While Lisp's compiler macros are different from regular Lisp macros, Scheme's macro system is powerful enough that the equivalent of Lisp's compiler macros can be implemented as regular Scheme macros. The partial evaluation of embedded interpreters then simply involves replacing the libraries of functions imple-18 Figure 2. Imprecise analysis of the format primitive. Figure 3. Imprecise analysis of regexp-match . Figure 4. Imprecise analysis of a SchemeQL query. 19 menting the interpreters with libraries of semantically equivalent macros 1 . This has the additional advantage that it can be done by the author of the library of functions, as opposed to the compiler's or analyzer's implementor in the case of ad-hoc extensions. Of course, the partial evaluation of embedded interpreters is only possible when their input programs are known statically. For example , it is not possible to expand a call to format if the formatting string given as its first argument is computed at runtime. The programmer therefore makes a trade-off between the precision of analyses and how dynamic the code can be. In practice, though, the embedded programs are often specified statically in user code. Combined with the simplicity of implementing partial evaluation with macros, this makes for a useful technique for improving the precision of analyses at a low cost. In the next two sections, we describe some of the important features of the Scheme macro system and then explain how we make use of this system to partially evaluate the interpreters of these embedded languages to improve the results of static analysis. Macros in Scheme Scheme has a powerful macro system for extending the language with derived expression forms that can be rewritten as expressions in the core language. Macros serve as a means of syntactic abstraction . Programmers can generalize syntactic patterns in ways that are not possible with functional abstraction. This technology also provides a hook into the standard compiler tool chain by allowing programmers to implement additional program transformations before compilation. In this section we describe the basics of standard Scheme macros and introduce identifier macros, a generalization of the contexts in which macros can be matched. 5.1 Rule-Based Macros The define-syntax special form allows the programmer to extend Scheme with derived expression forms. Before compilation or execution of a Scheme program, all occurrences of these derived forms are replaced with their specified expansions. The syntax-rules form specifies macro expansions as rewrite rules. Consider the following simple macro, which defines a short-circuit logical or as a derived form: (define-syntax or (syntax-rules () [(or e1 e2) (let ([tmp e1]) (if tmp tmp e2))])) The macro defines a single rewrite rule, consisting of a pattern and a template. The pattern matches the or keyword in operator position followed by two pattern variables e1 and e2 , each matching an arbitrary subexpression in argument position. The template directs the macro expansion to replace occurrences of the matched pattern with a let -expression constructed from the matched subexpressions . 1 The transformation is not strictly speaking partial evaluation: the reductions performed by the macros are not exactly the ones performed by the embedded interpreters. However, the macros share the techniques and issues of partial evaluation since they simulate parts of the interpreters, and it is therefore useful to describe them as such. Notice that this or form cannot be defined as a regular function in Scheme. The second argument is only evaluated if the first argument evaluates to false. Since Scheme has a strict evaluation semantics, a functional or would necessarily evaluate both of its arguments before computing a result. Controlling the evaluation of expressions is an important use of Scheme macros. Macros can also abstract over other syntactic forms in ways that functions cannot by expanding into second-class language constructs such as define . 5.2 Lexical Scope Macros written with the standard Scheme syntax-rules mechanism are both hygienic and referentially transparent. Hygienic macro expansion guarantees that binding forms inside the definition of the macro template do not capture free variables in macro arguments. Consider the following use of our or macro: 2 (or other tmp) (let ([tmp 1 other]) (if tmp 1 tmp 1 tmp)) Hygienic expansion automatically renames the variable bound inside the expanded macro template to avoid capturing the free variable in the macro argument. Referential transparency complements hygiene by ensuring that free variables inside the macro template cannot be captured by the context of the macro call site. For example, if the context that invokes or rebinds the if name, the expansion algorithm renames the binding in the caller's context to avoid capturing the variable used in the template body: (let ([if 3]) (or if #f)) (let ([if 1 3]) (let ([tmp if 1 ]) (if tmp tmp #f))) The combination of hygiene and referential transparency produces macros that are consistent with Scheme's rules of lexical scope and can be invoked anywhere in a program without the danger of unexpected variable capture. 3 5.3 Identifier Macros The syntax-rules form only matches expressions in which the macro name occurs in "application position," i.e., as the operator in an application expression. References to a syntax-rules macro in other contexts result in syntax errors: (fold or #f ls) syntax error PLT Scheme's syntax-id-rules form is similar to syntax-rules but matches occurrences of the macro keyword in arbitrary expression contexts: in operator position, operand position, or as the target of an assignment. 2 We use the convention of representing macro expansion with a double-arrow ( ) and ordinary (runtime) evaluation with a single-arrow ( ). 3 Macros can also be defined in and exported from modules in PLT Scheme [11]. 20 The following macro demonstrates a hypothetical use of syntax-id-rules : (define-syntax clock (syntax-id-rules (set!) [(set! clock e) (set-clock! e)] [(clock e) (make-time-stamp (get-clock) e)] [clock (get-clock)])) The list of identifiers following syntax-id-rules , which was empty in our previous examples, now includes the set! identifier , indicating that set! is to be treated as a keyword rather than a pattern variable. The first rewrite rule matches expressions in which the clock name occurs as the target of an assignment. The second rule is familiar, matching the macro in application position. The final rule matches the identifier clock in any context not matched by the previous two rules. In addition to the usual application context, we can use the clock macro in an argument position: (+ clock 10) (+ (get-clock) 10) or as a set! target: (set! clock 5) (set-clock! 5) 5.4 Programmatic Macros The language of patterns and templates recognized by syntax-rules and syntax-id-rules is actually a special case of Scheme macros. In general, the define-syntax form binds a transformer procedure (define-syntax name (lambda (stx) etc . The argument to the transformer procedure is a syntax object, which is similar to an S-expression representing quoted code, but which also encapsulates information about the lexical context of the code, such as source file location and variable bindings. This context information is essential in allowing DrScheme's language tools to trace errors and binding relationships back to the original source location in the user's code where a macro is invoked. Because syntax objects are so similar to quoted data, the standard library includes the syntax-object-&gt;datum procedure, which strips the lexical information from a syntax object and returns its corresponding datum. For example, the datum corresponding to a syntax object representing a literal number is its numeric value, the datum corresponding to an identifier is a symbol representing the identifier's name, and so on. A syntax transformer procedure accepts as its argument a syntax object representing the expression that invoked the macro, and produces a new syntax object, which the macro expansion algorithm uses to replace the original expression. All Scheme macros are syntax transformers; although the syntax-rules and syntax-id-rules forms do not use the lambda notation, they are themselves implemented as macros that expand to syntax transformer procedures. The syntax-case facility allows the construction of macros with pattern matching, as with syntax-rules and syntax-id-rules , but with arbitrary expressions in place of templates for the result expressions. For example, the above or macro would be defined as: (define-syntax or (lambda (stx) (syntax-case stx () [(or e1 e2) #'(let ([tmp e1]) (if tmp tmp e2))]))) The macro is almost the same as before, but for two refinements. First, the syntax-case form takes the argument stx explicitly, whereas syntax-rules implicitly defines a transformer procedure and operates on the procedure argument. Second, the result expression is prefixed by the syntax-quoting #' operator, which is analogous to Scheme's quote operator ' . Whereas an expression prefixed with ' evaluates to a quoted S-expression, a #' expression becomes a quoted syntax object that also includes lexical information . Similarly, the quasisyntax operator #` and unsyntax operator #, behave for syntax objects like the quasiquote and unquote operators for S-expressions, respectively. The use of arbitrary computations in the result expression allows macros to expand differently based on the results of actual computations : (define-syntax swap (lambda (stx) (syntax-case stx () [(swap a b) (if (and (identifier? #'a) (identifier? #'b)) #'(let ([tmp b]) (set! b a) (set! a tmp)) (raise-syntax-error 'swap &quot;expects identifiers&quot; stx))]))) In this example, if swap is not given identifiers as arguments, the raise-syntax-error function uses the lexical information in the stx syntax object to highlight the original swap expression in the user's code. Conditional matching can also be achieved using pattern guards, which can inspect a matched expression and determine whether to accept a match: (define-syntax swap (lambda (stx) (syntax-case stx () [(swap a b) (and (identifier? #'a) (identifier? #'b)) #'(let ([tmp b]) (set! b a) (set! a tmp))]))) The pattern guard is a new expression, inserted between the pattern and the result expressions. A guarded match only succeeds if its guard does not evaluate to false; when a guard fails, the pattern matcher falls through to attempt the next pattern in the list. Macros for Interpreters In this section, we present a general technique for specializing embedded interpreters with macros, and explain how we apply this technique to the three embedded languages described in Section 3. 21 The technique can be summarized in the following steps: 1. Write the interpreter compositionally as a module of library functions. 2. Replace the interpreter's main function with a macro that unfolds the case dispatch on the input (the embedded program) when it is known statically. 3. Default to the original function when the input is not known at compile time. Writing the interpreters compositionally serves two purposes. First, by delegating the interpretation of the program constructs that make up an embedded program to separate functions, it becomes possible to share code between the original interpreter and the macro that replaces it. This effectively limits the macro's responsibility to a simple dispatch. Second, compositionality makes it easier to guarantee that unfolding terminates, since the recursive macro calls always operate on smaller terms. 6.1 Format Strings The implementation of a string formatter involves a number of simple library functions to convert each possible type of argument to strings. Each formatting tag corresponds to one of these combinators . For example, the &quot;~c&quot; tag corresponds to a combinator, format/char , which accepts a character and converts it to a string, the &quot;~x&quot; tag corresponds to format/hex , which converts integers to their hexadecimal representation, and so forth. The string formatter then simply dispatches to these combinators based on the content of the formatting string: (define (format s . args) (cond [(string=? s &quot;&quot;) &quot;&quot;] [(string=? (substring s 0 2) &quot;~c&quot;) (string-append (format/char (car args)) (apply format (substring s 2) (cdr args)))] etc . )) The interpreter accepts the formatting string s and, based on formatting tags like &quot;~c&quot; that it finds, decomposes the string into a series of applications of the corresponding combinators to successive arguments of format (represented by args ). It reassembles the transformed pieces with the standard string-append function. In order to specialize the format interpreter, we replace it with a macro that re-uses its associated combinators: (define (format/dynamic s . args) as before ) (define-syntax format (lambda (stx) (syntax-case stx () [(format s-exp a1 a2 ...) (string? (syntax-object-&gt;datum #'s-exp)) (let ([s (syntax-object-&gt;datum #'s-exp)]) (cond [(string=? s &quot;&quot;) #'&quot;&quot;] [(string=? (substring s 0 2) &quot;~c&quot;) #`(string-append (format/char a1) (format #,(substring s 2) a2 ...))] etc . ))] [(format s-exp a1 a2 ...) #'(format/dynamic s-exp a1 a2 ...)] [format (identifier? #'format) #'format/dynamic]))) The partial evaluation works by unfolding the interpreter's top-level case dispatch on the program text. Rather than delaying the inspection of the string to runtime, the macro precomputes the result of the decomposition statically whenever the string is given as a literal. We can identify literal strings through the use of a pattern guard. More precisely, the macro can inspect the syntax object s-exp , corresponding to format 's first argument, and determine whether it can be converted to a string via syntax-object-&gt;datum . When the conversion succeeds, the pattern guard allows the match to succeed , and partial evaluation proceeds. After the macro expansion, the resulting program text consists of the application of string-append to the calls to the library functions , with no references to the interpreter: (format &quot;~c = 0x~x&quot; c n) (string-append (format/char c) &quot; = 0x&quot; (format/hex n)) In order for the replacement of the original function with a macro to be unobservable, the macro must behave exactly like the original function in all contexts. When format is applied to a dynamic formatting string, the macro defaults to the original functional implementation . Similarly, when format is passed as an argument to a higher-order function, we use the technique of identifier macros to refer to the original function. 4 6.2 Regular Expressions One of PLT Scheme's regular expression engines uses the two-continuation model of backtracking [1]. A regular expression "matcher" is represented as a function that accepts a success continuation and a failure continuation. When a matcher succeeds in matching its input, it applies its success continuation to the accepted input, and when it fails to match, it invokes its failure continuation. This allows the interpretation of the alternation operator " | " to try each alternate pattern sequentially: an alternation matcher tries to match its first pattern with a failure continuation to try the second pattern. Thus if the first pattern fails, the matcher invokes the failure continuation, which tries the second pattern. Otherwise, the failure continuation is disregarded and the matcher applies its success continuation , which skips the second pattern and returns the result of the first match. Each of the regular expression constructions corresponds to a functional combinator that produces a matcher. These combinators can express the standard operators of regular expressions: success , failure, alternation, concatenation, and repetition (i.e., Kleene star). There is also a submatch combinator for the parenthesized subpatterns in the original regular expression. A successful regexp-match returns a list with the entire matched string followed by each submatch corresponding to a parenthesized subpattern . Any subpattern that does not match corresponds to an entry of false ( #f ) in the result list. For example, the following successful 4 The case of set! is not critical since, in PLT Scheme, imported module references cannot be the target of an assignment. 22 match contains a failed submatch: (regexp-match &quot;a((b)|(c))&quot; &quot;ac&quot;) (list &quot;ac&quot; &quot;c&quot; #f &quot;c&quot;) Regardless of the contents of the second argument, there is always exactly one element in the result list for each parenthesized subpattern in the regular expression. The submatch operator accomplishes this by wrapping a given matcher with continuations that add either the result of a successful match or false to a list of indexed submatches accumulated during the match. The initial (success ) continuation for regexp-match sorts the accumulated list of indexed submatches, adding false entries for all submatches that were never reached because of backtracking. Partial evaluation of the regular expression library works by unfolding the definitions of the combinators as well as the contents of the initial continuation. Each application of a combinator gets replaced by an application of a copy of the body of the combinator's definition. 5 The recursive code that constructs the result list in the success continuation gets expanded into an explicit chain of cons expressions: (regexp-match &quot;a((b)|(c))&quot; input) ((build-matcher input) (lambda (subs) (cons (lookup subs 0) (cons (lookup subs 1) (cons (lookup subs 2) (cons (lookup subs 3) null))))) (lambda () #f)) Since the size of the result list is known, it is possible to unfold recursive definitions, such as the initial continuation that constructs the match result, to make the structure of the result explicit. Finally, in the cases where the embedded program is not known statically , or when regexp-match is used in non-application contexts, the macro expands to the original functional definition. 6.3 SchemeQL The SchemeQL language differs from the other examples in that its programs are not embedded as strings but rather as special forms recognized by a library of macros. This means that for queries that select from a fixed set of columns, the length of cursor rows is always known statically; the column names are specified as a sequence of identifiers in the syntax of the query form. Just as the interpreters for the string-based embedded programs perform a case dispatch on the contents of program strings, the SchemeQL macros dispatch on the shape of the query expressions. The cases where partial evaluation is possible can be captured by inserting additional rules into the original library's macros. Partial evaluation of SchemeQL queries uses the same technique as for the regular expression library: the recursive function that constructs a cursor row is unfolded into an explicit chain of cons expressions . Since we know the length of the cursor row statically, the unfolding is guaranteed to terminate. 5 It is convenient to define the Kleene star operator recursively by p = (pp )| . However, this non-compositional definition leads to an infinite macro expansion, so the macro must carefully avoid unfolding such a definition. Since the SchemeQL library is implemented as macros, there is no need to capture the cases where the query forms are used in non-application contexts. Adding special cases to the existing macro does not affect its set of allowable contexts. Similarly, the cases where the row length is not known statically are already handled by the existing SchemeQL macros. Static Analysis for Scheme MrFlow's value flow analysis is an extension of an ordinary set-based closure analysis like Palsberg's [22]. For every expression in a program, MrFlow statically computes a conservative approximation of the set of values to which the expression might evaluate at runtime. From a given expression it creates a graph that simulates the flow of values inside the expression. The analysis simulates evaluation by propagating abstract values in this graph until reaching a fixed point. From the set of abstract values that propagate to a given node, the analysis reconstructs a type that is then displayed to the user through DrScheme's graphical interface. Extensions to the basic analysis include, among other things: analyzing functions that can take any number of arguments, analyzing assignments to variables ( set! ), and analyzing generative data structure definitions. MrFlow also supports all the primitives defined in R 5 RS [17]. The vast majority of these primitives are defined using a special, type-like language embedded inside the analyzer . For a given primitive, the corresponding type translates to a graph that simulates the primitive's internal flows. The analysis then proceeds just like for any other expression. The few remaining primitives need special handling because of their imperative nature ( set-car! or vector-fill! ) and are analyzed in an ad-hoc manner . By default, MrFlow analyzes the format primitive based on the following pseudo-type description: (string top *-&gt; string) The * in the *-&gt; constructor means that the primitive is a function that can take any number of arguments as input beyond the ones explicitly specified. In the present case, the function must receive a string as its first argument, followed by any number of arguments of any type (represented by the pseudo-type top ), and returns a string. Given such a description, the only errors MrFlow detects are when the primitive is given something other than a string as first argument, or if it is given no argument at all. After partial evaluation, the application of format is replaced by calls to its individual library functions such as format/char and format/hex . These functions have respectively the pseudo-types (char -&gt; string) and (integer -&gt; string) Using this more precise information, MrFlow can detect arguments to the original format call that have the wrong type. Checking that the format primitive receives the right number of arguments for a given formatting string happens during partial evaluation, so the analyzer never sees arity errors in the expanded code. Since DrScheme's syntax object system keeps track of program terms through the macro expansions [11], MrFlow is then able to trace detected errors back to the original guilty terms in the user's 23 program and flag them graphically. Arrows representing the flow of values can also be displayed interactively in terms of the original program, allowing the user to track in the program the sources of the values that triggered the errors. In essence, the only requirement for MrFlow to analyze the partially evaluated code of format is to specify the pseudo-types for the library functions introduced by the transformations, like format/char 6 . Similarly, it is enough to define pseudo-types for the functions used in the partially evaluated form of SchemeQL's query to have MrFlow automatically compute precise results without any further modifications. The partial evaluation for regular expressions is more challenging. Consider the example from Section 1: (let ([r (regexp-match &quot;http://([a-z.]*)/([a-z]*)/&quot; line)]) (if r (process-url (third r) (dispatch (second r))) (log-error))) After the call to regexp-match , the variable r can be either a list of three elements or false. Based on its conservative pseudo-type specification for regexp-match , MrFlow computes that r can be either a list of unknown length or false. This in turn triggers two errors for each of the second and third primitives: one error because the primitive might be applied to false when it expected a list, and one error because it might be applied to a list that is too short. The second kind of false positives can be removed by partially evaluating regexp-match to make the structure of the result more explicit to MrFlow, as described in Section 6.2. The analysis then determines that the primitive returns either a list of three elements or false and in turn checks that second and third are applied to a list with enough elements. Still, the possible return values of regexp-match may contain false. Indeed, false will be the value returned at runtime if the line given to regexp-match does not match the pattern. The programmer has to test for such a condition explicitly before processing the result any further. The only way for MrFlow not to show a false positive for second and third , because of the presence of this false value, is to make the analysis aware of the dependency between the test of r and the two branches of the if -expression. This form of flow-sensitive analysis for if -expressions is difficult to implement in general since there is no bound to the complexity of the tested expression . In practice, however, an appreciable proportion of these tests are simple enough that an ad-hoc solution is sufficient. In the case where the test is simply a variable reference it is enough to create two corresponding ghost variables, one for each branch of the if , establish filtering flows between the variable r and the two ghost variables, and make sure each ghost variable binds the r variable references in its respective branch of the if -expression. The filtering flows prevent the false abstract value from flowing into the then branch of the if -expression and prevent everything but the false value from flowing into the else branch. Only the combination of this flow sensitivity for if -expressions with the partial evaluation of regexp-match gives analysis results with no false positives. 6 Specifying such pseudo-types will not even be necessary once MrFlow knows how to analyze PLT Scheme contracts. This is the subject of a forthcoming paper. Once flow-sensitive analysis of if -expressions is added and pseudo-type descriptions of the necessary primitives are provided to the analysis, partial evaluation makes all the false positives described in Section 3 disappear, as we illustrate in the next section. Improvement of Static Analysis Partially evaluating format eliminates the possibility of runtime arity errors, since the macro transformations can statically check such invariants. It also allows MrFlow to detect type errors that it could not detect before, since the corresponding invariants were described only in the embedded formatting language. These invariants are now explicit at the Scheme level in the transformed program through the use of simpler primitives like format/char or format/integer . Figure 5 shows the same program as in Figure 2, but after applying partial evaluation. The format primitive is now blamed for two type errors that before could be found only at runtime. The error messages show that the user simply gave the arguments n and c in the wrong order. Similarly, specializing the regular expression engine with respect to a pattern eliminates false positives. The length of the list returned by regexp-match cannot be directly computed by the analysis since that information is hidden inside the regular expression pattern. As a result, the applications of second and third in Figure 3 are flagged as potential runtime errors (we have omitted the fairly large error messages from the figure). After specialization, the structure of the value returned by regexp-match is exposed to the analysis and MrFlow can then prove that if regexp-match returns a list, it must contain three elements. The false positives for second and third disappear in Figure 6. Of course, regexp-match can also return false at runtime, and the analysis correctly predicts this regardless of whether partial evaluation is used or not. Adding flow sensitivity for if -expressions as described in Section 7 removes these last spurious errors in Figure 6. Partial evaluation now allows the precise analysis of SchemeQL queries as well. Figure 7 shows the precise analysis of the same program as in Figure 4, this time after partial evaluation. As with regexp-match , the analysis previously computed that cursor-car could return a list of any length, and therefore flagged the call to third as a potential runtime error. This call is now free of spurious errors since the partial evaluation exposes enough structure of the list returned by cursor-car that MrFlow can compute its exact length and verify that third cannot fail at runtime. While the results computed by the analysis become more precise, partially evaluating the interpreters for any of the three embedded languages we use in this paper results in code that is bigger than the original program. Bigger code in turn means that analyses will take more time to complete. There is therefore a trade-off between precision and efficiency of the analyses. We intend to turn that trade-off into a user option in MrFlow. The user might also exercise full control over which embedded languages are partially evaluated and where by using either the functional or macro versions of the embedded languages' interpreters, switching between the two through the judicious use of a module system, for example [11]. Note that partial evaluation does not always benefit all analyses. In the regexp-match example from Figure 6, spurious errors disappear because MrFlow has been able to prove that the list r is of length three and therefore that applying the primitives second or 24 Figure 5. Precise analysis of the format primitive. Figure 6. Precise analysis of regexp-match . Figure 7. Precise analysis of a SchemeQL query. 25 third to r cannot fail. If the analysis were a Hindley-Milner-like type system, though, no difference would be seen whether partial evaluation were used or not. Indeed, while such a type system could statically prove that the arguments given to second or third are lists, is would not attempt to prove that they are lists of the required length and a runtime test would still be required. Using partial evaluation to expose such a property to the analysis would therefore be useless. Simply put, making invariants from embedded programs explicit in the host language only matters if the system analyzing the host language cares about those invariants. This does not mean partial evaluation is always useless when used in conjunction with a Hindley-Milner type system, though. Partially evaluating format , for example, would allow the type system to verify that the formatting string agrees with the types of the remaining arguments. This is in contrast to the ad-hoc solution used in OCaml [19] to type check the printf primitive, or the use of dependent types in the case of Cayenne [4]. Related Work Our work is analogous to designing type-safe embedded languages such as the one for printf [21, 4]. Both problems involve determining static information about programs based on the values of embedded programs. In some cases, designers of typed languages simply extend the host language to include specific embedded languages. The OCaml language, for example, contains a special library for printf [19] and uses of printf are type-checked in an ad-hoc manner. Similarly, the GCC compiler for the C language uses ad-hoc checking to find errors in printf format strings. Danvy [7] and Hinze [14] suggest implementations of printf in ML and Haskell, respectively, that obviate the need for dependent types by recasting the library in terms of individual combinators. In our system, those individual combinators are automatically introduced during macro expansion. The C++ language [26] likewise avoids the problem of checking invariants for printf by breaking its functionality into smaller operations that do not require the use of an embedded formatting language. A work more closely related to ours is the Cayenne language [4]. Augustsson uses a form of partial evaluation to specialize dependent types into regular Haskell-like types that can then be used by the type system to check the user's program. Our macro system uses macro-expansion time computation to specialize expressions so that the subsequent flow analysis can compute precise value flow results. Augustsson's dependent type system uses computation performed at type-checking time to specialize dependent types so that the rest of the type checking can compute precise type information. The specialization is done in his system through the use of type-computing functions that are specified by the user and evaluated by the type system. The main difference is that his system is used to compute specialized types and verify that the program is safe. Once the original program has been typed it is just compiled as-is with type checking turned off. This means that in the case of format , for example, the formatting string is processed twice: once at type checking time to prove the safety of the program, and once again at run time to compute the actual result. Our system is used to compute specialized expressions. This means that the evaluation of the format 's string needs to be done only once. Once specialized, the same program can either be run or analyzed to prove its safety. In both cases the format string will not have to be reprocessed since it has been completely replaced by more specialized code. Another difference is that in our system, non-specialized programs are still valid programs that can be analyzed, proved safe, and run (though the result of the analysis will probably be more conservative than when analyzing the corresponding partially evaluated program, so proving safety might be more difficult). This is not possible in Cayenne since programs with dependent types cannot be run without going through the partial evaluation phase first. Much work has gone into optimization of embedded languages. Hudak [15], Elliott et al [8], Backhouse [5], Christensen [6], and Veldhuizen [27] all discuss the use of partial evaluation to improve the efficiency of embedded languages, although none makes the connection between partial evaluation and static analysis. In Back-house's thesis he discusses the need to improve error checking for embedded languages, but he erroneously concludes that "syntactic analyses cannot be used due to the embedded nature of domain-specific embedded languages." The Lisp programming language ([25], Section 8.4) provides for "compiler macros" that programmers can use to create optimized versions of existing functions. The compiler is not required to use them, though. To our knowledge, there is no literature showing how to use these compiler macros to improve the results of static analyses. Lisp also has support for inlining functions, which might help monovariant analyses by duplicating the code of a function at all its call sites, thereby simulating polyvariant analyses. Bigloo [23] is a Scheme compiler that routinely implements embedded languages via macros and thus probably provides some of the benefits presented in this paper to the compiler's internal analyses . The compiler has a switch to "enable optimization by macro expansion," though there does not seem to be any documentation or literature describing the exact effect of using that switch. Conclusion Programs in embedded languages contain invariants that are not automatically enforced by their host language. We have shown that using macros to partially evaluate interpreters of little languages embedded in Scheme with respect to their input programs can recapture these invariants and convey them to a flow analysis. Because it is based on macros, this technique does not require any ad-hoc modification of either interpreters or analyses and is thus readily available to programmers. This makes it a sweet spot in the programming complexity versus precision landscape of program analysis. We intend to investigate the relationship between macros and other program analyses in a similar manner. Acknowledgments We thank Matthias Felleisen, Mitchell Wand, and Kenichi Asai for the discussions that led to this work and for their helpful feedback . Thanks to Matthew Flatt for his help with the presentation of Scheme macros. Thanks to Dale Vaillancourt for proofreading the paper and to Ryan Culpepper for his macrological wizardry. References [1] H. Abelson and G. J. Sussman. The Structure and Interpretation of Computer Programs. MIT Press, Cambridge, MA, 1985. [2] A. Aiken. Introduction to set constraint-based program analysis . Science of Computer Programming, 35:79111, 1999. 26 [3] K. Arnold, J. Gosling, and D. Holmes. The Java Programming Language. Addison-Wesley, 3d edition, 2000. [4] L. Augustsson. Cayenne--a language with dependent types. In Proceedings of the third ACM SIGPLAN international conference on Functional programming, pages 239250. ACM Press, 1998. [5] K. Backhouse. Abstract Interpretation of Domain-Specific Embedded Languages. PhD thesis, Oxford University, 2002. [6] N. H. Christensen. Domain-specific languages in software development and the relation to partial evaluation. PhD thesis , DIKU, Dept. of Computer Science, University of Copenhagen , Universitetsparken 1, DK-2100 Copenhagen East, Denmark, July 2003. [7] O. Danvy. Functional unparsing. Journal of Functional Programming , 8(6):621625, 1998. [8] C. Elliott, S. Finne, and O. de Moor. Compiling embedded languages. In SAIG, pages 927, 2000. [9] R. B. Findler, J. Clements, M. F. Cormac Flanagan, S. Krishnamurthi , P. Steckler, and M. Felleisen. DrScheme: A progamming environment for scheme. Journal of Functional Programming, 12(2):159182, March 2002. [10] C. Flanagan and M. Felleisen. Componential set-based analysis . ACM Trans. on Programming Languages and Systems, 21(2):369415, Feb. 1999. [11] M. Flatt. Composable and compilable macros: you want it when? In Proceedings of the seventh ACM SIGPLAN international conference on Functional programming, pages 7283. ACM Press, 2002. [12] P. Graunke, S. Krishnamurthi, S. V. D. Hoeven, and M. Felleisen. Programming the web with high-level programming languages. In Programming Languages and Systems , 10th European Symposium on Programming, ESOP 2001, Proceedings, volume 2028 of Lecture Notes in Computer Science, pages 122136, Berlin, Heidelberg, and New York, 2001. Springer-Verlag. [13] N. Heintze. Set Based Program Analysis. PhD thesis, Carnegie-Mellon Univ., Pittsburgh, PA, Oct. 1992. [14] R. Hinze. Formatting: a class act. Journal of Functional Programming, 13(5):935944, 2003. [15] P. Hudak. Modular domain specific languages and tools. In Proceedings of Fifth International Conference on Software Reuse, pages 134142, June 1998. [16] S. N. Kamin. Research on domain-specific embedded languages and program generators. In R. Cleaveland, M. Mis-love , and P. Mulry, editors, Electronic Notes in Theoretical Computer Science, volume 14. Elsevier, 2000. [17] R. Kelsey, W. Clinger, and J. R. [editors]. Revised 5 report on the algorithmic language Scheme. Higher-Order and Symbolic Computation, 11(1):7104, August 1998. Also appeared in SIGPLAN Notices 33:9, September 1998. [18] B. W. Kernighan and D. M. Ritchie. The C programming language . Prentice Hall Press, 1988. [19] X. Leroy. The Objective Caml System, release 3.07, 2003. http://caml.inria.fr/ocaml/htmlman . [20] P. Meunier. http://www.plt-scheme.org/software/ mrflow . [21] M. Neubauer, P. Thiemann, M. Gasbichler, and M. Sperber. Functional logic overloading. In Proceedings of the 29th ACM SIGPLAN-SIGACT symposium on Principles of programming languages, pages 233244. ACM Press, 2002. [22] J. Palsberg. Closure analysis in constraint form. Proc. ACM Trans. on Programming Languages and Systems, 17(1):47 62, Jan. 1995. [23] M. Serrano and P. Weis. Bigloo: A portable and optimizing compiler for strict functional languages. In Static Analysis Symposium, pages 366381, 1995. [24] O. Shivers. A universal scripting framework, or Lambda: the ultimate "little language". In Proceedings of the Second Asian Computing Science Conference on Concurrency and Parallelism , Programming, Networking, and Security, pages 254 265. Springer-Verlag, 1996. [25] G. L. Steele. COMMON LISP: the language. Digital Press, 12 Crosby Drive, Bedford, MA 01730, USA, 1984. With contributions by Scott E. Fahlman and Richard P. Gabriel and David A. Moon and Daniel L. Weinreb. [26] B. Stroustrup. The C++ Programming Language, Third Edition . Addison-Wesley Longman Publishing Co., Inc., 1997. [27] T. L. Veldhuizen. C++ templates as partial evaluation. In Partial Evaluation and Semantic-Based Program Manipulation, pages 1318, 1999. [28] N. Welsh, F. Solsona, and I. Glover. SchemeUnit and SchemeQL: Two little languages. In Proceedings of the Third Workshop on Scheme and Functional Programming, 2002. 27

macros;interpreter;value flow analysis;flow analysis;set-based analysis;partial evaluation;embedded language;Partial evaluation;regular expression;embedded languages;Scheme

108

IncSpan: Incremental Mining of Sequential Patterns in Large Database

Many real life sequence databases grow incrementally. It is undesirable to mine sequential patterns from scratch each time when a small set of sequences grow, or when some new sequences are added into the database. Incremental algorithm should be developed for sequential pattern mining so that mining can be adapted to incremental database updates . However, it is nontrivial to mine sequential patterns incrementally, especially when the existing sequences grow incrementally because such growth may lead to the generation of many new patterns due to the interactions of the growing subsequences with the original ones. In this study, we develop an efficient algorithm, IncSpan, for incremental mining of sequential patterns, by exploring some interesting properties. Our performance study shows that IncSpan outperforms some previously proposed incremental algorithms as well as a non-incremental one with a wide margin.

INTRODUCTION Sequential pattern mining is an important and active research topic in data mining [1, 5, 4, 8, 13, 2], with broad applications, such as customer shopping transaction analysis , mining web logs, mining DNA sequences, etc. There have been quite a few sequential pattern or closed sequential pattern mining algorithms proposed in the previous work, such as [10, 8, 13, 2, 12, 11], that mine frequent subsequences from a large sequence database efficiently. These algorithms work in a one-time fashion: mine the entire database and obtain the set of results. However, in many applications, databases are updated incrementally. For example , customer shopping transaction database is growing daily due to the appending of newly purchased items for existing customers for their subsequent purchases and/or insertion of new shopping sequences for new customers. Other examples include Weather sequences and patient treatment sequences which grow incrementally with time. The existing sequential mining algorithms are not suitable for handling this situation because the result mined from the old database is no longer valid on the updated database, and it is intolerably inefficient to mine the updated databases from scratch. There are two kinds of database updates in applications: (1) inserting new sequences (denoted as INSERT), and (2) appending new itemsets/items to the existing sequences (denoted as APPEND). A real application may contain both. It is easier to handle the first case: INSERT. An important property of INSERT is that a frequent sequence in DB = DB db must be frequent in either DB or db (or both). If a sequence is infrequent in both DB and db, it cannot be frequent in DB , as shown in Figure 1. This property is similar to that of frequent patterns, which has been used in incremental frequent pattern mining [3, 9, 14]. Such incremental frequent pattern mining algorithms can be easily extended to handle sequential pattern mining in the case of INSERT. It is far trickier to handle the second case, APPEND, than the first one. This is because not only the appended items may generate new locally frequent sequences in db, but also that locally infrequent sequences may contribute their occurrence count to the same infrequent sequences in the original database to produce frequent ones. For example, in the appended database in Figure 1, suppose |DB|=1000 and |db|=20, min sup=10%. Suppose a sequence s is in-527 Research Track Poster s infrequent s infrequent DB s is infrequent in DB' sup(s)=99 db=20 sup(s)=1 s is frequent in DB' db |DB| = 1000 Figure 1: Examples in INSERT and APPEND database frequent in DB with 99 occurrences (sup = 9.9%). In addition , it is also infrequent in db with only 1 occurrence (sup = 5%). Although s is infrequent in both DB and db, it becomes frequent in DB with 100 occurrences. This problem complicates the incremental mining since one cannot ignore the infrequent sequences in db, but there are an exponential number of infrequent sequences even in a small db and checking them against the set of infrequent sequences in DB will be very costly. When the database is updated with a combination of INSERT and APPEND, we can treat INSERT as a special case of APPEND treating the inserted sequences as appended transactions to an empty sequence in the original database. Then this problem is reduced to APPEND. Therefore, we focus on the APPEND case in the following discussion. In this paper, an efficient algorithm, called IncSpan, is developed, for incremental mining over multiple database increments. Several novel ideas are introduced in the algorithm development: (1) maintaining a set of "almost frequent " sequences as the candidates in the updated database, which has several nice properties and leads to efficient techniques , and (2) two optimization techniques, reverse pattern matching and shared projection, are designed to improve the performance. Reverse pattern matching is used for matching a sequential pattern in a sequence and prune some search space. Shared projection is designed to reduce the number of database projections for some sequences which share a common prefix. Our performance study shows that IncSpan is efficient and scalable. The remaining of the paper is organized as follows. Section 2introduces the basic concepts related to incremental sequential pattern mining. Section 3 presents the idea of buffering patterns, several properties of this technique and the associated method. Section 4 formulates the IncSpan algorithm with two optimization techniques. We report and analyze performance study in Section 5, introduce related work in Section 6. We conclude our study in Section 7. PRELIMINARY CONCEPTS Let I = {i 1 , i 2 , . . . , i k } be a set of all items. A subset of I is called an itemset. A sequence s = t 1 , t 2 , . . . , t m (t i I) is an ordered list. The size, |s|, of a sequence is the number of itemsets in the sequence. The length, l(s), is the total number of items in the sequence, i.e., l(s) = n i=1 |t i |. A sequence = a 1 , a 2 , . . . , a m is a sub-sequence of another sequence = b 1 , b 2 , . . . , b n , denoted as (if = , written as ), if and only if i 1 , i 2 , . . . , i m , such that 1 i 1 &lt; i 2 &lt; . . . &lt; i m n and a 1 b i 1 , a 2 b i 2 , . . . , and a m b i m . A sequence database, D = {s 1 , s 2 , . . . , s n }, is a set of sequences . The support of a sequence in D is the number of sequences in D which contain , support() = |{s|s D and s }|. Given a minimum support threshold, min sup, a sequence is frequent if its support is no less than min sup; given a factor 1, a sequence is semi-frequent if its support is less than min sup but no less than min sup; a sequence is infrequent if its support is less than min sup. The set of frequent sequential pattern, F S, includes all the frequent sequences; and the set of semi-frequent sequential pattern SF S, includes all the semi-frequent sequences. EXAMPLE 1. The second column of Table 1 is a sample sequence database D. If min sup = 3, F S = { (a) : 4, (b) : 3, (d) : 4, (b)(d) : 3 }. Seq ID. Original Part Appended Part 0 (a)(h) (c) 1 (eg) (a)(bce) 2 (a)(b)(d) (ck)(l) 3 (b)(df )(a)(b) 4 (a)(d) 5 (be)(d) Table 1: A Sample Sequence Database D and the Appended part Given a sequence s = t 1 , . . . , t m and another sequence s a = t 1 , . . . , t n , s = s s a means s concatenates with s a . s is called an appended sequence of s, denoted as s a s. If s a is empty, s = s, denoted as s = a s. An appended sequence database D of a sequence database D is one that (1) s i D , s j D such that s i a s j or s i = a s j , and (2 ) s i D, s j D such that s j a s i or s j = a s i . We denote LDB = {s i |s i D and s i a s j }, i.e., LDB is the set of sequences in D which are appended with items/itemsets. We denote ODB = {s i |s i D and s i a s j }, i.e., ODB is the set of sequences in D which are appended with items/itemsets in D . We denote the set of frequent sequences in D as F S . EXAMPLE 2. The third column of Table 1 is the appended part of the original database. If min sup = 3, F S = { (a) : 5, (b) : 4, (d) : 4, (b)(d) : 3, (c) : 3, (a)(b) : 3, (a)(c) : 3 }. A sequential pattern tree T is a tree that represents the set of frequent subsequences in a database. Each node p in T has a tag labelled with s or i. s means the node is a starting item in an itemset; i means the node is an intermediate item in an itemset. Each node p has a support value which represents the support of the subsequence starting from the root of T and ending at the node p. Problem Statement. Given a sequence database D, a min sup threshold, the set of frequent subsequences F S in D, and an appended sequence database D of D, the problem of incremental sequential pattern mining is to mine the set of frequent subsequences F S in D based on F S instead of mining on D from scratch. BUFFER SEMI-FREQUENT PATTERNS In this section, we present the idea of buffering semi-frequent patterns, study its properties, and design solutions of how to incrementally mine and maintain F S and SF S. 528 Research Track Poster &lt;&gt; &lt;d&gt;s:4 &lt;b&gt;s:3 &lt;a&gt;s:4 &lt;e&gt;s:2 &lt;d&gt;s:3 &lt;d&gt;s:2 &lt;b&gt;s:2 Figure 2: The Sequential Pattern Tree of F S and SF S in D 3.1 Buffering Semi-frequent Patterns We buffer semi-frequent patterns, which can be considered as a statistics-based approach. The technique is to lower the min sup by a buffer ratio 1 and keep a set SF S in the original database D. This is because since the sequences in SF S are "almost frequent ", most of the frequent subsequences in the appended database will either come from SF S or they are already frequent in the original database. With a minor update to the original database, it is expected that only a small fraction of subsequences which were infrequent previously would become frequent. This is based on the assumption that updates to the original database have a uniform probability distribution on items. It is expected that most of the frequent subsequences introduced by the updated part of the database would come from the SF S. The SF S forms a kind of boundary (or "buffer zone") between the frequent subsequences and infrequent subsequences. EXAMPLE 3. Given a database D in Example 1, min sup = 3, = 0.6. The sequential pattern tree T representing F S and SF S in D is shown in Figure 2. F S are shown in solid line and SF S in dashed line. When the database D is updated to D , we have to check LDB to update support of every sequence in F S and SF S. There are several possibilities: 1. A pattern which is frequent in D is still frequent in D ; 2. A pattern which is semi-frequent in D becomes frequent in D ; 3. A pattern which is semi-frequent in D is still semi-frequent in D ; 4. Appended database db brings new items. 5. A pattern which is infrequent in D becomes frequent in D ; 6. A pattern which is infrequent in D becomes semi-frequent in D ; Case (1)(3) are trivial cases since we already keep the information. We will consider case (4)(6) now. Case (4): Appended database db brings new items. For example, in the database D , (c) is a new item brought by db. It does not appear in D. Property: An item which does not appear in D and is brought by db has no information in F S or SF S. Solution: Scan the database LDB for single items. For a new item or an originally infrequent item in D, if it becomes frequent or semi-frequent, insert it into F S or SF S. Then use the new frequent item as prefix to construct projected database and discover frequent and semi-frequent sequences recursively. For a frequent or semi-frequent item in D, update its support. Case (5): A pattern which is infrequent in D becomes frequent in D . For example, in the database D , (a)(c) is an example of case (5). It is infrequent in D and becomes frequent in D . We do not keep (a)(c) in F S or SF S, but we have the information of its prefix (a) . Property: If an infrequent sequence p in D becomes frequent in D , all of its prefix subsequences must also be frequent in D . Then at least one of its prefix subsequences p is in F S. Solution: Start from its frequent prefix p in F S and construct p-projected database, we will discover p . Formally stated, given a frequent pattern p in D , we want to discover whether there is any pattern p with p as prefix where p was infrequent in D but is frequent in D . A sequence p which changes from infrequent to frequent must have sup(p ) &gt; (1 - )min sup. We claim if a frequent pattern p has support in LDB sup LDB (p) (1 - )min sup, it is possible that some subsequences with p as prefix will change from infrequent to frequent. If sup LDB (p) &lt; (1 - )min sup, we can safely prune search with prefix p. Theorem 1. For a frequent pattern p, if its support in LDB sup LDB (p) &lt; (1 - )min sup, then there is no sequence p having p as prefix changing from infrequent in D to frequent in D . Proof : p was infrequent in D, so sup D (p ) &lt; min sup (1) If sup LDB (p) &lt; (1 - )min sup, then sup LDB (p ) sup LDB (p) &lt; (1 - )min sup Since sup LDB (p ) = sup ODB (p ) + sup(p ). Then we have sup LDB (p ) sup LDB (p ) &lt; (1 - )min sup. (2) Since sup D (p ) = sup D (p ) + sup(p ), combining (1) and (2), we have sup D (p ) &lt; min sup. So p cannot be frequent in D . Therefore, if a pattern p has support in LDB sup LDB (p) &lt; (1 - )min sup, we can prune search with prefix p. Otherwise , if sup LDB (p) (1-)min sup, it is possible that some sequences with p as prefix will change from infrequent to frequent . In this case, we have to project the whole database D using p as prefix. If |LDB| is small or is small, there are very few patterns that have sup LDB (p) (1 - )min sup, making the number of projections small. In our example, sup LDB (a) = 3 &gt; (1 - 0.6) 3, we have to do the projection with (a) as prefix. And we discover " (a)(c) : 3" which was infrequent in D. For another example , sup LDB (d) = 1 &lt; (1 - 0.6) 3, there is no sequence with d as prefix which changes from infrequent to frequent, so we can prune the search on it. Theorem 1 provides an effective bound to decide whether it is necessary to project a database. It is essential to guarantee the result be complete. We can see from the projection condition, sup LDB (p) (1 - )min sup, the smaller is, the larger buffer we keep, the fewer database projections the algorithm needs. The choice of is heuristic. If is too high, then the buffer is small and we have to do a lot of database projections to discover sequences outside of the buffer. If is set very low, we will keep many subsequences in the buffer. But mining the buffering patterns using min sup would be much more inefficient than with min sup. We will show this 529 Research Track Poster &lt;&gt; &lt;d&gt;s:4 &lt;b&gt;s:4 &lt;a&gt;s:5 &lt;c&gt;s:3 &lt;d&gt;s:3 &lt;c&gt;s:3 &lt;d&gt;s:2 &lt;b&gt;s:3 &lt;e&gt;i:2 &lt;e&gt;s:2 Figure 3: The Sequential Pattern Tree of F S and SF S in D tradeoff through experiments in Section 5. Case (6): A pattern which is infrequent in D becomes semi-frequent in D . For example, in the database D , (be) is an example of case (6). It is infrequent in D and becomes semi-frequent in D . Property: If an infrequent sequence p becomes semi-frequent in D , all of its prefix subsequences must be either frequent or semi-frequent. Then at least one of its prefix subsequences, p, is in F S or SF S. Solution: Start from its prefix p in F S or SF S and construct p-projected database, we will discover p . Formally stated, given a pattern p, we want to discover whether there is any pattern p with p as prefix where p was infrequent but is semi-frequent in D . If the prefix p is in F S or SF S, construct p-projected database and we will discover p in p-projected database. Therefore, for any pattern p from infrequent to semi-frequent, if its prefix is in F S or SF S, p can be discovered. In our example, for the frequent pattern (b) , we do the projection on (b) and get a semi-frequent pattern (be) : 2 which was infrequent in D. We show in Figure 3 the sequential pattern tree T including F S and SF S after the database updates to D . We can compare it with Figure 2to see how the database update affects F S and SF S. INCSPAN DESIGN AND IMPLEMENTATION In this section, we formulate the IncSpan algorithm which exploits the technique of buffering semi-frequent patterns. We first present the algorithm outline and then introduce two optimization techniques. 4.1 IncSpan: Algorithm Outline Given an original database D, an appended database D , a threshold min sup, a buffer ratio , a set of frequent sequences F S and a set of semi-frequent sequences SF S, we want to discover the set of frequent sequences F S in D . Step 1: Scan LDB for single items, as shown in case (4). Step 2: Check every pattern in F S and SF S in LDB to adjust the support of those patterns. Step 2.1: If a pattern becomes frequent, add it to F S . Then check whether it meets the projection condition. If so, use it as prefix to project database, as shown in case (5). Step 2.2: If a pattern is semi-frequent, add it to SF S . The algorithm is given in Figure 4. 4.2 Reverse Pattern Matching Reverse pattern matching is a novel optimization technique . It matches a sequential pattern against a sequence from the end towards the front. This is used to check sup-Algorithm . IncSpan(D , min sup, , F S, SF S) Input: An appended database D , min sup, , frequent sequences F S in D, semi-frequent sequences SF S in D. Output: F S and SF S . 1: F S = , SF S = 2 : Scan LDB for single items; 3: Add new frequent item into F S ; 4: Add new semi-frequent item into SF S ; 5: for each new item i in F S do 6: PrefixSpan(i, D |i, min sup, F S , SF S ); 7: for every pattern p in F S or SF S do 8: check sup(p); 9: if sup(p) = sup D (p) + sup(p) min sup 10: insert(F S , p); 11: if sup LDB (p) (1 - )min sup 12: PrefixSpan(p, D |p, min sup, F S , SF S ); 13: else 14: insert(SF S , p); 15: return; Figure 4: IncSpan algorithm s s a s' Figure 5: Reverse Pattern Matching port increase of a sequential pattern in LDB. Since the appended items are always at the end part of the original sequence, reverse pattern matching would be more efficient than projection from the front. Given an original sequence s, an appended sequence s = s s a , and a sequential pattern p, we want to check whether the support of p will be increased by appending s a to s. There are two possibilities: 1. If the last item of p is not supported by s a , whether p is supported by s or not, sup(p) is not increased when s grows to s . Therefore, as long as we do not find the last item of p in s a , we can prune searching. 2. If the last item of p is supported by s a , we have to check whether s supports p. We check this by continuing in the reverse direction. If p is not supported by s , we can prune searching and keep sup(p) unchanged. Otherwise we have to check whether s supports p. If s supports p, keep sup(p) unchanged; otherwise, increase sup(p) by 1. Figure 5 shows the reverse pattern matching. 4.3 Shared Projection Shared Projection is another optimization technique we exploit. Suppose we have two sequences (a)(b)(c)(d) and (a)(b)(c)(e) , and we need to project database using each as prefix. If we make two database projections individually , we do not take advantage of the similarity between the two subsequences. Actually the two projected databases up to subsequence (a)(b)(c) , i.e., D | (a)(b)(c) are the same. 530 Research Track Poster From D | (a)(b)(c) , we do one more step projection for item d and e respectively. Then we can share the projection for (a)(b)(c) . To use shared projection, when we detect some subsequence that needs projecting database, we do not do the projection immediately. Instead we label it. After finishing checking and labelling all the sequences, we do the projection by traversing the sequential pattern tree. Tree is natural for this task because the same subsequences are represented using shared branches. PERFORMANCE STUDY A comprehensive performance study has been conducted in our experiments. We use a synthetic data generator provided by IBM. The synthetic dataset generator can be re-trieved from an IBM website, http://www.almaden.ibm.com /cs/quest. The details about parameter settings can be re-ferred in [1]. All experiments are done on a PowerEdge 6600 server with Xeon 2.8 , 4G memory. The algorithms are written in C++ and compiled using g++ with -O3 optimization . We compare three algorithms: IncSpan, an incremental mining algorithm ISM [7], and a non-incremental algorithm PrefixSpan[8]. Figure 6 (a) shows the running time of three algorithms when min sup changes on the dataset D10C10T2.5N10, 0.5% of which has been appended with transactions. IncSpan is the fastest, outperforming PrefixSpan by a factor of 5 or more, and outperforming ISM even more. ISM even cannot finish within a time limit when the support is low. Figure 6 (b) shows how the three algorithms can be affected when we vary the percentage of sequences in the database that have been updated. The dataset we use is D10C10T2.5N10, min sup=1%. The buffer ratio = 0.8. The curves show that the time increases as the incremental portion of the database increases. When the incremental part exceeds 5% of the database, PrefixSpan outperforms IncSpan. This is because if the incremental part is not very small, the number of patterns brought by it increases, making a lot overhead for IncSpan to handle. In this case, mining from scratch is better. But IncSpan still outperforms ISM by a wide margin no matter what the parameter is. Figure 6 (c) shows the memory usage of IncSpan and ISM. The database is D10C10T2.5N10, min sup varies from 0.4% to 1.5%, buffer ratio = 0.8. Memory usage of IncSpan increases linearly as min sup decreases while memory used by ISM increases dramatically. This is because the number of sequences in negative border increases sharply as min sup decreases. This figure verifies that negative border is a memory-consuming approach. Figure 7 (a) shows how the IncSpan algorithm can be affected by varying buffer ratio . Dataset is D10C10T2.5N10, 5% of which is appended with new transactions. We use PrefixSpan as a baseline. As we have discussed before, if we set very high, we will have fewer pattern in SF S, then the support update for sequences in SF S on LDB will be more efficient. However, since we keep less information in SF S, we may need to spend more time on projecting databases. In the extreme case = 1, SF S becomes empty. On the other hand, if we set the very low, we will have a large number of sequences in SF S, which makes the support update stage very slow. Experiment shows, when = 0.8, it achieves the best performance. Figure 7 (b) shows the performance of IncSpan to handle multiple (5 updates in this case) database updates. Each time the database is updated, we run PrefixSpan to mine from scratch. We can see from the figure, as the increments accumulate, the time for incremental mining increases, but increase is very small and the incremental mining still outperforms mining from scratch by a factor of 4 or 5. This experiment shows that IncSpan can really handle multiple database updates without significant performance degrading . Figure 7 (c) shows the scalability of the three algorithms by varying the size of database. The number of sequences in databases vary from 10,000 to 100,000. 5% of each database is updated. min sup=0.8%. It shows that all three algorithms scale well with the database size. RELATED WORK In sequential pattern mining, efficient algorithms like GSP [10], SPADE [13], PrefixSpan [8], and SPAM [2] were developed . Partition [9] and FUP [3] are two algorithms which promote partitioning the database, mining local frequent itemsets , and then consolidating the global frequent itemsets by cross check. This is based on that a frequent itemset must be frequent in at least one local database. If a database is updated with INSERT, we can use this idea to do the incremental mining. Zhang et al. [14] developed two algorithms for incremental mining sequential patterns when sequences are inserted into or deleted from the original database. Parthasarathy et al. [7] developed an incremental mining algorithm ISM by maintaining a sequence lattice of an old database. The sequence lattice includes all the frequent sequences and all the sequences in the negative border. However , there are some disadvantages for using negative border: (1) The combined number of sequences in the frequent set and the negative border is huge; (2) The negative border is generated based on the structural relation between sequences . However, these sequences do not necessarily have high support. Therefore, using negative border is very time and memory consuming. Masseglia et al. [6] developed another incremental mining algorithm ISE using candidate generate-and-test approach. The problem of this algorithm is (1) the candidate set can be very huge, which makes the test-phase very slow; and (2) its level-wise working manner requires multiple scans of the whole database. This is very costly, especially when the sequences are long. CONCLUSIONS In this paper, we investigated the issues for incremental mining of sequential patterns in large databases and addressed the inefficiency problem of mining the appended database from scratch. We proposed an algorithm IncSpan by exploring several novel techniques to balance efficiency and reusability. IncSpan outperforms the non-incremental method (using PrefixSpan) and a previously proposed incremental mining algorithm ISM by a wide margin. It is a promising algorithm to solve practical problems with many real applications. There are many interesting research problems related to IncSpan that should be pursued further. For example, incremental mining of closed sequential patterns, structured 531 Research Track Poster 0.01 0.1 1 10 100 1000 0.03 0.06 0.1 0.4 0.6 0.8 1 1.5 minsup (%) Ti m e (s ) IncSpan PrefixSpan ISM (a) varying min sup 0.01 0.1 1 10 100 0.5 1 2 3 4 5 Percent of growing seq (%) Ti m e (s ) IncSpan PrefixSpan ISM (b) varying percentage of updated sequences 1 10 100 1000 10000 0.4 0.6 0.8 1 1.5 minsup (%) Mem o ry U s a g e ( M B ) ISM IncSpan (c) Memory Usage under varied min sup Figure 6: Performance study 0 8 16 24 32 40 0.4 0.5 0.6 0.7 0.8 0.9 1 PrefixSpan varying buffer ratio u Ti m e ( s ) T ime (a) varying buffer ratio 0 30 60 90 120 1 2 3 4 5 Increment of database Ti m e (s ) IncSpan PrefixSpan (b) multiple increments of database 0.01 0.1 1 10 100 1000 10 20 50 80 100 No. of S equences in 1000 Ti m e ( s ) IncSpan PrefixSpan ISM (c) varying # of sequences (in 1000) in DB Figure 7: Performance study patterns in databases and/or data streams are interesting problems for future research. REFERENCES [1] R. Agrawal and R. Srikant. Mining sequential patterns. In Proc. 1995 Int. Conf. Data Engineering (ICDE'95), pages 314, March 1995. [2 ] J. Ayres, J. E. Gehrke, T. Yiu, and J. Flannick. Sequential pattern mining using bitmaps. In Proc. 2002 ACM SIGKDD Int. Conf. Knowledge Discovery in Databases (KDD'02), July 2 002 . [3] D. Cheung, J. Han, V. Ng, and C. Wong. Maintenance of discovered association rules in large databases: An incremental update technique. In Proc. of the 12th Int. Conf. on Data Engineering (ICDE'96), March 1996. [4] M. Garofalakis, R. Rastogi, and K. Shim. SPIRIT: Sequential pattern mining with regular expression constraints. In Proc. 1999 Int. Conf. Very Large Data Bases (VLDB'99), pages 223234, Sept 1999. [5] H. Mannila, H. Toivonen, and A. I. Verkamo. Discovering frequent episodes in sequences. In Proc. 1995 Int. Conf. Knowledge Discovery and Data Mining (KDD'95), pages 210215, Aug 1995. [6] F. Masseglia, P. Poncelet, and M. Teisseire. Incremental mining of sequential patterns in large databases. Data Knowl. Eng., 46(1):97121, 2003. [7] S. Parthasarathy, M. Zaki, M. Ogihara, and S. Dwarkadas. Incremental and interactive sequence mining. In Proc. of the 8th Int. Conf. on Information and Knowledge Management (CIKM'99), Nov 1999. [8] J. Pei, J. Han, B. Mortazavi-Asl, H. Pinto, Q. Chen, U. Dayal, and M.-C. Hsu. PrefixSpan: Mining sequential patterns efficiently by prefix-projected pattern growth. In Proc. 2001 Int. Conf. Data Engineering (ICDE'01), pages 215224, April 2001. [9] A. Savasere, E. Omiecinski, and S. Navathe. An efficient algorithm for mining association rules in large databases. In Proc. 1995 Int. Conf. Very Large Data Bases (VLDB'95), Sept 1995. [10] R. Srikant and R. Agrawal. Mining sequential patterns: Generalizations and performance improvements. In Proc. of the 5th Int. Conf. on Extending Database Technology (EDBT'96), Mar 1996. [11] J. Wang and J. Han. Bide: Efficient mining of frequent closed sequences. In Proc. of 2004 Int. Conf. on Data Engineering (ICDE'04), March 2004. [12] X. Yan, J. Han, and R. Afshar. CloSpan: Mining closed sequential patterns in large datasets. In Proc. 2003 SIAM Int.Conf. on Data Mining (SDM'03), May 2003. [13] M. Zaki. SPADE: An efficient algorithm for mining frequent sequences. Machine Learning, 40:3160, 2001. [14] M. Zhang, B. Kao, D. Cheung, and C. Yip. Efficient algorithms for incremental updates of frequent sequences. In Proc. of Pacific-Asia Conf. on Knowledge Discovery and Data Mining (PAKDD'02), May 2002. 532 Research Track Poster

database updates;sequence database;shared projection;frequent itemsets;optimization;buffering pattern;sequential pattern;buffering patterns;reverse pattern matching;incremental mining

109

Index Structures and Algorithms for Querying Distributed RDF Repositories

A technical infrastructure for storing, querying and managing RDF data is a key element in the current semantic web development. Systems like Jena, Sesame or the ICS-FORTH RDF Suite are widely used for building semantic web applications. Currently, none of these systems supports the integrated querying of distributed RDF repositories. We consider this a major shortcoming since the semantic web is distributed by nature. In this paper we present an architecture for querying distributed RDF repositories by extending the existing Sesame system. We discuss the implications of our architecture and propose an index structure as well as algorithms for query processing and optimization in such a distributed context.

MOTIVATION The need for handling multiple sources of knowledge and information is quite obvious in the context of semantic web applications. First of all we have the duality of schema and information content where multiple information sources can adhere to the same schema. Further, the re-use, extension and combination of multiple schema files is considered to be common practice on the semantic web [7]. Despite the inherently distributed nature of the semantic web, most current RDF infrastructures (for example [4]) store information locally as a single knowledge repository, i.e., RDF models from remote sources are replicated locally and merged into a single model. Distribution is virtually retained through the use of namespaces to distinguish between different models. We argue that many interesting applications on the semantic web would benefit from or even require an RDF infrastructure that supports real distribution of information sources that can be accessed from a single point. Beyond Copyright is held by the author/owner(s). WWW2004 , May 1722, 2004, New York, New York, USA. ACM 1-58113-844-X/04/0005. the argument of conceptual adequacy, there are a number of technical reasons for real distribution in the spirit of distributed databases: Freshness: The commonly used approach of using a local copy of a remote source suffers from the problem of changing information . Directly using the remote source frees us from the need of managing change as we are always working with the original. Flexibility: Keeping different sources separate from each other provides us with a greater flexibility concerning the addition and removal of sources. In the distributed setting, we only have to adjust the corresponding system parameters. In many cases, it will even be unavoidable to adopt a distributed architecture, for example in scenarios in which the data is not owned by the person querying it. In this case, it will often not be permitted to copy the data. More and more information providers, however, create interfaces that can be used to query the information. The same holds for cases where the information sources are too large to just create a single model containing all the information, but they still can be queried using a special interface (Musicbrainz is an example of this case). Further, we might want to include sources that are not available in RDF, but that can be wrapped to produce query results in RDF format. A typical example is the use of a free-text index as one source of information. Sometimes there is not even a fixed model that could be stored in RDF, because the result of a query is only calculated at runtime (Google, for instance, provides a programming interface that could be wrapped into an RDF source). In all these scenarios, we are forced to access external information sources from an RDF infrastructure without being able to create a local copy of the information we want to query. On the semantic web, we almost always want to combine such external sources with each other and with additional schema knowledge. This confirms the need to consider an RDF infrastructure that deals with information sources that are actually distributed across different locations. In this paper, we address the problem of integrated access to distributed RDF repositories from a practical point of view. In particular , starting from a real-life use case where we are considering a number of distributed sources that contain research results in the form of publications, we take the existing RDF storage and retrieval system Sesame and describe how the architecture and the query processing methods of the system have to be extended in order to move to a distributed setting. 631 The paper is structured as follows. In Section 2 we present an extension of the Sesame architecture to multiple, distributed repositories and discuss basic assumptions and implications of the architecture . Section 3 presents source index hierarchies as suitable mechanisms to support the localization of relevant data during query processing. In Section 4 we introduce a cost model for processing queries in the distributed architecture, and show its use in optimizing query execution as a basis for the two-phase optimization heuristics for join ordering. Section 5 reviews previous work on index structures for object-oriented data bases. It also summarizes related work on query optimization particularly focusing on the join ordering problem. We conclude with a discussion of open problems and future work. INTEGRATION ARCHITECTURE Before discussing the technical aspects of distributed data and knowledge access, we need to put our work in context by introducing the specific integration architecture we have to deal with. This architecture limits the possible ways of accessing and processing data, and thereby provides a basis for defining some requirements for our approach. It is important to note that our work is based on an existing RDF storage and retrieval system, which more or less predefines the architectural choices we made. In this section, we describe an extension of the Sesame system [4] to distributed data sources. The Sesame architecture is flexible enough to allow a straightforward extension to a setting where we have to deal with multiple distributed RDF repositories. In the current setting, queries, expressed in Sesame's query language SeRQL, are directly passed from the query engine to an RDF API (SAIL) that abstracts from the specific implementation of the repository. In the distributed setting , we have several repositories that can be implemented in different ways. In order to abstract from this technical heterogeneity, it is useful to introduce RDF API implementations on top of each repository, making them accessible in the same way. The specific problem of a distributed architecture is now that information relevant to a query might be distributed over the different sources. This requires to locate relevant information, retrieve it, and combine the individual answers. For this purpose, we introduce a new component between the query parser and the actual SAILs the mediator SAIL (see Figure 1). In this work, we assume that local repositories are implemented using database systems that translate queries posed to the RDF API into SQL queries and use the database functionality to evaluate them (compare [5]). This assumption has an important influence on the design of the distributed query processing: the database engines underlying the individual repositories have the opportunity to perform local optimization on the SQL queries they pose to the data. Therefore we do not have to perform optimizations on sub-queries that are to be forwarded to a single source, because the repository will deal with it. Our task is rather to determine which part of the overall query has to be sent to which repository. In the remainder of this paper, we describe an approach for querying distributed RDF sources that addresses these requirements implied by the adopted architecture. We focus our attention on index structures and algorithms implemented in the mediator SAIL. Figure 1: Distribution Architecture. INDEX STRUCTURES As discussed above, in order to be able to make use of the optimization mechanisms of the database engines underlying the different repositories, we have to forward entire queries to the different repositories. In the case of multiple external models, we can further speed up the process by only pushing down queries to information sources we can expect to contain an answer. The ultimate goal is to push down to a repository exactly that part of a more complex query for which a repository contains an answer. This part can range from a single statement template to the entire query. We can have a situation where a subset of the query result can directly be extracted from one source, and the rest has to be extracted and combined from different sources. This situation is illustrated in the following example. E XAMPLE 1. Consider the case where we want to extract information about research results. This information is scattered across a variety of data sources containing information about publications , projects, patents, etc. In order to access these sources in a uniform way, we use the OntoWeb research ontology. Figure 2 shows parts of this ontology. Figure 2: Part of the OntoWeb Ontology. Suppose we now want to ask for the titles of articles by employees of organizations that have projects in the area "RDF". The path expression of a corresponding SeRQL query would be the following 1 : 1 For the sake of readability we omit namespaces whenever they do not play a technical role. 632 {A} title {T}; author {W} affiliation {O} carriesOut {P} topic {'RDF'} Now, let's assume that we have three information sources I , P , and Q . I is a publication data base that contains information about articles, titles, authors and their affiliations. P is a project data base with information about industrial projects, topics, and organizations. Finally, Q is a research portal that contains all of the above information for academic research. If we want to answer the query above completely we need all three information sources. By pushing down the entire query to Q we get results for academic research. In order to also retrieve the information for industrial research, we need to split up the query, push the fragment {A} title {T}; author {W} affiliation {O} to I , the fragment {O} carriesOut {P} topic {'RDF'} to P , and join the result based on the identity of the organization . The example illustrates the need for sophisticated indexing structures for deciding which part of a query to direct to which information source. On the one hand we need to index complex query patterns in order to be able to push down larger queries to a source; on the other hand we also need to be able to identify sub-queries needed for retrieving partial results from individual sources. In order to solve this problem we build upon existing work on indexing complex object models using join indices [14]. The idea of join indices is to create additional database tables that explic-itly contain the result of a join over a specific property. At runtime, rather than computing a join, the system just accesses the join index relation which is less computationally expensive. The idea of join indices has been adapted to deal with complex object models. The resulting index structure is a join index hierarchy [21]. The most general element in the hierarchy is an index table for elements connected by a certain path p HXXn I of length n. Every following level contains all the paths of a particular length from 2 paths of length n I at the second level of the hierarchy to n paths of length 1 at the bottom of the hierarchy. In the following, we show how the notion of join index hierarchies can be adapted to deal with the problem of determining information sources that contain results for a particular sub-query. 3.1 Source Index Hierarchies The majority of work in the area of object oriented databases is focused on indexing schema-based paths in complex object models. We can make use of this work by relating it to the graph-based interpretation of RDF models. More specifically, every RDF model can be seen as a graph where nodes correspond to resources and edges to properties linking these resources. The result of a query to such a model is a set of subgraphs corresponding to a path expression . While a path expression does not necessarily describe a single path, it describes a tree that can be created by joining a set of paths. Making use of this fact, we first decompose the path expression into a set of expressions describing simple paths, then forward the simpler path expressions to sources that contain the corresponding information using a path-based index structure, and join retrieved answers to create the result. The problem with using path indices to select information sources is the fact that the information that makes up a path might be distributed across different information sources (compare Example 1). We therefore have to use an index structure that also contains information about sub-paths without loosing the advantage of indexing complete paths. An index structure that combines these two characteristics is the join index hierarchy proposed in [21]. We therefore take their approach as a basis for defining a source index hierarchy. D EFINITION 1 (S CHEMA P ATH ). Let q a hY iY vY sY tY li be a labelled graph of an RDF model where is a set of nodes, i a set of edges, v a set of labels, sY t X i 3 and l X i 3 v. For every e P i, we have s@eA a r I Y t@eA a r P and l@eA a l e if and only if the model contains the triple @r I Y l e Y r P A. A path in G is a list of edges e H Y Y e n I such that t@e i A a s@e iCI A for all i a HY Y n P. Let p a e H Y Y e n I be a path, the corresponding schema path is the list of labels l H Y Y l n I such that l i a l@e i A. The definition establishes the notion of a path for RDF models. We can now use path-based index structures and adapt them to the task of locating path instances in different RDF models. The basic structure we use for this purpose is an index table of sources that contain instances of a certain path. D EFINITION 2 (S OURCE I NDEX ). Let p be a schema path; a source index for p is a set of pairs @s k Y n k A where s k is an information source (in particular, an RDF model) and the graph of s k contains exactly n k paths with schema path p and n k b H. A source index can be used to determine information sources that contain instances of a particular schema path. If our query contains the path p, the corresponding source index provides us with a list of information sources we have to forward the query to in order to get results. The information about the number of instance paths can be used to estimate communication costs and will be used for join ordering (see Section 4). So far the index satisfies the requirement of being able to list complete paths and push down the corresponding queries to external sources. In order to be able to retrieve information that is distributed across different sources, we have to extend the structure based on the idea of a hierarchy of indices for arbitrary sub-paths. The corresponding structure is defined as follows. D EFINITION 3 (S OURCE I NDEX H IERARCHY ). Let p a l H Y Y l n I be a schema path. A source index hierarchy for p is an n-tuple h n Y Y I i where n is a source index for p i is the set of all source indices for sub-paths of p with length i that have at least one entry. The most suitable way to represent such index structure is a hierarchy , where the source index of the indexed path is the root element . The hierarchy is formed in such a way that the subpart rooted at the source index for a path p always contains source indices for all sub-paths of p. This property will later be used in the query answering algorithm. Forming a lattice of source indices, a source index hierarchy contains information about every possible schema sub-path. Therefore we can locate all fragments of paths that might be combined into a query result. At the same time, we can first concentrate on complete path instances and successively investigate smaller fragments using the knowledge about the existence of longer paths. We illustrate this principle in the following example. 633 E XAMPLE 2. Let us reconsider the situation in Example 1. The schema path we want to index is given by the list (author, affiliation , carriesOut, topic). The source index hierarchy for this path therefore contains source indices for the paths p HXXQ : (author, affiliation, carriesOut, topic) p HXXP : (author, affiliation, carriesOut), p IXXQ : (affiliation, carriesOut, topic) p HXXI :(author, affiliation), p IXXP :(affiliation, carriesOut), p PXXQ :(carriesOut, topic) p H :(author), p I :(affiliation), p P :(carriesOut), p Q (topic) Starting from the longest path, we compare our query expression with the index (see Figure 3 for an example of index contents). We immediately get the information that Q contains results. Turning to sub-paths, we also find out that I contains results for the sub-path (author, affiliation) and P for the sub-path (carriesOut, topic) that we can join in order to compute results, because together both sub-paths make up the path we are looking for. The source indices also contain information about the fact that Q contains results for all sub-paths of our target path. We still have to take this information into account, because in combination with fragments from other sources we might get additional results. However, we do not have to consider joining sub-paths from the same source, because these results are already covered by longer paths. In the example we see that P will return far less results than I (because there are less projects than publications). We can use this information to optimize the process of joining results. A key issue connected with indexing information sources is the trade-off between required storage space and computational properties of index-based query processing. Compared to index structures used to speed up query processing within an information source, a source index is relatively small as it does not encode information about individual elements in a source. Therefore, the size of the index is independent of the size of the indexed information sources. The relevant parameters in our case are the number of sources s and the lengths of the schema path n. More specifically, in the worst case a source index hierarchy contains source indices for every sub-path of the indexed schema path. As the number of all sub-path of a path is n iaI i, the worst-case 2 space complexity of a source index hierarchy is y@s n P A. We conclude that the length of the indexed path is the significant parameter here. 3.2 Query Answering Algorithm Using the notion of a source index hierarchy, we can now define a basic algorithm for answering queries using multiple sources of information. The task of this algorithm is to determine all possible combinations of sub-paths of the given query path. For each of these combinations, it then has to determine the sources containing results for the path fragments, retrieve these results, and join them into a result for the complete path. The main task is to guarantee that we indeed check all possible combinations of sub-paths for the 2 It is the case where all sources contain results for the complete schema path. query path. The easiest way of guaranteeing this is to use a simple tree-recursion algorithm that retrieves results for the complete path, then splits the original path, and joins the results of recursive calls for the sub-paths. In order to capture all possible splits this has to be done for every possible split point in the original path. The corresponding semi-formal algorithm is given below (Algorithm 1). Algorithm 1 Compute Answers. Require: A schema path p a l H Y Y l n I Require: A source index hierarchy h a @ n Y Y I A for p for all sources s k in source index n do ANSWERS := instances of schema path p in source s k RESULT := result nswers end for if n ! P then for all i a I n I do p HXXi I := l H Y l i I p iXXn I := l i Y l n I h HXXi I := Sub-hierarchy of h rooted at the source index for p HXXi I h iXXn I := Sub-hierarchy of h rooted at the source index for p iXXn I res I := gomputeenswers@p HXXi I Y h HXXi I A res P := gomputeenswers@p iXXn I Y h iXXn I A RESULT := result join@res I Y res P A end for end if return result Note that Algorithm 1 is far from being optimal with respect to runtime performance. The straightforward recursion scheme does not take specific actions to prevent unnecessary work and it neither selects an optimal order for joining sub-paths. We can improve this situation by using knowledge about the information in the different sources and performing query optimization. QUERY OPTIMIZATION In the previous section we described a light-weight index structure for distributed RDF querying. Its main task is to index schema paths w.r.t. underlying sources that contain them. Compared to instance-level indexing, our approach does not require creating and maintaining oversized indices since there are far fewer sources than there are instances. Instance indexing would not scale in the web environment and as mentioned above in many cases it would not even be applicable, e.g., when sources do not allow replication of their data (which is what instance indices essentially do). The downside of our approach, however, is that query answering without the index support at the instance level is much more computationally intensive. Moreover, in the context of semantic web portal applications the queries are not man-entered anymore but rather generated by a portal's front-end (triggered by the user) and often exceed the size 3 which can be easily computed by using brute force. Therefore we focus in this section on query optimization as an important part of a distributed RDF query system. We try to avoid re-inventing the wheel and once again seek for inspiration in the database field, making it applicable by "relationizing" the RDF model. Each single schema path p i of length 1 (also called 1-pth ) can be perceived as a relation with two attributes: the source vertex 3 Especially, the length of the path expression. 634 Figure 3: Source index hierarchy for the given query path. s@p i A and the target vertex t@p i A. A schema path of length more than 1 is modelled as a set of relations joined together by the identity of the adjacent vertices, essentially representing a chain query of joins as defined in Definition 4. This relational view over an RDF graph offers the possibility to re-use the extensive research on join optimization in databases, e.g. [1, 8, 9, 17, 20]. Taking into account the (distributed) RDF context of the join ordering problem there are several specifics to note when devising a good query plan. As in distributed databases, communication costs significantly contribute to the overall cost of a query plan. Since in our case the distribution is assumed to be realized via an IP network with a variable bandwidth, the communications costs are likely to contribute substantially to the overall processing costs, which makes the minimization of data transmission across the network very important. Unless the underlying sources provide join capabilities, the data transmission cannot be largely reduced: all (selected) bits of data from the sources are joined by the mediator and hence must be transmitted via the network. There may exist different dependencies (both structural and ex-tensional ) on the way the data is distributed. If the information about such dependencies is available, it essentially enables the optimizer to prune join combinations which cannot yield any results. The existence of such dependencies can be (to some extent) com-puted/discovered prior to querying, during the initial integration phase. Human insight is, however, often needed in order to avoid false dependency conclusions, which could potentially influence the completeness of query answering. The performance and data statistics are both necessary for the optimizer to make the right decision. In general, the more the optimizer knows about the underlying sources and data, the better optimized the query plan is. However, taking into account the autonomy of the sources, the necessary statistics do not have to be always available. We design our mediator to cope with incomplete statistical information in such a way that the missing parameters are estimated as being worse than those that are known (pessimistic approach ). Naturally, the performance of the optimizer is then lower but it increases steadily when the estimations are made more realistic based on the actual response from the underlying sources; this is also known as optimizer calibration. As indicated above, the computational capabilities of the underlying sources may vary considerably. We distinguish between those sources that can only retrieve the selected local data (pull up strategy ) and those that can perform joins of their local and incoming external data (push down strategy), thus offering computational services that could be used to achieve both a higher degree of parallelism and smaller data transmission over the network, e.g., by applying semi-join reductions [1]. At present, however, most sources are capable only of selecting the desired data within their extent, i.e., they do not offer the join capability. Therefore, further we focus mainly on local optimization at the mediator's side. For this purpose we need to perceive an RDF model as a set of relations on which we can apply optimization results from the area of relational databases. In this context the problem of join ordering arises, when we want to compute the results for schema paths from partial results obtained from different sources. Creating the result for a schema corresponds to the problem of computing the result of a chain query as defined below: D EFINITION 4 (C HAIN Q UERY ). Let p be a schema path composed from the 1-paths p I Y Y p n . The chain query of p is the n-join p I FG t@p I Aas@p P A p P FG t@p P Aas@p Q A p Q FG p n , where s@p i A and t@p i A are returning an identity of a source and target node, respectively. As the join condition and attributes follow the same pattern for all joins in the chain query, we omit them whenever they are clear from the context. In other words, to follow a path p of length 2 means performing a join between the two paths of length 1 which p is composed from. The problem of join optimization is to determine the right order in which the joins should be computed, such that the overall response time for computing the path instances is minimized. 4 Note that a chain query in Definition 4 does not include explicit joins, i.e., those specified in the here clause, or by assigning the same variable names along the path expression. When we append these explicit joins, the shape of the query usually changes from a linear chain to a query graph containing a circle or a star, making the join ordering problem NP-hard [15]. 4 In case the sources offer also join capabilities the problem is not only in which order but also where the joins should take place. 635 4.1 Space Complexity Disregarding the solutions obtained by the commutativity of joins, each query execution plan can be associated with a sequence of numbers that represents the order in which the relations are joined. We refer to this sequence as footprint of the execution plan. E XAMPLE 3. For brevity reasons, assume the following name substitutions in the model introduced in Example 1: the concept names Article, Employee, Organization, Project, ResearchTopic become a, b, c, d, e, respectively; the property names author, affiliation , carriesOut, topic are substituted with 1, 2, 3, 4, respectively. Figure 4 presents two possible execution plans and their footprints. Figure 4: Two possible query executions and their footprints. If also the order of the join operands matters, i.e., the commutativity law is considered, the sequence of the operands of each join is recorded in the footprint as well. The solution space consists of query plans (their footprints) which can be generated. We distinguish two cases: first the larger solution space of bushy trees and then its subset consisting of right-deep trees. If we allow for an arbitrary order of joins the resulting query plans are so-called bushy trees where the operands of a join can be both a base relation 5 or a result of a previous join. For a query with n joins there are n3 possibilities of different query execution plans if we disregard the commutativity of joins and cross products . Note that in the case of bushy trees, there might be several footprints associated with one query tree. For instance, the bushy tree in Example 3 can be evaluated in different order yielding two more footprints: (2, 4, 1, 3) or (4, 2, 1, 3). In our current approach, these footprints would be equivalent w.r.t. the cost they represent. However, treating them independently allows us to consider in the future also semi-join optimization [1] where their cost might differ considerably. If the commutativity of join is taken into account, there are Pn n n3 P n different possibilities of ordering joins and their individual constituents [22]. However, in case of memory-resident databases where all data fits in main memory, the possibilities generated by the commutativity law can be for some join methods neglected as they mainly play a role in the cost model minimizing disk-memory operations ; we discuss this issue further in Subsection 4.2. We adopt the memory-only strategy as in our context there are always only two 5 A base relation is that part of the path which can be retrieved directly from one source. attributes per relation, both of them being URI references which, when the namespace prefix is stored separately, yield a very small size. Of course, the assumption we make here is that the Sesame server is equipped with a sufficient amount of memory to accommodate all intermediate tuples of relations appearing in the query. A special case of a general execution plan is a so-called right-deep tree which has the left-hand join operands consisting only of base relations. For a footprint that starts with the r-th join there are n r possibilities of finishing the joining sequence. Thus there are in total n I iaH n I i a P n I possibilities of different query execution plans. 6 . In this specially shaped query tree exists an execution pipeline of length n I that allows both for easier parallelizing and for shortening the response time [8] This property is very useful in the context of the WWW where many applications are built in a producer-consumer paradigm. 4.2 Cost Model The main goal of query optimization is to reduce the computational cost of processing the query both in terms of the transmission cost and the cost of performing join operations on the retrieved result fragments. In order to determine a good strategy for processing a query, we have to be able to exactly determine the cost of a query execution plan and to compare it to costs of alternative plans. For this purpose, we capture the computational costs of alternative query plans in a cost model that provides the basis for the optimization algorithm that is discussed later. As mentioned earlier, we adopt the memory-resident paradigm, and the cost we are trying to minimize is equivalent to minimizing the total execution time. There are two main factors that influence the resulting cost in our model. First is the cost of data transmission to the mediator, and second is the data processing cost. D EFINITION 5 (T RANSMISSION C OST ). The transmission cost of path instances of the schema path p from a source to the mediator is modelled as g p a ginit C jpj vngth p ksk g where ginit represents the cost of initiating the data transmission, jpj denotes the cardinality, vngth p stands for the length of the schema path p, ksk is the size of a URI at the source X 7 and g represents transmission cost per data unit from to the mediator. Since we apply all reducing operations (e.g., selections and projections ) prior to the data transmission phase, the data processing mainly consists of join costs. The cost of a join operation is influ-enced by the cardinality of the two operands and the join-method which is utilized. As we already pointed out, there are no instance indices at the mediator side that would allow us to use some join "shortcuts". In the following we consider two join methods: a nested loop join and a hash join both without additional indexing support. D EFINITION 6 (N ESTED LOOP JOIN COST ). The processing cost of a nested loop join of two relations pY r is defined as xtg pYr a jpjjrju@pY rA, where jxj denotes the cardinality of the relation x and u@pY rA represents the cost of the identity comparison. 6 The number corresponds to a sum of the n I-th line in the Pascal triangle. 7 Different sources may model URIs differently, however, we assume that at the mediator all URIs are represented in the same way. 636 Note that the nested loop join allows for a more sophisticated definition of object equality than a common URI comparison. In particular, if necessary, the basic URI comparison can be complemented by (recursive) comparisons of property values or mapping look-ups. This offers room to address the issue of URI diversity also known as the designation problem, when two different URIs refer to the same real-life object. D EFINITION 7 (H ASH JOIN COST ). The processing cost of a hash join of two relations pY r is defined as rtg pYr a s jpj C jrj f, where jxj denotes the cardinality of the relation x, s represents the cost of inserting a path instance in the hash table (the building factor), models the cost of retrieving a bucket from the hash table, and f stands for the average number of path instances in the bucket. Unlike the previous join method, the hash join algorithm assumes that the object equality can be determined by a simple URI comparison, in other words that the URI references are consistent across the sources. Another difference is that in the case of the nested loop join for in-memory relations the join commutativity can be neglected, as the query plan produced from another query plan by the commutativity law will have exactly the same cost. However , in the case of the hash join method the order of operands influences the cost and thus the solution space must also include those solutions produced by the commutativity law. D EFINITION 8 (Q UERY PLAN COST ). The overall cost of a query plan consists of the sum of all communication costs and all join processing costs of the query tree. g a n iaI g p i C g , where g represents the join processing cost of the query tree and it is computed as a sum of recurrent applications of the formula in Definition 6 or 7 depending on which join method is utilized. To compute the cardinality of non-base join arguments, a join selectivity is used. The join selectivity ' is defined as a ratio between the tuples retained by the join and those created by the Cartesian product: ' a jpFGrj jprj . As it is not possible to determine the precise join selectivity before the query is evaluated, ' for each sub-path join is assumed to be estimated and available in the source index hierarchy. After the evaluation of each query initial ' estimates are improved and made more realistic. 4.3 Heuristics for join ordering While the join ordering problem in the context of a linear/chain query can be solved in a polynomial time [12], we have to take into account the more complex problem when also the explicit joins are involved which is proven to be NP-hard [15]. It is apparent that evaluating all possible join strategies for achieving the global optimum becomes quickly unfeasible for a larger n. In these cases we have to rely on heuristics that compute a "good-enough" solution given the constraints. In fact, this is a common approach for op-timizers in interactive systems. There, optimization is often about avoiding bad query plans in very short time, rather than devoting a lot of the precious CPU time to find the optimal plan, especially, when it is not so uncommon that the optimal plan improves the heuristically obtained solutions only marginally. Heuristics for the join ordering problem have been studied exten-sively in the database community. In this work we adopt the results of comparing different join ordering heuristics from [17]. Inspired from this survey, we chose to apply the two-phase optimization consisting of the iterative improvement (II) algorithm followed by the simulated annealing (SA) algorithm [20]. This combination performs very well on the class of queries we are interested in, both in the bushy and the right-deep tree solution space, and degrades gracefully under time constrains. The II algorithm is a simple greedy heuristics which accepts any improvement on the cost function. The II randomly generates several initial solutions, taking them as starting points for a walk in the chosen solution space. The actual traversal is performed by applying a series of random moves from a predefined set. The cost function is evaluated for every such move, remembering the best solution so far. The main idea of this phase is to descent rapidly into several local minima assuring aforementioned graceful degradation . For each of the sub-optimal solutions, the second phase of the SA algorithm is applied. The task of the SA phase is to explore the "neighborhood" of a prosperous solution more thor-oughly , hopefully lowering the cost. Algorithm 2 Simulated annealing algorithm Require: start solution solution Require: start temperature s empr solution := solution estolution := solution tempr := s empr ost := gost@estolutionA mingost := ost repeat repeat newolution := NEW(solution) newgost := gost@newolutionA if newgost ost then solution := newolution ost := newCost else if e @newgost ostA tempr ! exh@HXXIA then solution := newolution ost := newCost end if if ost ` mingost then estolution := solution mingost := ost end if until equilibrium reached DECREASE( tempr) until frozen return estolution The pseudo-code of the SA phase is presented in Algorithm 2. It takes a starting point/solution from the II phase, and similarly to II performs random moves from a predefined set accepting all cost improvements. However, unlike the II, the SA algorithm can accept with a certain probability also those moves that result in a solution with a higher cost than the current best solution. The probability of such acceptance depends on the temperature of the system and the cost difference. The idea is that at the beginning the system is hot and accepts easier the moves yielding even solutions with higher costs. However, as the temperature decreases the system is becoming more stable, strongly preferring those solutions with lower costs. The SA algorithm improves on the II heuristics by making the stop condition less prone to get trapped in a local minimum; SA stops when the temperature drops below a certain threshold or if the best solution so far was not improved in a number of consecutive 637 temperature decrements, the system is considered frozen. There are two sets of moves: one for the bushy solution space and one for the right-deep solution space; for details we refer the reader to [20]. Figure 5: Acceptance probability with respect to the temperature and the cost difference. Figure 5 shows the acceptance probability dependency in the SA phase computed for the range of parameters that we used in our experiments . As we adopted the two-phase algorithm our simulations were able to reproduce the trends in results presented in [17]; due to the lack of space we omit the detail performance analysis and the interested reader is referred to the aforementioned survey. RELATED WORK In this paper we focused mainly on basic techniques such as indexing and join ordering. Relevant related work is described in the remainder of this section. More advanced techniques such as site selection and dynamic data placement are not considered, because they are not supported by the current architecture of the system. We also do not consider techniques that involve view-based query answering techniques [6] because we are currently not considering the problem of integrating heterogeneous data. 5.1 Index Structures for Object Models There has been quite a lot of research on indexing object oriented databases. The aim of this work was to speed up querying and navigation in large object databases. The underlying idea of many existing approaches is to regard an object base as a directed graph, where objects correspond to nodes, and object properties to links [16]. This view directly corresponds to RDF data, that is often also regarded as a directed graph. Indices over such graph structures now describe paths in the graph based on a certain pattern normally provided by the schema. Different indexing techniques vary on the kind of path patterns they describe and on the structure of the index. Simple index structures only refer to a single property and organize objects according to the value of that property. Nested indices and path indices cover a complete path in the model that might contain a number of objects and properties [2]. In RDF as well as in object oriented databases, the inheritance relation plays a special role as it is connected with a predefined semantics. Special index structures have been developed to speed up queries about such hierarchies and have recently been rediscovered for indexing RDF data [5]. In the area of object-oriented database systems, these two kinds of indexing structures have been combined resulting in the so-called nested inheritance indices [3] and generalized nested inheritance indices [16]. These index structures directly represent implications of inheritance reasoning, an approach that is equivalent to indexing the deductive closure of the model. 5.2 Query Optimization There is a long tradition of work on distributed databases in general [13] and distributed query processing in particular [10]. The dominant problem is the generation of an optimal query plan that reduces execution costs as much as possible while guaranteeing completeness of the result. As described by Kossmann in [10], the choice of techniques for query plan generation depends on the architecture of the distributed system. He discusses basic techniques as well as methods for client-server architectures and for heterogeneous databases. Due to our architectural limitations (e.g., limited source capabilities) we focused on join-ordering optimization which can be performed in a centralized manner by the mediator . While some restricted cases of this problem can be solved in a polynomial time [12, 11], the general problem of finding an optimal plan for evaluating join queries has been proven to be NP-hard [15]. The approaches to tackle this problem can be split into several categories [17]: deterministic algorithms, randomized algorithms, and genetic algorithms. Deterministic algorithms often use techniques of dynamic programming (e.g. [12]), however, due to the complexity of the problem they introduce simplifications, which render them as heuristics. Randomized algorithms (e.g. [20, 19]), perform a random walk in the solution space according to certain rules. After the stop-condition is fulfilled, the best solution found so far is declared as the result. Genetic algorithms (e.g. [18]) perceive the problem as biological evolution; they usually start with a random population (set of solutions) and generate offspring by applying a crossover and mutation. Subsequently, the selection phase eliminates weak members of the new population. LIMITATIONS AND FUTURE WORK The work reported in this paper can be seen as a very first step towards a solution for the problem of distributed processing of RDF queries. We motivated the overall problem and proposed some data structures and algorithms that deal with the most fundamental problems of distributed querying in a predefined setting. We identified a number of limitations of the current proposal with respect to the generality of the approach and assumptions made. These limitations also set the agenda for future work to be done on distributed RDF querying and its support in Sesame. Implementation Currently, our work on distributed query processing is of a purely theoretical nature. The design and evaluation of the methods described are based on previous work reported in the literature and on worst-case complexity estimations. The next step is to come up with a test implementation of a distributed RDF storage system. The implementation will follow the architecture introduced in the beginning of the paper and will be built on top of the Sesame storage and retrieval engine. The implementation will provide the basis for a more practical evaluation of our approach and will allow us to make assertions about the real system behavior in the presence of different data sets and different ways they are distributed. Such a practical evaluation will be the basis for further optimization of the methods. Schema-Awareness One of the limitations of the approach described in this paper concerns schema aware querying in a distributed setting. Even if every single repository is capable of computing the deductive closure of the model it contains, the overall 638 result is not necessarily complete, as schema information in one repository can have an influence on information in other repositories . This information could lead to additional conclusions if taken into account during query processing. In order to be able to deal with this situation, we need to do some additional reasoning within the mediator in order to detect and process dependencies between the different models. Object Identity One of the basic operations of query processing is the computation of joins of relations that correspond to individual properties. The basic assumption we make at this point is that we are able to uniquely determine object identity. Identity is essential because it is the main criterion that determines whether to connect two paths or not. From a pragmatic point of view, the URI of an RDF resource provides us with an identity criterion. While this may be the case in a single repository, it is not clear at all whether we can make this assumption in a distributed setting as different repositories can contain information about the same real world object (e.g., a paper) and assign different URIs to it. To deal with this situation we have to develop heuristics capable of deciding whether two resources describe the same real world object. Query Model In order to be able to design efficient index structures we restricted ourselves to path queries as a query model that is directly supported. We argued above that tree-shaped queries can be easily split into a number of path queries that have to be joined afterwards. Nevertheless, this simplification does not apply to the optimization part which is capable of processing also different query shapes. An important aspect of future work is to extend our indexing approach to more expressive query models that also include tree and graph shaped queries which can be found in existing RDF query languages. It remains to be seen whether the same kind of structures and algorithms can be used for more complex queries or whether we have to find alternatives. Architecture The starting point of our investigation was a particular architecture, namely a distributed repository where the data is accessed at a single point but stored in different repositories. We further made the assumption that these repositories are read-only, i.e., they only provide answers to path queries that they are known to contain some information about. An interesting question is how more flexible architectures can be supported. We think of architectures where information is accessed from multiple points and repositories are able to forward queries. Further we can imagine grid-based architectures where components can perform local query processing on data received from other repositories. A prominent example of such more flexible architectures are peer-to-peer systems. This would also bring a new potential for optimization as peers may collaborate on query evaluation which in turn may help in reducing both the communication and processing costs. REFERENCES [1] P. Bernstein and D. Chiu. Using semi-joins to solve relational queries. Journal of the ACM, 28:2540, 1981. [2] E. Bertino. An indexing technique for object-oriented databases. In Proceedings of the Seventh International Conference on Data Engineering, April 8-12, 1991, Kobe, Japan, pages 160170. IEEE Computer Society, 1991. [3] E. Bertino and P. Foscoli. Index organizations for object-oriented database systems. TKDE, 7(2):193209, 1995. [4] J. Broekstra, A. Kampman, and F. van Harmelen. Sesame: A generic architecture for storing and querying rdf and rdf schema. In The Semantic Web - ISWC 2002, volume 2342 of LNCS, pages 5468. Springer, 2002. [5] V. Christophides, D. Plexousakisa, M. Scholl, and S. Tourtounis. On labeling schemes for the semantic web. In Proceedings of the 13th World Wide Web Conference, pages 544555, 2003. [6] A. Halevy. Answering queries using views - a survey. The VLDB Journal, 10(4):270294, 2001. [7] J. Hendler. Agents and the semantic web. IEEE Intelligent Systems, (2), 2001. [8] H. Hsiao, M. Chen, and P. Yu. Parallel execution of hash joins in parallel databases. IEEE Transactions on Parallel and Distributed Systems, 8:872883, 1997. [9] Y. Ioannidis and E. Wong. Query optimization by simulated annealing. In ACM SIGMOD International Conference on Management of Data, pages 922. ACM:Press, 1987. [10] D. Kossmann. The state of the art in distributed query processing. ACM Computing Surveys, 32(4):422469, 2000. [11] G. Moerkotte. Constructing optimal bushy trees possibly containing cross products for order preserving joins is in p, tr-03-012. Technical report, University of Mannheim, 2003. [12] K. Ono and G. M. Lohman. Measuring the complexity of join enumeration in query optimization. In 16th International Conference on Very Large Data Bases, pages 314325. Morgan Kaufmann, 1990. [13] M. Ozsu and P. Valduriez. Principles of Distributed Database Systems. Prentice Hall, 1991. [14] D. Rotem. Spatial join indices. In Proceedings of International Conference on Data Engineering, 1991. [15] W. Scheufele and G. Moerkotte. Constructing optimal bushy processing trees for join queries is np-hard, tr-96-011. Technical report, University of Mannheim, 1996. [16] B. Shidlovsky and E. Bertino. A graph-theoretic approach to indexing in object-oriented databases. In S. Y. W. Su, editor, Proceedings of the Twelfth International Conference on Data Engineering, February 26 - March 1, 1996, New Orleans, Louisiana, pages 230237. IEEE Computer Society, 1996. [17] M. Steinbrunn, G. Moerkotte, and A. Kemper. Heuristic and randomized optimization for join ordering problem. The VLDB Journal, 6:191208, 1997. [18] M. Stillger and M. Spiliopoulou. Genetic programming in database query optimization. In J. R. Koza, D. E. Goldberg, D. B. Fogel, and R. L. Riolo, editors, Genetic Programming 1996: Proceedings of the First Annual Conference, pages 388393. MIT Press, 1996. [19] A. Swami. Optimization of large join queries: combining heuristics and combinatorial techniques. In ACM SIGMOD International Conference on Management of Data, pages 367376. ACM:Press, 1989. [20] A. Swami and A. Gupta. Optimization of large join queries. In ACM SIGMOD International Conference on Management of Data, pages 817. ACM:Press, 1988. [21] Z. Xie and J. Han. Join index hierarchies for supporting efficient navigations in object-oriented databases. In Proceedings of the International Conference on Very Large Data Bases, pages 522533, 1994. [22] C. Yu and W. Meng. Principles of Database Query Processing for Advanced Applications. Morgan Kaufmann Publishers, 1998. 639

index structure;external sources;query optimization;distributed architecture;repositories;RDF;infrastructure;RDF Querying;Optimization;Index Structures;semantic web;join ordering problem

11

A Functional Correspondence between Evaluators and Abstract Machines

We bridge the gap between functional evaluators and abstract machines for the λ-calculus, using closure conversion, transformation into continuation-passing style, and defunctionalization. We illustrate this approach by deriving Krivine's abstract machine from an ordinary call-by-name evaluator and by deriving an ordinary call-by-value evaluator from Felleisen et al.'s CEK machine. The first derivation is strikingly simpler than what can be found in the literature. The second one is new. Together, they show that Krivine's abstract machine and the CEK machine correspond to the call-by-name and call-by-value facets of an ordinary evaluator for the λ-calculus. We then reveal the denotational content of Hannan and Miller's CLS machine and of Landin's SECD machine. We formally compare the corresponding evaluators and we illustrate some degrees of freedom in the design spaces of evaluators and of abstract machines for the λ-calculus with computational effects. Finally, we consider the Categorical Abstract Machine and the extent to which it is more of a virtual machine than an abstract machine

Introduction and related work In Hannan and Miller's words [23, Section 7], there are fundamental differences between denotational definitions and definitions of abstract machines. While a functional programmer tends to be familiar with denotational definitions [36], he typically wonders about the following issues: Design: How does one design an abstract machine? How were existing abstract machines, starting with Landin's SECD machine , designed? How does one make variants of an existing abstract machine? How does one extend an existing abstract machine to a bigger source language? How does one go about designing a new abstract machine? How does one relate two abstract machines? Correctness: How does one prove the correctness of an abstract machine? Assuming it implements a reduction strategy, should one prove that each of its transitions implements a part of this strategy? Or should one characterize it in reference to a given evaluator, or to another abstract machine? A variety of answers to these questions can be found in the literature . Landin invented the SECD machine as an implementation model for functional languages [26], and Plotkin proved its correctness in connection with an evaluation function [30, Section 2]. Krivine discovered an abstract machine from a logical standpoint [25], and Cregut proved its correctness in reference to a reduction strategy; he also generalized it from weak to strong normalization [7]. Curien discovered the Categorical Abstract Machine from a categorical standpoint [6, 8]. Felleisen et al. invented the CEK machine from an operational standpoint [16, 17, 19]. Hannan and Miller discovered the CLS machine from a proof-theoretical standpoint [23]. Many people derived, invented, or (re-)discovered Krivine's machine. Many others proposed modifications of existing machines. And recently, Rose presented a method to construct abstract machines from reduction rules [32], while Hardin, Maranget, and Pagano presented a method to extract the reduction strategy of a machine by extracting axioms from its transitions and structural rules from its architecture [24]. In this article, we propose one constructive answer to all the questions above. We present a correspondence between functional evaluators and abstract machines based on a two-way derivation consisting of closure conversion, transformation into continuation-passing style (CPS), and defunctionalization. This two-way derivation lets us connect each of the machines above with an evaluator, and makes it possible to echo variations in the evaluator into variations in the abstract machine, and vice versa. The evaluator clarifies the reduction strategy of the corresponding machine. The abstract machine makes the evaluation steps explicit in a transition system. 8 Some machines operate on -terms directly whereas others operate on compiled -terms expressed with an instruction set. Accordingly , we distinguish between abstract machines and virtual machines in the sense that virtual machines have an instruction set and abstract machines do not; instead, abstract machines directly operate on source terms and do not need a compiler from source terms to instructions. (Gregoire and Leroy make the same point when they talk about a compiled implementation of strong reduction [21].) Prerequisites: ML, observational equivalence, abstract machines, -interpreters, CPS transformation, defunctionalization , and closure conversion. We use ML as a meta-language, and we assume a basic familiarity with Standard ML and reasoning about ML programs. In particular , given two pure ML expressions e and e' we write e e' to express that e and e' are observationally equivalent. Most of our implementations of the abstract machines raise compiler warnings about non-exhaustive matches. These are inherent to programming abstract machines in an ML-like language. The warnings could be avoided with an option type or with an explicit exception, at the price of readability and direct relation to the usual mathematical specifications of abstract machines. It would be helpful to the reader to know at least one of the machines considered in the rest of this article, be it Krivine's machine , the CEK machine, the CLS machine, the SECD machine, or the Categorical Abstract Machine. It would also be helpful to have already seen a -interpreter written in a functional language [20, 31, 35, 39]. In particular, we make use of Strachey's notions of expressible values, i.e., the values obtained by evaluating an expression, and denotable values, i.e., the values denoted by identifiers [38]. We make use of the CPS transformation [12, 33]: a term is CPS-transformed by naming all its intermediate results, sequentializing their computation, and introducing continuations. Plotkin was the first to establish the correctness of the CPS transformation [30]. We also make use of Reynolds's defunctionalization [31]: defunctionalizing a program amounts to replacing each of its function spaces by a data type and an apply function; the data type enumerates all the function abstractions that may give rise to inhabitants of this function space in this program [15]. Nielsen, Banerjee, Heintze, and Riecke have established the correctness of defunctionalization [3, 29]. A particular case of defunctionalization is closure conversion: in an evaluator, closure conversion amounts to replacing each of the function spaces in expressible and denotable values by a tuple, and inlining the corresponding apply function. We would like to stress that all the concepts used here are elementary ones, and that the significance of this article is the one-fits-all derivation between evaluators and abstract machines. Overview: The rest of this article is organized as follows. We first consider a call-by-name and a call-by-value evaluator, and we present the corresponding machines, which are Krivine's machine and the CEK machine. We then turn to the CLS machine and the SECD machine, and we present the corresponding evaluators. We finally consider the Categorical Abstract Machine. For simplicity, we do not cover laziness and sharing, but they come for free by threading a heap of updateable thunks in a call-by-name evaluator [2]. Call-by-name, call-by-value, and the calculus We first go from a call-by-name evaluator to Krivine's abstract machine (Section 2.1) and then from the CEK machine to a call-by-value evaluator (Section 2.2). Krivine's abstract machine operates on de Bruijn-encoded -terms, and the CEK machine operates on -terms with names. Starting from the corresponding evaluators, it is simple to construct a version of Krivine's abstract machine that operates on -terms with names, and a version of the CEK machine that operates on de Bruijn-encoded -terms (Section 2.3). The derivation steps consist of closure conversion, transformation into continuation-passing style, and defunctionalization of continuations . Closure converting expressible and denotable values makes the evaluator first order. CPS transforming the evaluator makes its control flow manifest as a continuation. Defunctionalizing the continuation materializes the control flow as a first-order data structure. The result is a transition function, i.e., an abstract machine. 2.1 From a call-by-name evaluator to Krivine's machine Krivine's abstract machine [7] operates on de Bruijn-encoded terms . In this representation, identifiers are represented by their lexical offset, as traditional since Algol 60 [40]. datatype term = IND of int (* de Bruijn index *) | ABS of term | APP of term * term Programs are closed terms. 2.1.1 A higher-order and compositional call-by-name evaluator Our starting point is the canonical call-by-name evaluator for the -calculus [35, 37]. This evaluator is compositional in the sense of denotational semantics [34, 37, 41] and higher order ( Eval0.eval ). It is compositional because it solely defines the meaning of each term as a composition of the meaning of its parts. It is higher order because the data types Eval0.denval and Eval0.expval contain functions: denotable values ( denval ) are thunks and expressible values ( expval ) are functions. An environment is represented as a list of denotable values. A program is evaluated in an empty environment ( Eval0.main ). structure Eval0 = struct datatype denval = THUNK of unit -&gt; expval and expval = FUNCT of denval -&gt; expval (* eval : term * denval list -&gt; expval *) fun eval (IND n, e) = let val (THUNK thunk) = List.nth (e, n) in thunk () end | eval (ABS t, e) = FUNCT (fn v =&gt; eval (t, v :: e)) | eval (APP (t0, t1), e) = let val (FUNCT f) = eval (t0, e) in f (THUNK (fn () =&gt; eval (t1, e))) end (* main : term -&gt; expval *) fun main t = eval (t, nil) end 9 An identifier denotes a thunk. Evaluating an identifier amounts to forcing this thunk. Evaluating an abstraction yields a function . Evaluating an application requires the evaluation of the sub-expression in position of function; the intermediate result is a function , which is applied to a thunk. 2.1.2 From higher-order functions to closures We now closure-convert the evaluator of Section 2.1.1. In Eval0 , the function spaces in the data types of denotable and expressible values are only inhabited by instances of the abstractions fn v =&gt; eval (t, v :: e) in the meaning of abstractions , and fn () =&gt; eval (t1, e) in the meaning of applications. Each of these -abstractions has two free variables: a term and an environment. We defunctionalize these function spaces into closures [15, 26, 31], and we inline the corresponding apply functions. structure Eval1 = struct datatype denval = THUNK of term * denval list and expval = FUNCT of term * denval list (* eval : term * denval list -&gt; expval *) fun eval (IND n, e) = let val (THUNK (t, e')) = List.nth (e, n) in eval (t, e') end | eval (ABS t, e) = FUNCT (t, e) | eval (APP (t0, t1), e) = let val (FUNCT (t, e')) = eval (t0, e) in eval (t, (THUNK (t1, e)) :: e') end (* main : term -&gt; expval *) fun main t = eval (t, nil) end The definition of an abstraction is now Eval1.FUNCT (t, e) instead of fn v =&gt; Eval0.eval (t, v :: e) , and its use is now Eval1.eval (t, (Eval1.THUNK (t1, e)) :: e') instead of f (Eval0.THUNK (fn () =&gt; Eval0.eval (t1, e))) . Similarly, the definition of a thunk is now Eval1.THUNK (t1, e) instead of Eval0.THUNK (fn () =&gt; Eval0.eval (t1, e)) and its use is Eval1.eval (t, e') instead of thunk () . The following proposition is a corollary of the correctness of defunctionalization . P ROPOSITION 1 ( FULL CORRECTNESS ). For any ML value p : term denoting a program, evaluating Eval0.main p yields a value FUNCT f and evaluating Eval1.main p yields a value FUNCT (t, e) such that f fn v =&gt; Eval1.eval (t, v :: e) 2.1.3 CPS transformation We transform Eval1.eval into continuation-passing style. 1 Doing so makes it tail recursive. 1 Since programs are closed, applying List.nth cannot fail and therefore it denotes a total function. We thus keep it in direct style [14]. structure Eval2 = struct datatype denval = THUNK of term * denval list and expval = FUNCT of term * denval list (* eval : term * denval list * (expval -&gt; 'a) *) (* -&gt; 'a *) fun eval (IND n, e, k) = let val (THUNK (t, e')) = List.nth (e, n) in eval (t, e', k) end | eval (ABS t, e, k) = k (FUNCT (t, e)) | eval (APP (t0, t1), e, k) = eval (t0, e, fn (FUNCT (t, e')) =&gt; eval (t, (THUNK (t1, e)) :: e', k)) (* main : term -&gt; expval *) fun main t = eval (t, nil, fn v =&gt; v) end The following proposition is a corollary of the correctness of the CPS transformation. (Here observational equivalence reduces to structural equality over ML values of type expval .) P ROPOSITION 2 ( FULL CORRECTNESS ). For any ML value p : term denoting a program, Eval1.main p Eval2.main p 2.1.4 Defunctionalizing the continuations The function space of the continuation is inhabited by instances of two -abstractions: the initial one in the definition of Eval2.main , with no free variables, and one in the meaning of an application, with three free variables. To defunctionalize the continuation, we thus define a data type cont with two summands and the corresponding apply cont function to interpret these summands. structure Eval3 = struct datatype denval = THUNK of term * denval list and expval = FUNCT of term * denval list and cont = CONT0 | CONT1 of term * denval list * cont (* eval : term * denval list * cont -&gt; expval *) fun eval (IND n, e, k) = let val (THUNK (t, e')) = List.nth (e, n) in eval (t, e', k) end | eval (ABS t, e, k) = apply_cont (k, FUNCT (t, e)) | eval (APP (t0, t1), e, k) = eval (t0, e, CONT1 (t1, e, k)) and apply_cont (CONT0, v) = v | apply_cont (CONT1 (t1, e, k), FUNCT (t, e')) = eval (t, (THUNK (t1, e)) :: e', k) (* main : term -&gt; expval *) fun main t = eval (t, nil, CONT0) end The following proposition is a corollary of the correctness of defunctionalization . (Again, observational equivalence reduces here to structural equality over ML values of type expval .) 10 P ROPOSITION 3 ( FULL CORRECTNESS ). For any ML value p : term denoting a program, Eval2.main p Eval3.main p We identify that cont is a stack of thunks, and that the transitions are those of Krivine's abstract machine. 2.1.5 Krivine's abstract machine To obtain the canonical definition of Krivine's abstract machine, we abandon the distinction between denotable and expressible values and we use thunks instead, we represent the defunctionalized continuation as a list of thunks instead of a data type, and we inline apply cont . structure Eval4 = struct datatype thunk = THUNK of term * thunk list (* eval : term * thunk list * thunk list *) (* -&gt; term * thunk list *) fun eval (IND n, e, s) = let val (THUNK (t, e')) = List.nth (e, n) in eval (t, e', s) end | eval (ABS t, e, nil) = (ABS t, e) | eval (ABS t, e, (t', e') :: s) = eval (t, (THUNK (t', e')) :: e, s) | eval (APP (t0, t1), e, s) = eval (t0, e, (t1, e) :: s) (* main : term -&gt; term * thunk list *) fun main t = eval (t, nil, nil) end The following proposition is straightforward to prove. P ROPOSITION 4 ( FULL CORRECTNESS ). For any ML value p : term denoting a program, Eval3.main p Eval4.main p For comparison with Eval4 , the canonical definition of Krivine's abstract machine is as follows [7, 22, 25], where t denotes terms, v denotes expressible values, e denotes environments, and s denotes stacks of expressible values: Source syntax: t :: n t t 0 t 1 Expressible values (closures): v :: t e Initial transition, transition rules, and final transition: t t nil nil n e s t e s where t e nth e n t e t e :: s t t e :: e s t 0 t 1 e s t 0 e t 1 e :: s t e nil t e Variables n are represented by their de Bruijn index, and the abstract machine operates on triples consisting of a term, an environment, and a stack of expressible values. Each line in the canonical definition matches a clause in Eval4 . We conclude that Krivine's abstract machine can be seen as a defunctionalized , CPS-transformed, and closure-converted version of the standard call-by-name evaluator for the -calculus. This evaluator evidently implements Hardin, Maranget, and Pagano's K strategy [24, Section 3]. 2.2 From the CEK machine to a call-by-value evaluator The CEK machine [16, 17, 19] operates on -terms with names and distinguishes between values and computations in their syntax (i.e., it distinguishes trivial and serious terms, in Reynolds's words [31]). datatype term = VALUE of value | COMP of comp and value = VAR of string (* name *) | LAM of string * term and comp = APP of term * term Programs are closed terms. 2.2.1 The CEK abstract machine Our starting point reads as follows [19, Figure 2, page 239], where t denotes terms, w denotes values, v denotes expressible values, k denotes evaluation contexts, and e denotes environments: Source syntax: t :: w t 0 t 1 w :: x x t Expressible values (closures) and evaluation contexts: v :: x t e k :: stop fun v k arg t e k Initial transition, transition rules (two kinds), and final transition : t init t mt stop w e k eval k w e t 0 t 1 e k eval t 0 e arg t 1 e k arg t 1 e k v cont t 1 e fun v k fun x t e k v cont t e x v k stop v final v where x e e x x t e x t e Variables x are represented by their name, and the abstract machine consists of two mutually recursive transition functions. The first transition function operates on triples consisting of a term, an environment , and an evaluation context. The second operates on pairs consisting of an evaluation context and an expressible value. Environments are extended in the fun -transition, and consulted in . The empty environment is denoted by mt. This specification is straightforward to program in ML: 11 signature ENV = sig type 'a env val mt : 'a env val lookup : 'a env * string -&gt; 'a val extend : string * 'a * 'a env -&gt; 'a env end Environments are represented as a structure Env : ENV containing a representation of the empty environment mt , an operation lookup to retrieve the value bound to a name in an environment, and an operation extend to extend an environment with a binding. structure Eval0 = struct datatype expval = CLOSURE of string * term * expval Env.env datatype ev_context = STOP | ARG of term * expval Env.env * ev_context | FUN of expval * ev_context (* eval : term * expval Env.env * ev_context *) (* -&gt; expval *) fun eval (VALUE v, e, k) = continue (k, eval_value (v, e)) | eval (COMP (APP (t0, t1)), e, k) = eval (t0, e, ARG (t1, e, k)) and eval_value (VAR x, e) = Env.lookup (e, x) | eval_value (LAM (x, t), e) = CLOSURE (x, t, e) and continue (STOP, w) = w | continue (ARG (t1, e, k), w) = eval (t1, e, FUN (w, k)) | continue (FUN (CLOSURE (x, t, e), k), w) = eval (t, Env.extend (x, w, e), k) (* main : term -&gt; expval *) fun main t = eval (t, Env.mt, STOP) end 2.2.2 Refunctionalizing the evaluation contexts into continuations We identify that the data type ev context and the function continue are a defunctionalized representation. The corresponding higher-order evaluator reads as follows. As can be observed, it is in continuation-passing style. structure Eval1 = struct datatype expval = CLOSURE of string * term * expval Env.env (* eval : term * expval Env.env * (expval -&gt; 'a) *) (* -&gt; 'a *) fun eval (VALUE v, e, k) = k (eval_value (v, e)) | eval (COMP (APP (t0, t1)), e, k) = eval (t0, e, fn (CLOSURE (x, t, e')) =&gt; eval (t1, e, fn w =&gt; eval (t, Env.extend (x, w, e'), k))) and eval_value (VAR x, e) = Env.lookup (e, x) | eval_value (LAM (x, t), e) = CLOSURE (x, t, e) (* main : term -&gt; expval *) fun main t = eval (t, Env.mt, fn w =&gt; w) end The following proposition is a corollary of the correctness of defunctionalization . (Observational equivalence reduces here to structural equality over ML values of type expval .) P ROPOSITION 5 ( FULL CORRECTNESS ). For any ML value p : term denoting a program, Eval0.main p Eval1.main p 2.2.3 Back to direct style CPS-transforming the following direct-style evaluator yields the evaluator of Section 2.2.2 [10]. structure Eval2 = struct datatype expval = CLOSURE of string * term * expval Env.env (* eval : term * expval Env.env -&gt; expval *) fun eval (VALUE v, e) = eval_value (v, e) | eval (COMP (APP (t0, t1)), e) = let val (CLOSURE (x, t, e')) = eval (t0, e) val w = eval (t1, e) in eval (t, Env.extend (x, w, e')) end and eval_value (VAR x, e) = Env.lookup (e, x) | eval_value (LAM (x, t), e) = CLOSURE (x, t, e) (* main : term -&gt; expval *) fun main t = eval (t, Env.mt) end The following proposition is a corollary of the correctness of the direct-style transformation. (Again, observational equivalence reduces here to structural equality over ML values of type expval .) P ROPOSITION 6 ( FULL CORRECTNESS ). For any ML value p : term denoting a program, Eval1.main p Eval2.main p 2.2.4 From closures to higher-order functions We observe that the closures, in Eval2 , are defunctionalized representations with an apply function inlined. The corresponding higher-order evaluator reads as follows. structure Eval3 = struct datatype expval = CLOSURE of expval -&gt; expval (* eval : term * expval Env.env -&gt; expval *) fun eval (VALUE v, e) = eval_value (v, e) 12 | eval (COMP (APP (t0, t1)), e) = let val (CLOSURE f) = eval (t0, e) val w = eval (t1, e) in f w end and eval_value (VAR x, e) = Env.lookup (e, x) | eval_value (LAM (x, t), e) = CLOSURE (fn w =&gt; eval (t, Env.extend (x, w, e))) (* main : term -&gt; expval *) fun main t = eval (t, Env.mt) end The following proposition is a corollary of the correctness of defunctionalization . P ROPOSITION 7 ( FULL CORRECTNESS ). For any ML value p : term denoting a program, evaluating Eval2.main p yields a value CLOSURE (x, t, e) and evaluating Eval3.main p yields a value CLOSURE f such that fn w =&gt; Eval2.eval (t, Env.extend (x, w, e)) f 2.2.5 A higher-order and compositional call-by-value evaluator The result in Eval3 is a call-by-value evaluator that is compositional and higher-order. This call-by-value evaluator is the canonical one for the -calculus [31, 35, 37]. We conclude that the CEK machine can be seen as a defunctionalized, CPS-transformed, and closure-converted version of the standard call-by-value evaluator for -terms. 2.3 Variants of Krivine's machine and of the CEK machine It is easy to construct a variant of Krivine's abstract machine for terms with names, by starting from a call-by-name evaluator for -terms with names. Similarly, it is easy to construct a variant of the CEK machine for -terms with de Bruijn indices, by starting from a call-by-value evaluator for -terms with indices. It is equally easy to start from a call-by-value evaluator for -terms with de Bruijn indices and no distinction between values and computations ; the resulting abstract machine coincides with Hankin's eager machine [22, Section 8.1.2]. Abstract machines processing -terms with de Bruijn indices often resolve indices with transitions: 0 v :: e s v :: s n 1 v :: e s n e s Compared to the evaluator of Section 2.1.1, the evaluator corresponding to this machine has List.nth inlined and is not compositional : fun eval (IND 0, denval :: e, s) = ... denval ... | eval (IND n, denval :: e, s) = eval (IND (n - 1), e, s) | ... 2.4 Conclusion We have shown that Krivine's abstract machine and the CEK abstract machine are counterparts of canonical evaluators for call-by-name and for call-by-value -terms, respectively. The derivation of Krivine's machine is strikingly simpler than what can be found in the literature. That the CEK machine can be derived is, to the best of our knowledge, new. That these two machines are two sides of the same coin is also new. We have not explored any other aspect of this call-by-name/call-by-value duality [9]. Using substitutions instead of environments or inlining one of the standard computational monads (state, continuations, etc. [39]) in the call-by-value evaluator yields variants of the CEK machine that have been documented in the literature [16, Chapter 8]. For example , inlining the state monad in a monadic evaluator yields a state-passing evaluator. The corresponding abstract machine has one more component to represent the state. In general, inlining monads provides a generic recipe to construct arbitrarily many new abstract machines. It does not seem as straightforward, however, to construct a "monadic abstract machine" and then to inline a monad; we are currently studying the issue. On another note, one can consider an evaluator for strictness-annotated -terms--represented either with names or with indices, and with or without distinction between values and computations. One is then led to an abstract machine that generalizes Krivine's machine and the CEK machine [13]. Finally, it is straightforward to extend Krivine's machine and the CEK machine to bigger source languages (with literals, primitive operations, conditional expressions, block structure, recursion, etc.), by starting from evaluators for these bigger languages. For example, all the abstract machines in "The essence of compiling with continuations" [19] are defunctionalized continuation-passing evaluators, i.e., interpreters. In the rest of this article, we illustrate further the correspondence between evaluators and abstract machines. The CLS abstract machine The CLS abstract machine is due to Hannan and Miller [23]. In the following, t denotes terms, v denotes expressible values, c denotes lists of directives (a term or the special tag ap ), e denotes environments , l denotes stacks of environments, and s denotes stacks of expressible values. Source syntax: t :: n t t 0 t 1 Expressible values (closures): v :: t e Initial transition, transition rules, and final transition: t t :: nil nil :: nil nil t :: c e :: l s c l t e :: s t 0 t 1 :: c e :: l s t 0 :: t 1 :: ap :: c e :: e :: l s 0 :: c v :: e :: l s c l v :: s n 1 :: c v :: e :: l s n :: c e :: l s ap :: c l v :: t e :: s t :: c v :: e :: l s nil nil v :: s v 13 Variables n are represented by their de Bruijn index, and the abstract machine operates on triples consisting of a list of directives, a stack of environments, and a stack of expressible values. 3.1 The CLS machine Hannan and Miller's specification is straightforward to program in ML: datatype term = IND of int (* de Bruijn index *) | ABS of term | APP of term * term Programs are closed terms. structure Eval0 = struct datatype directive = TERM of term | AP datatype env = ENV of expval list and expval = CLOSURE of term * env (* run : directive list * env list * expval list *) (* -&gt; expval *) fun run (nil, nil, v :: s) = v | run ((TERM (IND 0)) :: c, (ENV (v :: e)) :: l, s) = run (c, l, v :: s) | run ((TERM (IND n)) :: c, (ENV (v :: e)) :: l, s) = run ((TERM (IND (n - 1))) :: c, (ENV e) :: l, s) | run ((TERM (ABS t)) :: c, e :: l, s) = run (c, l, (CLOSURE (t, e)) :: s) | run ((TERM (APP (t0, t1))) :: c, e :: l, s) = run ((TERM t0) :: (TERM t1) :: AP :: c, e :: e :: l, s) | run (AP :: c, l, v :: (CLOSURE (t, ENV e)) :: s) = run ((TERM t) :: c, (ENV (v :: e)) :: l, s) (* main : term -&gt; expval *) fun main t = run ((TERM t) :: nil, (ENV nil) :: nil, nil) end 3.2 A disentangled definition of the CLS machine In the definition of Section 3.1, all the possible transitions are meshed together in one recursive function, run . Instead, let us factor run into several mutually recursive functions, each of them with one induction variable. In this disentangled definition, run c interprets the list of control directives, i.e., it specifies which transition to take if the list is empty, starts with a term, or starts with an apply directive. If the list is empty, the computation terminates. If the list starts with a term, run t is called, caching the term in the first parameter. If the list starts with an apply directive, run a is called. run t interprets the top term in the list of control directives. run a interprets the top value in the current stack. The disentangled definition reads as follows: structure Eval1 = struct datatype directive = TERM of term | AP datatype env = ENV of expval list and expval = CLOSURE of term * env (* run_c : directive list * env list * expval list *) (* -&gt; expval *) fun run_c (nil, nil, v :: s) = v | run_c ((TERM t) :: c, l, s) = run_t (t, c, l, s) | run_c (AP :: c, l, s) = run_a (c, l, s) and run_t (IND 0, c, (ENV (v :: e)) :: l, s) = run_c (c, l, v :: s) | run_t (IND n, c, (ENV (v :: e)) :: l, s) = run_t (IND (n - 1), c, (ENV e) :: l, s) | run_t (ABS t, c, e :: l, s) = run_c (c, l, (CLOSURE (t, e)) :: s) | run_t (APP (t0, t1), c, e :: l, s) = run_t (t0, (TERM t1) :: AP :: c, e :: e :: l, s) and run_a (c, l, v :: (CLOSURE (t, ENV e)) :: s) = run_t (t, c, (ENV (v :: e)) :: l, s) (* main : term -&gt; expval *) fun main t = run_t (t, nil, (ENV nil) :: nil, nil) end P ROPOSITION 8 ( FULL CORRECTNESS ). For any ML value p : term denoting a program, Eval0.main p Eval1.main p P ROOF . By fold-unfold [5]. The invariants are as follows. For any ML values t : term , e : expval list , and s : expval list , Eval1.run c (c, l, s) Eval0.run (c, l, s) Eval1.run t (t, c, l, s) Eval0.run ((TERM t) :: c, l, s) Eval1.run a (c, l, s) Eval0.run (AP :: c, l, s) 3.3 The evaluator corresponding to the CLS machine In the disentangled definition of Section 3.2, there are three possible ways to construct a list of control directives (nil, cons'ing a term, and cons'ing an apply directive). We could specify these constructions as a data type rather than as a list. Such a data type, together with run c , is in the image of defunctionalization ( run c is the apply functions of the data type). The corresponding higher-order evaluator is in continuation-passing style. Transforming it back to direct style yields the following evaluator: structure Eval3 = struct datatype env = ENV of expval list and expval = CLOSURE of term * env (* run_t : term * env list * expval list *) (* -&gt; env list * expval list *) fun run_t (IND 0, (ENV (v :: e)) :: l, s) = (l, v :: s) | run_t (IND n, (ENV (v :: e)) :: l, s) = run_t (IND (n - 1), (ENV e) :: l, s) | run_t (ABS t, e :: l, s) = (l, (CLOSURE (t, e)) :: s) 14 | run_t (APP (t0, t1), e :: l, s) = let val (l, s) = run_t (t0, e :: e :: l, s) val (l, s) = run_t (t1, l, s) in run_a (l, s) end and run_a (l, v :: (CLOSURE (t, ENV e)) :: s) = run_t (t, (ENV (v :: e)) :: l, s) (* main : term -&gt; expval *) fun main t = let val (nil, v :: s) = run_t (t, (ENV nil) :: nil, nil) in v end end The following proposition is a corollary of the correctness of defunctionalization and of the CPS transformation. (Here observational equivalence reduces to structural equality over ML values of type expval .) P ROPOSITION 9 ( FULL CORRECTNESS ). For any ML value p : term denoting a program, Eval1.main p Eval3.main p As in Section 2, this evaluator can be made compositional by refunctionalizing the closures into higher-order functions and by factoring the resolution of de Bruijn indices into an auxiliary lookup function. We conclude that the evaluation model embodied in the CLS machine is a call-by-value interpreter threading a stack of environments and a stack of intermediate results with a caller-save strategy (witness the duplication of environments on the stack in the meaning of applications) and with a left-to-right evaluation of sub-terms. In particular, the meaning of a term is a partial endofunction over a stack of environments and a stack of intermediate results. The SECD abstract machine The SECD abstract machine is due to Landin [26]. In the following , t denotes terms, v denotes expressible values, c denotes lists of directives (a term or the special tag ap ), e denotes environments, s denotes stacks of expressible values, and d denotes dumps (list of triples consisting of a stack, an environment and a list of directives). Source syntax: t :: x x t t 0 t 1 Expressible values (closures): v :: x t e Initial transition, transition rules, and final transition: t nil mt t :: nil nil s e x :: c d e x :: s e c d s e x t :: c d x t e :: s e c d s e t 0 t 1 :: c d s e t 1 :: t 0 :: ap :: c d x t e :: v :: s e ap :: c d nil e x v t :: nil d where d s e c :: d v :: s e nil s e d :: d v :: s e c d v :: s e nil nil v Variables x are represented by their name, and the abstract machine operates on quadruples consisting of a stack of expressible values, an environment, a list of directives, and a dump. Environments are consulted in the first transition rule, and extended in the fourth. The empty environment is denoted by mt. 4.1 The SECD machine Landin's specification is straightforward to program in ML. Programs are closed terms. Environments are as in Section 2.2. datatype term = VAR of string (* name *) | LAM of string * term | APP of term * term structure Eval0 = struct datatype directive = TERM of term | AP datatype value = CLOSURE of string * term * value Env.env fun run (v :: nil, e', nil, nil) = v | run (s, e, (TERM (VAR x)) :: c, d) = run ((Env.lookup (e, x)) :: s, e, c, d) | run (s, e, (TERM (LAM (x, t))) :: c, d) = run ((CLOSURE (x, t, e)) :: s, e, c, d) | run (s, e, (TERM (APP (t0, t1))) :: c, d) = run (s, e, (TERM t1) :: (TERM t0) :: AP :: c, d) | run ((CLOSURE (x, t, e')) :: v :: s, e, AP :: c, d) = run (nil, Env.extend (x, v, e'), (TERM t) :: nil, (s, e, c) :: d) | run (v :: nil, e', nil, (s, e, c) :: d) = run (v :: s, e, c, d) (* main : term -&gt; value *) fun main t = run (nil, Env.mt, (TERM t) :: nil, nil) end 4.2 A disentangled definition of the SECD machine As in the CLS machine, in the definition of Section 4.1, all the possible transitions are meshed together in one recursive function, run . Instead, we can factor run into several mutually recursive functions, each of them with one induction variable. These mutually recursive functions are in defunctionalized form: the one processing the dump is an apply function for the data type representing the dump (a list of stacks, environments, and lists of directives), and the one processing the control is an apply function for the data type representing the control (a list of directives). The corresponding higher-order evaluator is in continuation-passing style with two nested continuations and one control delimiter, reset [12, 18]. The delimiter resets the control continuation when evaluating the body of a abstraction . (More detail is available in a technical report [11].) 15 4.3 The evaluator corresponding to the SECD machine The direct-style version of the evaluator from Section 4.2 reads as follows: structure Eval4 = struct datatype value = CLOSURE of string * term * value Env.env (* eval : term * value list * value Env.env *) (* -&gt; value list * value Env.env *) fun eval (VAR x, s, e) = ((Env.lookup (x, e)) :: s, e) | eval (LAM (x, t), s, e) = ((CLOSURE (x, t, e)) :: s, e) | eval (APP (t0, t1), s, e) = let val (s, e) = eval (t1, s, e) val (s, e) = eval (t0, s, e) in apply (s, e) end and apply ((CLOSURE (x, t, e')) :: v :: s, e) = let val (v :: nil, _) = reset (fn () =&gt; eval (t, nil, Env.extend (x, v, e'))) in (v :: s, e) end (* main : term -&gt; value *) fun main t = let val (v :: nil, _) = reset (fn () =&gt; eval (t, nil, Env.mt)) in v end end The following proposition is a corollary of the correctness of defunctionalization and of the CPS transformation. (Here observational equivalence reduces to structural equality over ML values of type value .) P ROPOSITION 10 ( FULL CORRECTNESS ). For any ML value p : term denoting a program, Eval0.main p Eval4.main p As in Sections 2 and 3, this evaluator can be made compositional by refunctionalizing the closures into higher-order functions. We conclude that the evaluation model embodied in the SECD machine is a call-by-value interpreter threading a stack of intermediate results and an environment with a callee-save strategy (witness the dynamic passage of environments in the meaning of applications), a right-to-left evaluation of sub-terms, and a control delimiter. In particular, the meaning of a term is a partial endofunction over a stack of intermediate results and an environment. Furthermore, this evaluator evidently implements Hardin, Maranget, and Pagano's L strategy, i.e., right-to-left call by value, without us having to "guess" its inference rules [24, Section 4]. The denotational content of the SECD machine puts a new light on it. For example, its separation between a control register and a dump register is explained by the control delimiter in the evaluator ( reset in Eval4.eval ). 2 Removing this control delimiter gives rise to an abstract machine with a single stack component for control-not by a clever change in the machine itself, but by a straightforward simplification in the corresponding evaluator. Variants of the CLS machine and of the SECD machine It is straightforward to construct a variant of the CLS machine for -terms with names, by starting from an evaluator for -term with names. Similarly, it is straightforward to construct a variant of the SECD machine for -terms with de Bruijn indices, by starting from an evaluator for -term with indices. In the same vein, it is simple to construct call-by-name versions of the CLS machine and of the SECD machine, by starting from call-by-name evaluators. It is also simple to construct a properly tail recursive version of the SECD machine, and to extend the CLS machine and the SECD machine to bigger source languages, by extending the corresponding evaluator. The Categorical Abstract Machine What is the difference between an abstract machine and a virtual machine? Elsewhere [1], we propose to distinguish them based on the notion of instruction set: A virtual machine has an instruction set whereas an abstract machine does not. Therefore, an abstract machine directly operates on a -term, but a virtual machine operates on a compiled representation of a -term, expressed using an instruction set. (This distinction can be found elsewhere in the literature [21].) The Categorical Abstract Machine [6], for example, has an instruction set--categorical combinators--and therefore (despite its name) it is a virtual machine, not an abstract machine. In contrast , Krivine's machine, the CEK machine, the CLS machine, and the SECD machine are all abstract machines, not virtual machines, since they directly operate on -terms. In this section, we present the abstract machine corresponding to the Categorical Abstract Machine (CAM). We start from the evaluation model embodied in the CAM [1]. 6.1 The evaluator corresponding to the CAM The evaluation model embodied in the CAM is an interpreter threading a stack with its top element cached in a register, representing environments as expressible values (namely nested pairs linked as lists), with a caller-save strategy (witness the duplication of the register on the stack in the meaning of applications below), and with a left-to-right evaluation of sub-terms. In particular, the meaning of a term is a partial endofunction over the register and the stack. This evaluator reads as follows: datatype term = IND of int (* de Bruijn index *) | ABS of term | APP of term * term | NIL | CONS of term * term | CAR of term | CDR of term Programs are closed terms. 2 A rough definition of reset is fun reset t = t () . A more accurate definition, however, falls out of the scope of this article [12, 18]. 16 structure Eval0 = struct datatype expval = NULL | PAIR of expval * expval | CLOSURE of expval * (expval * expval list -&gt; expval * expval list) (* access : int * expval * expval list *) (* -&gt; expval * expval list *) fun access (0, PAIR (v1, v2), s) = (v2, s) | access (n, PAIR (v1, v2), s) = access (n - 1, v1, s) (* eval : term * expval * expval list *) (* -&gt; expval * expval list *) fun eval (IND n, v, s) = access (n, v, s) | eval (ABS t, v, s) = (CLOSURE (v, fn (v, s) =&gt; eval (t, v, s)), s) | eval (APP (t0, t1), v, s) = let val (v, v' :: s) = eval (t0, v, v :: s) val (v', (CLOSURE (v, f)) :: s) = eval (t1, v', v :: s) in f (PAIR (v, v'), s) end | eval (NIL, v, s) = (NULL, s) | eval (CONS (t1, t2), v, s) = let val (v, v' :: s) = eval (t1, v, v :: s) val (v, v' :: s) = eval (t2, v', v :: s) in (PAIR (v', v), s) end | eval (CAR t, v, s) = let val (PAIR (v1, v2), s) = eval (t, v, s) in (v1, s) end | eval (CDR t, v, s) = let val (PAIR (v1, v2), s) = eval (t, v, s) in (v2, s) end (* main : term -&gt; expval *) fun main t = let val (v, nil) = eval (t, NULL, nil) in v end end This evaluator evidently implements Hardin, Maranget, and Pagano's X strategy [24, Section 6]. 6.2 The abstract machine corresponding to the CAM As in Sections 2, 3, and 4, we can closure-convert the evaluator of Section 6.1 by defunctionalizing its expressible values, transform it into continuation-passing style, and defunctionalize its continuations . The resulting abstract machine reads as follows, where t denotes terms, v denotes expressible values, k denotes evaluation contexts, and s denotes stacks of expressible values. Source syntax: t :: n t t 0 t 1 nil cons t 1 t 2 car t cdr t Expressible values (unit value, pairs, and closures) and evaluation contexts: v :: null v 1 v 2 v t k :: CONT0 CONT1 t k CONT2 k CONT3 t k CONT4 k CONT5 k CONT6 k Initial transition, transition rules (two kinds), and final transition : t init t null nil CONT0 n v s k eval k n v s t v s k eval k v t s nil v s k eval k null s t 0 t 1 v s k eval t 0 v v :: s CONT1 t 1 k cons t 1 t 2 v s k eval t 1 v v :: s CONT3 t 2 k car t v s k eval t v s CONT5 k cdr t v s k eval t v s CONT6 k CONT1 t k v v :: s cont t v v :: s CONT2 k CONT2 k v v t :: s cont t v v s k CONT3 t 1 k v v :: s cont t 1 v v :: s CONT4 k CONT4 k v v :: s cont k v v s CONT5 k v 1 v 2 s cont k v 1 s CONT6 k v 1 v 2 s cont k v 2 s CONT0 v nil final v where 0 v 1 v 2 v 2 n v 1 v 2 n 1 v 1 Variables n are represented by their de Bruijn index, and the abstract machine consists of two mutually recursive transition functions . The first transition function operates on quadruples consisting of a term, an expressible value, a stack of expressible values, and an evaluation context. The second transition function operates on triples consisting of an evaluation context, an expressible value, and a stack of expressible values. This abstract machine embodies the evaluation model of the CAM. Naturally, more intuitive names could be chosen instead of CONT0 , CONT1 , etc. Conclusion and issues We have presented a constructive correspondence between functional evaluators and abstract machines. This correspondence builds on off-the-shelf program transformations: closure conversion , CPS transformation, defunctionalization, and inlining. 3 We have shown how to reconstruct known machines (Krivine's machine , the CEK machine, the CLS machine, and the SECD machine ) and how to construct new ones. Conversely, we have revealed the denotational content of known abstract machines. We have shown that Krivine's abstract machine and the CEK machine correspond to canonical evaluators for the -calculus. We have also shown that they are dual of each other since they correspond to call-by-name and call-by-value evaluators in the same direct style. In terms of denotational semantics [27, 34], Krivine's machine and the CEK machine correspond to a standard semantics, whereas the CLS machine and the SECD machine correspond to a stack semantics of the -calculus. Finally, we have exhibited the abstract machine corresponding to the CAM, which puts the reader in a new position to answer the recurrent question as to whether the CLS machine is closer to the CAM or to the SECD machine. 3 Indeed the push-enter twist of Krivine's machine is obtained by inlining apply cont in Section 2.1.5. 17 Since this article was written, we have studied the correspondence between functional evaluators and abstract machines for call by need [2] and for Propositional Prolog [4]. In both cases, we derived sensible machines out of canonical evaluators. It seems to us that this correspondence between functional evaluators and abstract machines builds a reliable bridge between denotational definitions and definitions of abstract machines. On the one hand, it allows one to identify the denotational content of an abstract machine in the form of a functional interpreter. On the other hand, it gives one a precise and generic recipe to construct arbitrarily many new variants of abstract machines (e.g., with substitutions or environments, or with stacks) or of arbitrarily many new abstract machines, starting from an evaluator with any given computational monad [28]. Acknowledgments: We are grateful to Malgorzata Biernacka, Julia Lawall, and Henning Korsholm Rohde for timely comments. Thanks are also due to the anonymous reviewers. This work is supported by the ESPRIT Working Group APPSEM II ( http://www.appsem.org ). References [1] Mads Sig Ager, Dariusz Biernacki, Olivier Danvy, and Jan Midtgaard. From interpreter to compiler and virtual machine : a functional derivation. Technical Report BRICS RS-03 -14, DAIMI, Department of Computer Science, University of Aarhus, Aarhus, Denmark, March 2003. [2] Mads Sig Ager, Olivier Danvy, and Jan Midtgaard. A functional correspondence between call-by-need evaluators and lazy abstract machines. Technical Report BRICS RS-03-24 , DAIMI, Department of Computer Science, University of Aarhus, Aarhus, Denmark, June 2003. [3] Anindya Banerjee, Nevin Heintze, and Jon G. Riecke. Design and correctness of program transformations based on control-flow analysis. In Naoki Kobayashi and Benjamin C. Pierce, editors, Theoretical Aspects of Computer Software, 4th International Symposium, TACS 2001, number 2215 in Lecture Notes in Computer Science, Sendai, Japan, October 2001. Springer-Verlag. [4] Dariusz Biernacki and Olivier Danvy. From interpreter to logic engine: A functional derivation. Technical Report BRICS RS-03-25, DAIMI, Department of Computer Science, University of Aarhus, Aarhus, Denmark, June 2003. Accepted for presentation at LOPSTR 2003. [5] Rod M. Burstall and John Darlington. A transformational system for developing recursive programs. Journal of ACM, 24(1):4467, 1977. [6] Guy Cousineau, Pierre-Louis Curien, and Michel Mauny. The categorical abstract machine. Science of Computer Programming , 8(2):173202, 1987. [7] Pierre Cregut. An abstract machine for lambda-terms normalization . In Mitchell Wand, editor, Proceedings of the 1990 ACM Conference on Lisp and Functional Programming, pages 333340, Nice, France, June 1990. ACM Press. [8] Pierre-Louis Curien. Categorical Combinators, Sequential Algorithms and Functional Programming. Progress in Theoretical Computer Science. Birkhauser, 1993. [9] Pierre-Louis Curien and Hugo Herbelin. The duality of computation . In Philip Wadler, editor, Proceedings of the 2000 ACM SIGPLAN International Conference on Functional Programming , SIGPLAN Notices, Vol. 35, No. 9, pages 233 243, Montreal, Canada, September 2000. ACM Press. [10] Olivier Danvy. Back to direct style. Science of Computer Programming, 22(3):183195, 1994. [11] Olivier Danvy. A lambda-revelation of the SECD machine. Technical Report BRICS RS-02-53, DAIMI, Department of Computer Science, University of Aarhus, Aarhus, Denmark, December 2002. [12] Olivier Danvy and Andrzej Filinski. Representing control, a study of the CPS transformation. Mathematical Structures in Computer Science, 2(4):361391, 1992. [13] Olivier Danvy and John Hatcliff. CPS transformation after strictness analysis. ACM Letters on Programming Languages and Systems, 1(3):195212, 1993. [14] Olivier Danvy and John Hatcliff. On the transformation between direct and continuation semantics. In Stephen Brookes, Michael Main, Austin Melton, Michael Mislove, and David Schmidt, editors, Proceedings of the 9th Conference on Mathematical Foundations of Programming Semantics, number 802 in Lecture Notes in Computer Science, pages 627648, New Orleans, Louisiana, April 1993. Springer-Verlag. [15] Olivier Danvy and Lasse R. Nielsen. Defunctionalization at work. In Harald Sndergaard, editor, Proceedings of the Third International ACM SIGPLAN Conference on Principles and Practice of Declarative Programming (PPDP'01), pages 162174, Firenze, Italy, September 2001. ACM Press. Extended version available as the technical report BRICS RS-01 -23. [16] Matthias Felleisen and Matthew Flatt. Programming languages and lambda calculi. Unpublished lecture notes. http://www.ccs.neu.edu/home/matthias/3810-w02/ readings.html , 1989-2003. [17] Matthias Felleisen and Daniel P. Friedman. Control operators, the SECD machine, and the -calculus. In Martin Wirsing, editor , Formal Description of Programming Concepts III, pages 193217. Elsevier Science Publishers B.V. (North-Holland), Amsterdam, 1986. [18] Andrzej Filinski. Representing monads. In Hans-J. Boehm, editor, Proceedings of the Twenty-First Annual ACM Symposium on Principles of Programming Languages, pages 446 457, Portland, Oregon, January 1994. ACM Press. [19] Cormac Flanagan, Amr Sabry, Bruce F. Duba, and Matthias Felleisen. The essence of compiling with continuations. In David W. Wall, editor, Proceedings of the ACM SIGPLAN'93 Conference on Programming Languages Design and Imple-mentation , SIGPLAN Notices, Vol. 28, No 6, pages 237247, Albuquerque, New Mexico, June 1993. ACM Press. [20] Daniel P. Friedman, Mitchell Wand, and Christopher T. Haynes. Essentials of Programming Languages, second edition . The MIT Press, 2001. [21] Benjamin Gregoire and Xavier Leroy. A compiled implementation of strong reduction. In Simon Peyton Jones, editor , Proceedings of the 2002 ACM SIGPLAN International Conference on Functional Programming, SIGPLAN Notices, Vol. 37, No. 9, pages 235246, Pittsburgh, Pennsylvania, September 2002. ACM Press. 18 [22] Chris Hankin. Lambda Calculi, a guide for computer scientists , volume 1 of Graduate Texts in Computer Science. Oxford University Press, 1994. [23] John Hannan and Dale Miller. From operational semantics to abstract machines. Mathematical Structures in Computer Science, 2(4):415459, 1992. [24] Ther`ese Hardin, Luc Maranget, and Bruno Pagano. Functional runtime systems within the lambda-sigma calculus. Journal of Functional Programming, 8(2):131172, 1998. [25] Jean-Louis Krivine. Un interpr`ete du -calcul. Brouil-lon . Available online at http://www.logique.jussieu.fr/ ~krivine , 1985. [26] Peter J. Landin. The mechanical evaluation of expressions. The Computer Journal, 6(4):308320, 1964. [27] Robert E. Milne and Christopher Strachey. A Theory of Programming Language Semantics. Chapman and Hall, London, and John Wiley, New York, 1976. [28] Eugenio Moggi. Notions of computation and monads. Information and Computation, 93:5592, 1991. [29] Lasse R. Nielsen. A denotational investigation of defunctionalization . Technical Report BRICS RS-00-47, DAIMI, Department of Computer Science, University of Aarhus, Aarhus, Denmark, December 2000. [30] Gordon D. Plotkin. Call-by-name, call-by-value and the calculus . Theoretical Computer Science, 1:125159, 1975. [31] John C. Reynolds. Definitional interpreters for higher-order programming languages. Higher-Order and Symbolic Computation , 11(4):363397, 1998. Reprinted from the proceedings of the 25th ACM National Conference (1972). [32] Kristoffer H. Rose. Explicit substitution tutorial & survey . BRICS Lecture Series LS-96-3, DAIMI, Department of Computer Science, University of Aarhus, Aarhus, Denmark, September 1996. [33] Amr Sabry and Matthias Felleisen. Reasoning about programs in continuation-passing style. Lisp and Symbolic Computation , 6(3/4):289360, 1993. [34] David A. Schmidt. Denotational Semantics: A Methodology for Language Development. Allyn and Bacon, Inc., 1986. [35] Guy L. Steele Jr. and Gerald J. Sussman. The art of the interpreter or, the modularity complex (parts zero, one, and two). AI Memo 453, Artificial Intelligence Laboratory , Massachusetts Institute of Technology, Cambridge, Massachusetts , May 1978. [36] Joseph Stoy. Some mathematical aspects of functional programming. In John Darlington, Peter Henderson, and David A. Turner, editors, Functional Programming and its Applications. Cambridge University Press, 1982. [37] Joseph E. Stoy. Denotational Semantics: The Scott-Strachey Approach to Programming Language Theory. The MIT Press, 1977. [38] Christopher Strachey. Fundamental concepts in programming languages. Higher-Order and Symbolic Computation, 13(1/2):149, 2000. [39] Philip Wadler. The essence of functional programming (in-vited talk). In Andrew W. Appel, editor, Proceedings of the Nineteenth Annual ACM Symposium on Principles of Programming Languages, pages 114, Albuquerque, New Mexico , January 1992. ACM Press. [40] Mitchell Wand. A short proof of the lexical addressing algorithm . Information Processing Letters, 35:15, 1990. [41] Glynn Winskel. The Formal Semantics of Programming Languages . Foundation of Computing Series. The MIT Press, 1993. 19

call-by-name;Interpreters;closure conversion;evaluator;defunctionalization;call-by-value;transformation into continuation-passing style (CPS);abstract machines;abstract machine

110

Indexing Multi-Dimensional Time-Series with Support for Multiple Distance Measures

Although most time-series data mining research has concentrated on providing solutions for a single distance function, in this work we motivate the need for a single index structure that can support multiple distance measures. Our specific area of interest is the efficient retrieval and analysis of trajectory similarities. Trajectory datasets are very common in environmental applications, mobility experiments, video surveillance and are especially important for the discovery of certain biological patterns. Our primary similarity measure is based on the Longest Common Subsequence (LCSS) model, that offers enhanced robustness, particularly for noisy data, which are encountered very often in real world applications . However, our index is able to accommodate other distance measures as well, including the ubiquitous Euclidean distance, and the increasingly popular Dynamic Time Warping (DTW). While other researchers have advocated one or other of these similarity measures, a major contribution of our work is the ability to support all these measures without the need to restructure the index. Our framework guarantees no false dismissals and can also be tailored to provide much faster response time at the expense of slightly reduced precision/recall. The experimental results demonstrate that our index can help speed-up the computation of expensive similarity measures such as the LCSS and the DTW.

INTRODUCTION In this work we present an efficient and compact, external memory index for fast detection of similar trajectories. Trajectory data are prevalent in diverse fields of interest such as meteorology, GPS tracking, wireless applications, video tracking [5] and motion capture [18]. Recent advances in mobile computing, sensor and GPS technology have made it possible to collect large amounts of spatiotemporal data and The research of this author was supported by NSF ITR 0220148, NSF CAREER 9907477, NSF IIS 9984729, and NRDRP Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SIGKDD '03, August 24-27, 2003, Washington, DC, USA. Copyright 2003 ACM 1-58113-737-0/03/0008... $ 5.00. there is increasing interest in performing data analysis tasks over such data [17]. In mobile computing, users equipped with mobile devices move in space and register their location at different time instances to spatiotemporal databases via wireless links. In environmental information systems, tracking animals and weather conditions is very common and large datasets can be created by storing locations of observed objects over time. Human motion data generated by tracking simultaneously various body joints are also multidimensional trajectories. In this field of computer graphics fundamental operations include the clustering of similar movements, leading to a multitude of applications such as interactive generation of motions [2]. Spatiotemporal data are also produced by migrating particles in biological sciences, where the focus can be on the discovery of subtle patterns during cellular mitoses [19]. In general, any dataset that involves storage of multiple streams (attributes) of data can be considered and treated as a multidimensional trajectory. One very common task for such data is the discovery of objects that follow a certain motion pattern, for purposes of clustering or classification. The objective here is to efficiently organize trajectories on disk, so that we can quickly answer k-Nearest-Neighbors (kNN) queries. A frequent obstacle in the analysis of spatiotemporal data, is the presence of noise, which can be introduced due to electromagnetic anomalies, transceiver problems etc. Another impediment is that objects may move in a similar way, but at different speeds. So, we would like our similarity model to be robust to noise, support elastic and imprecise matches. Choosing the Euclidean distance as the similarity model is unrealistic, since its performance degrades rapidly in the presence of noise and this measure is also sensitive to small variations in the time axis. We concentrate on two similarity models: the first is an extension of Dynamic Time Warping for higher dimensions. We note that DTW has been used so far for one-dimensional time series. Here we present a formulation for sequences of arbitrary dimensions. The second distance measure is a modification of the Longest Common Subsequence (LCSS), specially adapted for continuous values. Both measures represent a significant improvement compared to the Euclidean distance. However, LCSS is more robust than DTW under noisy conditions [20] as figure 1 shows. Euclidean matching completely disregards the variations in the time axis, while DTW performs excessive matchings, therefore distorting the true distance between sequences . The LCSS produces the most robust and intuitive correspondence between points. By incorporating warping in time as a requirement to 216 0 20 40 60 80 100 120 Euclidean Matching 0 20 40 60 80 100 120 Time Warping 0 20 40 60 80 100 120 Longest Common Subsequence Figure 1: A lucid example about the quality matching of the LCSS compared to other distance functions. The Euclidean distance performs an inflexible matching, while the DTW gives many superfluous and spurious matchings, in the presence of noise. our model, our algorithms are automatically challenged with quadratic execution time. Moreover, these flexible functions are typically non-metric, which makes difficult the design of indexing structures. To speed up the execution of a similarity function, one can devise a low cost, upper bounding function (since the LCSS model captures the similarity, which is inversely analogous to the distance). We utilize a fast prefiltering scheme that will return upper bound estimates for the LCSS similarity between the query and the indexed trajectories. In addition to providing similarity measures that guarantee no false dismissals, we also propose approximate similarity estimates that significantly reduce the index response time. Finally, we show that the same index can support other distance measures as well. Our technique works by splitting the trajectories in multidimensional MBRs and storing them in an R-tree. For a given query, we construct a Minimum Bounding Envelope (MBE) that covers all the possible matching areas of the query under warping conditions. This MBE is decomposed into MBRs and then probed in the R-tree index. Using the index we can discover which trajectories could potentially be similar to the query. The index size is compact and its construction time scales well with the trajectory length and the database size, therefore our method can be utilized for massive datamining tasks. The main contributions of the paper are: We present the first external memory index for multidimensional trajectories, that supports multiple distance functions (such as LCSS, DTW and Euclidean), without the need to rebuild the index. We give efficient techniques for upper(lower) bounding and for approximating the LCSS(DTW) for a set of trajectories . We incorporate these techniques in the design of an efficient indexing structure for the LCSS and the DTW. We provide a flexible method that allows the user to specify queries of variable warping length, and the technique can be tuned to optimize the retrieval time or the accuracy of the solution. RELATED WORK There has been a wealth of papers that use an L p distance family function to perform similarity matching for 1D time-series. Work on multidimensional sequences can be found in [14, 9]. However, they support only Euclidean distance, which, as mentioned in the introduction, cannot capture flexible similarities. Although the vast majority of database/data mining research on time series data mining has focused on Euclidean distance, virtually all real world systems that use time series matching as a subroutine, use a similarity measure which allows warping. In retrospect, this is not very surprising, since most real world processes, particularly biological processes, can evolve at varying rates. For example, in bioinformat-ics , it is well understood that functionally related genes will express themselves in similar ways, but possibly at different rates. Because of this, DTW is used for gene expression data mining [1, 3]. Dynamic Time Warping is a ubiquitous tool in the biometric/surveillance community. It has been used for tracking time series extracted from video [7], classifying handwritten text [16] and even fingerprint indexing [13]. While the above examples testify to the utility of a time warped distance measure, they all echo the same complaint; DTW has serious scalability issues. Work that attempted to mitigate the large computational cost has appeared in [12] and [21], where the authors use lower bounding measures to speed up the execution of DTW. However, the lower bounds can be loose approximations of the original distance, when the data are normalized. In [15] a different approach is used for indexing Time Warping, by using suffix trees. Nonetheless , the index requires excessive disk space (about 10 times the size of the original data). The flexibility provided by DTW is very important, however its efficiency deteriorates for noisy data, since by matching all the points, it also matches the outliers distorting the true distance between the sequences. An alternative approach is the use of Longest Common Subsequence (LCSS), which is a variation of the edit distance. The basic idea is to match two sequences by allowing them to stretch, without rearranging the order of the elements but allowing some elements to be unmatched. Using the LCSS of two sequences , one can define the distance using the length of this subsequence [6]. In [20] an internal memory index for the LCSS has been proposed. It also demonstrated that while the LCSS presents similar advantages to DTW, it does not share its volatile performance in the presence of outliers. Closest in spirit to our approach, is the work of [10] which, however, only addresses 1D time-series. The author uses constrained DTW as the distance function, and surrounds the possible matching regions by a modified version of a Piecewise Approximation, which is later stored as equi-length MBRs in an R-tree. However, by using DTW, such an approach is susceptible to high bias of outliers. Also, the 217 fixed MBR size (although simplifies the index operations) can lead to degenerate approximations of the original sequence . Moreover, the embedding of the envelope in the indexed sequences can slow the index construction time and limit the user's query capabilities to a predefined warping length. The use of LCSS as our primary similarity measure, lends itself to a more natural use of the R-tree, where the similarity estimates are simply computed by calculating the MBR intersection areas. Since the index is not constructed for a specific warping window, the user can pose queries with variable warping length. The purpose of this paper is to reconcile the best of both worlds. We provide a framework that can support in the same index, the LCSS, DTW and Euclidean distance functions . The only aspect that changes, is the different representation of the query for each distance measure. DISTANCE MEASURES In this section we present details of how the Dynamic Time Warping and the LCSS model can be extended to describe the similarity between trajectories. 3.1 Dynamic Time Warping for 2D trajectories We describe an extension in 2D of the original DTW function as described by Berndt and Clifford [4]. Let A and B be two trajectories of moving objects with size n and m respectively, where A = ((a x,1 , a y,1 ), . . . , (a x,n , a y,n )) and B = ((b x,1 , b y,1 ), . . . , (b x,m , b y,m )). For a trajectory A, let Head(A) = ((a x,1 , a y,1 ), . . . , (a x,n 1 , a y,n 1 )). Definition 1. The Time Warping between 2-dimensional sequences A and B is: DT W (A, B) = L p ((a x,n , a y,n ), (b x,m , b y,m )) + min {DTW(Head(A), Head(B)), DT W (Head(A), B), DT W (A, Head(B)) } (1) where L p is any p-Norm. Using dynamic programming and constraining the matching region within , the time required to compute DTW is O((n + m)). In order to represent an accurate relationship of distances between sequences with different lengths, the quantity in equation 1 is normalized by the length of the warping path. The extension to n dimensions is similar. In figure 2 we show an example of time warping for two trajectories. 3.2 LCSS model for 2D trajectories The original LCSS model refers to 1D sequences, we must therefore extend it to the 2D case. In addition, the LCSS paradigm matches discrete values, however in our model we want to allow a matching, when the values are within a certain range in space and time (note that like this, we also avoid distant and degenerate matchings). Definition 2. Given an integer and a real number 0 &lt; &lt; 1, we define the LCSS , (A, B) as follows: LCSS , (A, B) = 0 if A or B is empty 1 + LCSS , (Head(A), Head(B)) if |a x,n b x,m | &lt; and |a y,n b y,m | &lt; and |n m| max(LCSS , (Head(A), B), LCSS , (A, Head(B))), otherwise 0 50 100 150 0 500 1000 1500 100 200 300 400 500 600 X movement Time Y movement Figure 2: The support of flexible matching in spatiotemporal queries is very important. However, we can observe that Dynamic Time Warping matches all points (so the outliers as well), therefore distorting the true distance. In contrast, the LCSS model can efficiently ignore the noisy parts. where sequences A and Head(A) are defined similarly as before. The constant controls the flexibility of matching in time and constant is the matching threshold is space. The aforementioned LCSS model has the same O((n+m)) computational complexity as the DTW, when we only allow a matching window in time [6]. The value of LCSS is unbounded and depends on the length of the compared sequences. We need to normalize it, in order to support sequences of variable length. The distance derived from the LCSS similarity can be defined as follows: Definition 3. The distance D , expressed in terms of the LCSS similarity between two trajectories A and B is given by: D , (A, B) = 1 LCSS , (A, B) min(n, m) (2) INDEX CONSTRUCTION Even though imposing a matching window can help speed up the execution, the computation can still be quadratic when is a significant portion of the sequence's length. Therefore, comparing a query to all the trajectories becomes intractable for large databases. We are seeking ways to avoid examining the trajectories that are very distant to our query. This can be accomplished by discovering a close match to our query, as early as possible. A fast pre-filtering step is employed that eliminates the majority of distant matches. Only for some qualified sequences will we execute the costly (but accurate) quadratic time algorithm. This philosophy has also been successfully used in [21, 10]. There are certain preprocessing steps that we follow: 1. The trajectories are segmented into MBRs, which are stored in an Rtree T. 2. Given a query Q, we discover the areas of possible matching by constructing its Minimum Bounding Envelope (M BE Q ). 3. M BE Q is decomposed into MBRs that are probed in the index T. 218 4. Based on the MBR intersections, similarity estimates are computed and the exact LCSS (or DTW) is performed only on the qualified trajectories. The above notions are illustrated in figure 3 and we explain in detail how they can be applied for the LCSS case in the sections that follow. E. LCSS Upper Bound Estimate = L1+L2+L3 A. Query Q C. Envelope Splitting B. Query Envelope D. Sequence MBRs L1 L2 L3 Figure 3: An example of our approach (in 1D for clarity); A query is extended into a bounding envelope , which in turn is also split into the resulting MBRs. Overlap between the query and the index MBRs suggest areas of possible matching. 4.1 Bounding the Matching Regions Let us first consider a 1D time-series and let a sequence A be (a x,1 , . . . , a x,n ). Ignoring for now the parameter , we would like to perform a very fast LCSS match between sequence A and some query Q. Suppore that we replicate each point Q i for time instances before and after time i. The envelope that includes all these points defines the areas of possible matching. Everything outside this envelope can never be matched. 10 20 30 40 50 60 70 40 pts 6 pts 2 Q A Figure 4: The Minimum Bounding Envelope (MBE) within in time and in space of a sequence. Everything that lies outside this envelope can never be matched. We call this envelope, the Minimum Bounding Envelope (MBE) of a sequence. Also, once we incorporate the matching within in space, this envelope should extent above and below the original envelope (figure 4). The notion of the bounding envelope can be trivially extended in more dimensions , where M BE(, ) for a 2D trajectory Q = ((q x,1 , q y,1 ), . . . , (q x,n , q y,n ) covers the area between the following time-series : EnvLow MBE(, ) EnvHigh, where: EnvHigh[i] = max(Q[j] + epsilon) , |ij| EnvLow[j] = min(Q[j] epsilon) , |ij| The LCSS similarity between the envelope of Q and a sequence A is defined as: LCSS(M BE Q , A) = n i=1 1 if A[i] within envelope 0 otherwise For example, in figure 4 the LCSS similarity between M BE Q and sequence A is 46, as indicated in the figure. This value represents an upper bound for the similarity of Q and A. We can use the M BE Q to compute a lower bound on the distance between trajectories: Lemma 1. For any two trajectories Q and A the following holds: D , (M BE Q , A) D , (Q, A), Proof (Sketch): D , (M BE Q , A) = 1 LCSS , (M BE Q ,A) min( |Q|,|A|) , therefore it is sufficient to show that: LCSS , (M BE Q , A) LCSS , (Q, A). This is true since M BE Q by construction contains all possible areas within and of the query Q. Therefore, no possible matching points will be missed. 2 The previous lemma provides us with the power to create an index that guarantees no false dismissals. However, this lower bound refers to the raw data. In the sections that follow , we will 'split' the M BE of a trajectory, into a number of Minimum Bounding Rectangles (MBRs), to accommodate their storage into a multidimensional R-tree. We will show that the above inequality still holds between trajectory MBRs. The MBR generation procedure is orthogonal to our approach , since any segmentation methodology can be applied to our framework. Therefore, the description of the potential MBR generation methods (and of our implementation choice) will be delayed until later. QUICK PRUNING OF DISSIMILAR TRA-JECTORIES Suppose that we have an index with the segmented trajectories and the user provides a query Q. Our goal is the discovery of the k closest trajectories to the given query, according to the LCSS similarity. A prefiltering step will aid the quick discovery of a close match to the query, helping us discard the distant trajectories without using the costly quadratic algorithm. Therefore, in this phase, we compute upper bound estimates of the similarity between the query and the indexed sequences using their MBRs. Below we describe the algorithm to find the closest trajectory to a given query: Input: Query Q, Index I with trajectory MBRs, Method Output: Most similar trajectory to Q. 219 Box Env = constructM BE , (Q); Vector V Q = CreateM BRs(Env); // V Q contains a number of boxes. Priority queue P Q ; // P Q keeps one entry per trajectory sorted // according to the similarity estimate for each box B in V Q : V = I.intersectionQuery(B); // V contains all trajectory MBRs that intersect with B. if Method == Exact: // upper bound P Q computeL-SimilarityEstimates(V, B); else: // approximate P Q computeV-SimilarityEstimates(V, B); BestSoF ar = 0; Best ; while P Q not empty: E PQ.top; if E.estimate &lt; BestSoF ar: break; else: D = computeLCCS , (Q, E); // exact if D &gt; BestSoF ar: BestSoF ar = D; Best E; Report Best; The above algorithm can be adjusted to return the kNN sequences, simply by comparing with the k th bestSoF ar match. Next, we examine the possible similarity estimates. Some of them guarantee that will find the best match (they lower bound the original distance or upper bound the original similarity), while other estimates provide faster but approximate results. 5.1 Similarity Estimates Here we will show how to compute estimates of the LCSS similarity, based on the geometric properties of the trajectory MBRs and their intersection. An upper bound estimate is provided by the length of the MBR intersection and an approximate estimate is given as a parameter of the intersecting volume. To formalize these notions, first we present several operators. Then we will use these operators to derive the estimates. 5.1.1 Estimates for the LCSS Each trajectory T can be decomposed into a number of MBRs. The i th 3D MBR of T consists of six numbers: M T,i = {t l , t h , x l , x h , y l , y h }. Now, let us define the operators (c) t , (p) t and V between two 3D MBRs M P,i and M R,j , belonging to objects P and R, respectively: 1. (c) t (M P,i , M R,j ) = ||Intersection|| t , where M R,j .x l M P,i .x l M R,j .x h and M R,j .x l M P,i .x h M R,j .x h and M R,j .y l M P,i .y l M R,j .y h and M R,j .y l M P,i .y h M R,j .y h or similarly by rotating M R,j M P,i Therefore, this operator computes the time intersection of two MBR when one fully contains the other in the x,y dimensions. 2. (p) t (M P,i , M R,j ) = ||Intersection|| t , otherwise 3. V (M P,i , M R,j ) = ||Intersection|| t ||Intersection|| x ||Intersection|| y We can use upper bound or approximate estimates for the similarity: Common Volume Intersection The Intersection of MBRs is fully contained within one MBR Intersection between two MBRs time y x Figure 5: Top left: Intersection recorded in list L t,partial . Top right: Intersection recorded in list L t,complete . Bottom left: Percentage of Volume Intersection kept in L V . 1. Upper bound estimates (L-similarity estimate). Such estimates are computed using the following data-structures: The list L t,complete , an element L(P ) of which is defined as: L(P ) = m n M Q,m (c) t M P,n where Q is a query and P is a trajectory in the index. So the list stores for each trajectory the total time that its MBRs intersected with the query's MBRs. We record into this list only the intersections, where a query MBR is fully contained in all spatial dimensions by a trajectory MBR (or vice versa -it is equivalent. See figure 5, top right). The list L t,partial , an element L(P ) of which is defined as: L(P ) = m n M Q,m (p) t M P,n This list records for each sequence the total intersection in time for those query MBRs that are not fully contained within the x,y dimensions by the trajectory MBRs (or vice versa. Figure 5, top left). Regarding a query Q, for any trajectory P the sum of L t,complete (P ) + L t,partial (P ) will provide an upper bound on the similarity of P and Q. The reason for the distinction of the L-similarity estimate in two separate lists derives from the fact that the estimates stored in list L t,partial can significantly overestimate the LCSS similarity. If one wishes to relax the accuracy, in favor of enhanced performance, it is instructive to give a weight 0 &lt; w p &lt; 1 to all estimates in list L t,partial . Even though now we may miss the best match to our query, we are going to find a close match in less time. This weighted approach is used when we are seeking for approximate, but very good quality answers, however it will not be explained further due to space limitations. 2. Approximate estimates (V-similarity estimate). This second estimate is based on the intersecting volume of the MBRs. This type of estimates are stored in list L V : Any element L V (P ) of list L V records similarity estimates between trajectory P and query Q, based on the total volume intersection between the MBRs of P and Q. L(P ) = 1 length(P ) m n M Q,m V M P,n ||M Q,m || V ||M Q,m || t 220 where ||M|| V denotes the volume of MBR M and ||M|| t its length on the time axis. The L-similarity overestimates the LCSS , between two sequences A and B and so it can be deployed for the design of an index structure. Lemma 2. The use of the L-similarity estimate upper bounds the LCSS , similarity between two sequences A and B and therefore does not introduce any false dismissals. The V-similarity estimate can be used for approximate query answering. Even though it does not guarantee the absence of false dismissals, the results will be close to the optimal ones with high probability. Also, because this estimate provides a tighter approximation to the original distance , we expect faster response time. Indeed, as we show in the experimental section, the index performance is boosted, while the error in similarity is frequently less then 5%. 5.2 Estimates for the DTW When the distance function used is the Time Warping, using the index we obtain a lower bound of the actual distance . In this case we have the inverse situation from the LCSS; instead of calculating the degree of overlap between the MBRs of the indexed trajectories and the query, we evaluate the distance between the MBRs. The overall distance between the MBRs underestimates the true distance of the trajectories, and no false dismissals are introduced. Using the MBRs we can also calculate upper bound estimates on the distance, which hadn't been exploited in previous work [10, 22]. Sequences with lower bound larger than the smallest upper bound can be pruned. With this additional prefiltering step we can gain on average an additional 10-15% speedup in the total execution time. Due to space limitations only a visual representation of this approach is provided in figure 6. MBR GENERATION Given a multidimensional time-series (or an MBE) our objective is to minimize the volume of the sequence using k MBRs. Clearly, the best approximation of a trajectory (or an MBE) using a fixed number of MBRs is the set of MBRs that completely contain the sequence and minimize the volume consumption. We can show the following lemma: Lemma 3. Minimizing the volume of the Minimum Bounding Envelope, minimizes the expected similarity approximation error. Three different approaches are considered: 1. k-Optimal. We can discover the k MBRs of a sequence that take up the least volume, using a dynamic programming algorithm that requires O(n 2 k) time ([8]), where n is the length of the given sequence. Since this approach is not reasonable for large databases, we are motivated to consider approximate and faster solutions. 2. Equi-Split. This technique produces MBRs of fixed length l. It is a simple approach with cost linear in the length of a sequence. However, in pathological cases increasing the number of splits can result to larger space utilization ,therefore the choice of the MBR length becomes a critical parameter (see figure 7 for an example). A. Query Q B. Query Envelope C. Envelope Splitting D. Sequence MBRs E. MINDIST(Q,R) F. MAXDIST(Q,R) Figure 6: A visual intuition of the DTW indexing technique (the one-dimensional case is shown for clarity). The original query (A) is enclosed in a minimum-bounding envelope (B) like the LCSS approach . The MBE is split into its MBRs using equi or greedy split (fig. (C)). The candidate sequences in the database have their MBRs stored in the index (D). Between the query and any sequence in the index, the minimum and maximum distance can be quickly determined by examining the distance between the MBRs and the query's bounding envelope , as represented by the arrows in (E) and (F). 3. Greedy-Split. The Greedy approach is our implementation choice in this paper. Initially we assign an MBR to each of the n sequence points and at each subsequent step we merge the consecutive MBRs that will introduce the least volume consumption. The algorithm has a running time of O(nlogn). We can see a sketch of the method in fig. 8. Al-ternatively , instead of assigning the same number of splits to all objects, according to our space requirements we can assign a total of K splits to be distributed among all objects. This method can provide better results, since we can assign more splits for the objects that will yield more space gain. Also, this approach is more appropriate when one is dealing with sequences of different lengths. The complexity of this approach is O(K + N logN ), for a total of N objects ([8]). Input: A spatiotemporal trajectory T and an integer k denoting the number of final MBRs. For 0 i &lt; n compute the volume of the MBR produced by merging T i and T i+1 . The results are stored in a priority queue. While #M BRs &lt; k: Using the priority queue, merge the pair of consecutive MBRs that yield the smallest increase in volume. Delete the two merged MBRs and insert the new one in the priority queue. Output: A set of MBRs that cover T . Figure 8: The greedy algorithm for producing k MBRs that cover the trajectory T . After a trajectory is segmented the MBRs can be stored in a 3D-Rtree. Using the greedy split each additional split will always lead to smaller (or equal) volume (figure 7). A similar greedy split algorithm is also used for splitting the MBE of the query trajectory Q. 221 (a) Equi-Split, 8 MBRs, Gain = 5.992 (b) Equi-Split, 9 MBRs, Gain = 5.004 (c) Greedy-Split, 8MBRs, Gain = 9.157 (d) Greedy-Split, 9MBRs, Gain = 10.595 Figure 7: (a): 8 MBRs produced using equi-Split. The volume gain over having 1 MBR is 5.992. (b): Segmenting into 9 MBRs decreases the volume gain to 5.004. So, disk space is wasted without providing a better approximation of the trajectory. (c): 8 MBRs using greedy-Split. The volume gain over having 1 MBR is 9.157. (d): Every additional split will yield better space utilization. Segmentation into 9 MBRs increases volume gain to 10.595. SUPPORTING MULTIPLE MEASURES The application of the Minimum Bounding Envelope only on the query suggests that user queries are not confined to a predefined and rigid matching window . The user can pose queries of variable warping in time. In some datasets, there is no need to perform warping, since the Euclidean distance performs acceptably [11]. In other datasets, by using the Euclidean distance we can find quickly some very close matches, while using warping we can distinguish more flexible similarities. So, we can start by using a query with = 0 (no bounding envelope), and increase it progressively in order to find more flexible matches (figure 9). Therefore, our framework offers the unique advantage that multiple distance functions can be supported in a single index . The index sequences have been segmented without any envelope applied on them and never have to be adjusted again. For different measures, the aspects that change are, the creation of the query envelope and the type of operation between MBRs. In order to pose queries based on Euclidean distance we follow the steps: The query is segmented with no envelope applied on it. The minDist and maxDist estimators for the Euclidean distance are derived by calculating the distance between the query and index MBRs, just like in the DTW case. 0 50 100 150 200 250 300 350 200 150 100 50 0 50 100 Index Trajectory Query Figure 9: By incorporating the bounding envelope on the query, our approach can support Euclidean distance, constrained or full warping. This is accomplished by progressively expanding the MBE. EXPERIMENTAL EVALUATION In this section we compare the effectiveness of various splitting methods and we demonstrate the superiority of our lower bounding technique (for the DTW) compared to other proposed lower bounds. We describe the datasets we used and present comprehensive experiments regarding the index performance for the two similarity estimates. In addition, we evaluate the accuracy of the approximate estimates. All experiments conducted were run on an AMD Athlon 1.4 Ghz with 1GB RAM and 60GB of hard drive. 1. ASL 2. Buoy Sensor 3. Video Track 1 4. Flutter 5. Marine Mammals 6. Word Tracking 7. Random Walk 8. Video Track 2 Figure 10: Datasets used for testing the efficiency of various MBR generation methods. 8.1 MBR Generation Comparison The purpose of our first experiment is to test the space consumption of the presented MBR generation methods. We have used eight datasets with diverse characteristics, in order to provide objective results. We evaluate the space consumption, by calculating the "Average Volume Gain" (AvgV olGain), which is defined as the percentage of volume when using i MBRs, over the volume when using only 1 MBR, normalized by the maximum gain provided over all methods (for various number of splits). Random Equi Greedy 1 2 3 4 5 6 7 8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Dataset Average Volume Gain Figure 11: The greedy-split MBR generation algorithm presents the highest volume gain, by producing MBRs that consume consistently less space, over a number of datasets and for diverse number of generated MBRs 222 DATASET EQ s20,d5 GR s20,d5 EQ s40,d5 GR s40,d5 EQ s20,d5 GR s20,d5 EQ s40,d5 GR s40,d5 LB-Kim LB-Yi LCSS DTW ASL 0.732 0.799 0.825 0.856 0.449 0.632 0.588 0.756 0.1873 0.2530 VT1 0.260 0.339 0.453 0.511 0.087 0.136 0.230 0.266 0.0838 0.1692 Marine 0.719 0.750 0.804 0.814 0.226 0.506 0.308 0.608 0.2587 0.4251 Word 0.627 0.666 0.761 0.774 0.311 0.361 0.466 0.499 0.0316 0.2116 Random 0.596 0.652 0.701 0.741 0.322 0.384 0.440 0.491 0.1389 0.2067 VT2 0.341 0.431 0.498 0.569 0.210 0.296 0.363 0.437 0.2100 0.5321 Table 1: Some indicative results of how close our similarity estimates are to the exact value (for 20 and 40 splits, & = 5%). For all datasets the greedy-split approach provides the closest similarity estimates to the actual similarity. AvgV olGain is a number between 0 and 1, where higher numbers indicate increased volume gain (or less space consumption ) against the competitive methods. In figure 11 we observe the average volume gain for the eight datasets. The greedy-split algorithm produced MBRs that took at least half the space, compared to equi-split. The equi-split offers slightly better results, than producing MBRs at random positions. The volume gain of greedy-split was less, only for the buoy sensor, which is a very busy and unstruc-tured signal. This experiment validates that our choice to use the greedy-split method was correct. Since, the indexed MBR trajectories will take less space, we also expect tighter similarity estimates, therefore fewer false positives. 8.2 Tightness of Bounds In table 1 we show how close our similarity estimates are (for LCSS and DTW) to the actual similarity between sequences . Numbers closer to 1, indicate higher similarity to the value returned by the exact algorithm. To our best knowledge, this paper introduces the first upper bounding technique for the LCSS. For DTW there have been a few approaches to provide a lower bound of the distance; we refer to them as LB-Kim [12] and LB-Yi [21]. These lower bounds originally referred to 1D time-series; here we extend them in more dimensions, in order to provide unambiguous results about the tightness of our estimates. Note that the previously proposed methods operate on the raw data. Our approach can still provide tighter estimates, while operating only on the trajectory MBRs. Using the raw data our experiments indicate that we are consistently 2-3 times better than the best alternative approach. However, since our index operates on the segmented time-series we only report the results on the MBRs. The greedy-split method approximates the similarity consistently tighter than the equi-split. In table 1 only the results for = 5% of the query's length are reported, but similar results are observed for increasing values of . It is evident from the table that using our method we can provide very tight lower bounds of the actual distance. 8.3 Matching Quality We demonstrate the usefulness of our similarity measures in a real world dataset. The Library of Congress maintains thousands of handwritten manuscripts, and there is an increasing interest to perform automatic transcribing of these documents. Given the multiple variations of each word and due to the manuscript degradations, this is a particularly challenging task and the need for a flexible and robust distance function is essential. We have applied the LCSS and DTW measures on word Figure 12: Results for a real world application. 3NN reported for each query, using Dynamic Time Warping to match features extracted from scanned manuscript words. images extracted from a 10 page scanned manuscript. 4-dimensional time-series features have originally been extracted for each word. Here we maintain the 2 least correlated time-series features and treat each word as a trajectory. In figure 12 we observe the 3-KNN results using DT W for various word queries. The results are very good, showing high accuracy even for similarly looking words. Analogous results have been obtained using the LCSS. 8.4 Index performance We tested the performance of our index using the upper bound and the approximate similarity estimates, and compared it to the sequential scan. Because of limited space, the majority of the figures record the index performance using the LCSS as a similarity measure. The performance measure used is the total computation time required for the index and the sequential scan to return the nearest neighbor for the same one hundred queries. For the linear scan, one can also perform early termination of the LCSS (or the DTW) computation. Therefore, the LCSS execution can be stopped at the point where one is sure that the current sequence will not be more similar to the query than the bestSoFar. We call this optimistic linear scan. Pessimistic linear scan, is the one than does not reuse the previously computed similarity values and can be an accurate time estimate, when the query match resides at the end of the dataset. We demonstrate the index performance relative to 223 1024 2048 4096 8192 16384 32768 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Dataset size Time Ratio Compared to Linear Scan =5% Optimistic Pessimistic Linear Scan 1024 2048 4096 8192 16384 32768 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Dataset size Time Ratio Compared to Linear Scan =10% Optimistic Pessimistic Linear Scan 1024 2048 4096 8192 16384 32768 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Dataset size Time Ratio Compared to Linear Scan =20% Optimistic Pessimistic Linear Scan Figure 13: Index performance. For small warping windows the index can be up to 5 times faster than sequential scan without compromising accuracy. The gray regions indicate the range of potential speedup. 1024 2048 4096 8192 16384 32768 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Dataset size Time Ratio Compared to Linear Scan =5% Optimistic Pessimistic Linear Scan 1024 2048 4096 8192 16384 32768 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Dataset size Time Ratio Compared to Linear Scan =10% Optimistic Pessimistic Linear Scan 1024 2048 4096 8192 16384 32768 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Dataset size Time Ratio Compared to Linear Scan =20% Optimistic Pessimistic Linear Scan Figure 14: Using the approximate similarity estimates the response time can be more than 7 times faster. both types of linear scan, because this provides a realistic upper or lower bound on the index speedup. The dataset we used contained 2 10 . . . 2 16 trajectories. Taking under consideration that the average trajectory size is around 500 points, this resulted to a database with more than 16 million 2D points. The trajectories have been normalized by subtracting the average value in each direction of movement. All data and queries can be obtained by emailing the first author. Mixed two-dimensional Time-series (2D-Mixed). This second dataset consists of time-series of variable length, ranging from less than 100 points to over 1000 points. The dataset is comprised by the aggregation of the eight datasets we used for comparing the MBR generation methods. Since the total number of these trajectories is less than 200, we have used them as seeds to generate increasingly larger datasets. We create multiple copies of the original trajectories by incorporating the following features: Addition of small variations in the original trajectory pattern Addition of random compression and decompression in time The small variations in the pattern were added by interpolating peaks of Gaussian noise using splines. In this manner we are able to create the smooth variations that existed in the original datasets. 8.4.1 Results on the upper bound Estimates The index performance is influenced be three parameters: the size of the dataset, the warping length (as a percentage of the query's length) and the number of splits. For all experiments the parameter (matching in space) was set to std/2 of the query, which provided good and intuitive results. Dataset size: In figure 13 we can observe how the performance of the index scales with the database size (for various lengths of matching window). We record the index response time relative to both optimistic and pessimistic linear scan. Therefore, the gray region in the figures indicates the range of possible speedup. It is evident that the early termination feature of the sequential scan can significantly assist its performance. The usefulness of an index becomes obvious for large dataset sizes, where the quadratic computational cost dominates the I/O cost of the index. For these cases our approach can be up to 5 times faster than linear scan. In figure 15 we also demonstrate the pruning power of the index, as a true indicator (not biased by any implementation details) about the efficacy of our index. Using the index we perform 2-5 times fewer LCSS computations than the linear scan. We observe similar speedup when using the DTW as the distance function in figure 17. Parameter : The index performance is better for smaller warping lengths (parameter ). The experiments record the performance for warping from 5% to 20% of the query's length. Increasing values signify larger bounding envelopes around the query, therefore larger space of search and less accurate similarity estimates. The graphs suggest that an index cannot not be useful under full warping (when the data are normalized). Number of Splits: Although greater number of MBRs for each trajectory implies better volume utilization, nonetheless more MBRs also lead to increased I/O cost. When we are referring to x% splits, it means that we have assigned a total of 100/x( n i=1 ( ||T i ||)) splits, for all sequences T i . In our figures we provide the 5% splits scenario for the MBRs, which offers better performance than 10% and 20% splits, since for the last two cases the I/O cost negates the effect of the better query approximation. The index space requirements for 5% splits is less than a quarter of the dataset size. 8.4.2 Results on the approximate Estimates Here we present the index performance when the volume intersections of the MBRs are used as estimates of the sim-224 1024 2048 4096 8192 16384 32768 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 Dataset size Ratio of LCSS performed by the index Pruning Power compared to Linear Scan 5% splits 10% splits 20% splits Linear Scan =5% =10% =20% Figure 15: Each gray band indicates (for a certain warping window ) the percentage of LCSS computations conducted by the index compared to linear scan. 1024 2048 4096 8192 16384 32768 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Dataset size Average similarity Error Similarity Error, 5% splits =5% =10% =20% Figure 16: Using the V-similarity estimate, we can retrieve answers faster with very high accuracy. The LCSS similarity is very close (2-10%) to the exact answer returned by the sequential scan. 1024 2048 4096 8192 16384 32768 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Dataset size Time Ratio Compared to Linear Scan =5% Optimistic Pessimistic Linear Scan Figure 17: Index Performance using DTW as the distance measure. ( = 5%). We can observe up to 5 times speedup. ilarity and the results are shown in figure 14. We observe that using this approximate similarity estimate, our index performance is boosted up. The use of the V-similarity estimate leads to more tight approximations of the original similarity compared to the L-similarity estimate, however now we may miss finding the best match. Naturally, comes the question of the quality of the results . We capture this by calculating the absolute difference between the similarity of the best match returned by the index, and the best match found by the sequential scan for each query. Then we average the results over a number of queries |q|. Therefore, the Average Similarity Error (ASE) is: ASE = 1 |q| |q| i=1 ( |BestMatch index BestM atch exhaustive |) The results are shown in figure 16. We can see that the similarity returned by the V-similarity estimate is approxi-mately within 5% of the actual similarity (5% splits used). Therefore, by providing two similarity estimates the user can decide for the trade-off between the expedited execution time and the quality of results. Since by using the latter estimator we can significantly increase the performance of the index, this is the approach we recommend for mining large datasets. CONCLUSIONS AND FUTURE WORK In this paper we have presented an external memory indexing method for discovering similar multidimensional time-series . The unique advantage of our approach is that it can accommodate multiple distance measures. The method guarantees no false dismissals and depicts a significant execution speed up for the LCSS and DTW compared to sequential scan. We have shown the tightness of our similarity estimates and demonstrated the usefulness of our measures for challenging real world applications. We hope that our effort can act as a bridge between metric and non-metric functions, as well as a tool for understanding better their strengths and weaknesses. In the future we plan to investigate the combination of several heuristics, in order to provide even tighter estimates. Acknowledgements: We would like to thank Margrit Betke for providing us the Video Track I and II datasets. We also feel obliged to T. Rath and R. Manmatha for kindly providing the manuscript words dataset. REFERENCES [1] J. Aach and G. Church. Aligning gene expression time series with time warping algorithms. In Bioinformatics, Volume 17, pages 495508, 2001. [2] O. Arikan and D. Forsyth. Interactive motion generation from examples. In Proc. of ACM SIGGRAPH, 2002. [3] Z. Bar-Joseph, G. Gerber, D. Gifford, T. Jaakkola, and I. Simon. A new approach to analyzing gene expression time series data. In Proc. of 6th RECOMB, pages 3948, 2002. [4] D. Berndt and J. Clifford. Using Dynamic Time Warping to Find Patterns in Time Series. In Proc. of KDD Workshop, 1994. [5] M. Betke, J. Gips, and P. Fleming. The camera mouse: Visual tracking of body features to provide computer access for people with severe disabilities. In IEEE Transactions on Neural Systems and Rehabilitation Engineering, Vol. 10, No. 1, 2002. [6] G. Das, D. Gunopulos, and H. Mannila. Finding Similar Time Series. In Proc. of the First PKDD Symp., pages 88100, 1997. [7] D. Gavrila and L. Davis. Towards 3-d model-based tracking and recognition of human movement: a multi-view approach. In Int. Workshop on Face and Gesture Recognition. [8] M. Hadjieleftheriou, G. Kollios, V. Tsotras, and D. Gunopulos. Efficient indexing of spatiotemporal objects. In Proc. of 8th EDBT, 2002. [9] T. Kahveci, A. Singh, and A. Gurel. Similarity searching for multi-attribute sequences. In Proc. of SSDBM, 2002. [10] E. Keogh. Exact indexing of dynamic time warping. In Proc. of VLDB, 2002. [11] E. Keogh and S. Kasetty. On the need for time series data mining benchmarks: A survey and empirical demonstration. In Proc. of SIGKDD, 2002. [12] S. Kim, S. Park, and W. Chu. An index-based approach for similarity search supporting time warping in large sequence databases. In In Proc. of 17th ICDE, 2001. [13] Z. Kov acs-Vajna. A fingerprint verification system based on triangular matching and dynamic time warping. In IEEE Transactions on PAMI, Vol. 22, No. 11. [14] S.-L. Lee, S.-J. Chun, D.-H. Kim, J.-H. Lee, and C.-W. Chung. Similarity Search for Multidimensional Data Sequences. Proc. of ICDE, pages 599608, 2000. [15] S. Park, W. Chu, J. Yoon, and C. Hsu. Efficient Searches for Similar Subsequences of Different Lengths in Sequence Databases. In Proc. of ICDE, pages 2332, 2000. [16] T. Rath and R. Manmatha. Word image matching using dynamic time warping. In Tec Report MM-38. Center for Intelligent Information Retrieval, University of Massachusetts Amherst, 2002. [17] J. F. Roddick and K. Hornsby. Temporal, Spatial and Spatio-Temporal Data Mining. 2000. [18] M. Shimada and K. Uehara. Discovery of correlation from multi-stream of human motion. In Discovery Science 2000. [19] R. E. Valdes-Perez and C. A. Stone. Systematic detection of subtle spatio-temporal patterns in time-lapse imaging ii. particle migrations. In Bioimaging 6(2), pages 7178, 1998. [20] M. Vlachos, G. Kollios, and D. Gunopulos. Discovering similar multidimensional trajectories. In Proc. of ICDE, 2002. [21] B.-K. Yi, H. V. Jagadish, and C. Faloutsos. Efficient retrieval of similar time sequences under time warping. In Proc. of ICDE, pages 201208, 1998. [22] Y. Zhu and D. Shasha. Query by humming: a time series database approach. In Proc. of SIGMOD, 2003. 225

Dynamic Time Warping;indexing;trajectory;distance function;Dynamic Time Warping (DTW);similarity;Longest Common Subsequence;Trajectories;Longest Common Subsequence (LCSS);measure

111

Information Retrieval for Language Tutoring: An Overview of the REAP Project

INTRODUCTION Typical Web search engines are designed to run short queries against a huge collection of hyperlinked documents quickly and cheaply, and are often tuned for the types of queries people submit most often [2]. Many other types of applications exist for which large, open collections like the Web would be a valuable resource. However, these applications may require much more advanced support from information retrieval technology than is currently available. In particular, an application may have to describe more complex information needs, with a varied set of properties and data models, including aspects of the user's context and goals. In this paper we present an overview of one such application, the REAP project, whose main purpose is to provide reader-specific practice for improved reading comprehension. (REAP stands for REAder-specific Practice.) A key component of REAP is an advanced search model that can find documents satisfying a set of diverse and possibly complex lexical constraints, including a passage's topic, reading level (e.g. 3rd grade), use of syntax (simple vs. complex sentence structures), and vocabulary that is known or unknown to the student. Searching is performed on a database of documents automatically gathered from the Web which have been analyzed and annotated with a rich set of linguistic metadata. The Web is a potentially valuable resource for providing reading material of interest to the student because of its extent, variety, and currency for popular topics.

SYSTEM DESCRIPTION Here we describe the high-level design of the REAP information retrieval system, including document database requirements and construction, annotations, and a brief description of the retrieval model. 2.1 Database Construction Our goal is to present passages that are interesting to students, whether they are on current topics such as pop stars or sports events, or related to particular classroom projects. To this end, we use the Web as our source of reading practice materials because of its extent, variety, and currency of information. We want coverage of topics in the database to be deeper in areas that are more likely to be of interest to students. Coverage of other areas is intended to be broad, but more shallow. We therefore gather documents for the database using focused crawling [3]. The current prototype uses a page's reading difficulty to set priority for all links from that page equally, based on the distance from the target reading level range. We plan to explore more refined use of annotations to direct the crawl on a link-by-link basis. In our prototype, we collected 5 million pages based on an initial set of 20,000 seed pages acquired from the Google Kids Directory [7]. Our goal is to have at least 20 million pages that focus on material for grades 1 through 8. The document database must be large enough that the most important lexical constraints are satisfied by at least a small number of pages. Data annotation is currently performed off-line at indexing time. The specific annotations for REAP are described in Section 2.2. Once the documents are acquired, they are indexed using an extended version of the Lemur IR Toolkit [9]. We chose Lemur because of its support for language model-based retrieval, its extensibility, and its support for incremental indexing, which is important for efficient updates to the database. Annotations are currently stored as Lemur properties, but later versions will take advantage of the enhancements planned for support of rich document structure, described in Section 2.3. 2.2 Linguistic Annotations In addition to the underlying text, the following linguistic annotations are specified as features to be indexed: Basic text difficulty within a document section or region. This is calculated using a new method based on a mixture of language models [4] that is more reliable for Web pages and other non-traditional documents than typical reading difficulty measures. Grammatical structure. This includes part-of-speech tags for individual words as well as higher-level parse structures, up to sentence level. Document-level attributes such as title, metadata keywords, and ratings. 544 Topic category. This would involve broad categories such as fiction/non-fiction [5] or specific topics, perhaps based on Open Directory. Named entity tags. We use BBN's Identifinder [1] for high-precision tagging of proper names. We may also look at more advanced attributes such as text coherence and cohesion [6]. 2.3 Query and Retrieval Models A typical information need for the REAP system might be described as follows: Find a Web page about soccer, in American English, with reading difficulty around the Grade 3 level. The text should use both passive- and active-voice sentence constructions and should introduce about 10% new vocabulary relative to the student's known-vocabulary model. The page's topic is less important than finding pages that practice the words: for example, an article on another sport that satisfies the other constraints would also be acceptable. Information needs in REAP will be modeled as mixtures of multiple word histograms, representing different sources of evidence, as well as document-level or passage-level constraints on attributes such as reading difficulty. There is precedent for using word histograms to specify information needs: indeed, query expansion is one example of this. More specifically, related work includes language model-based techniques such as relevance models [8]. No current Web-based search engine is able to make use of combinations of lexical constraints and language models in this way, on such a large scale. To support this, we are making extensions to Lemur that include: 1. Retrieval models for rich document structure, which includes nested fields of different datatypes where each field may be associated with its own language model. 2. More detailed retrieval models in which we skew language models towards the appropriate grade level, topic, or style. 3. The use of user model descriptions as context for a query. 2.4 User Profiles In the current prototype, we model a reader's topic interests, reading level, and vocabulary acquisition goals using simple language models. For example, we model the curriculum as a word histogram. Although crude, this captures word-frequency information associated with general reading difficulty, as well as capturing topics that are the focus of the curriculum at each grade level. We plan to add more complex aspects to user profiles, including more specific lexical constraints such as grammar constructs and text novelty. The models can be updated incrementally as the student's interests evolve and they make progress through the curriculum. EVALUATION METHODS Evaluation of the end-to-end REAP system will be via a series of three year-long studies with both adults and children. The adult studies will provide feedback on vocabulary matching and comprehension, and the child studies will test the hypothesis that children will read adaptively to texts that vary in vocabulary demands, where those texts that closely reflect the reader's interests and comprehension can be used to support improved comprehension and vocabulary growth. CONCLUSION The REAP project is intended to advance the state of the art in information retrieval, as well as research in reading comprehension, by bringing together practical user models of student interests, vocabulary knowledge and growth, and other aspects of reading, with interesting material from large, open collections like the World Wide Web. This type of system is a valuable new research tool for educational psychologists and learning scientists, because it gives much greater control over how instructional materials are selected. This in turn allows testing of instructional hypotheses, such as the effect of 10% vocabulary stretch, which have been impractical to test in the past. The work also has direct application to other areas of language learning, such as English as a Second Language training. More broadly, however, we believe the REAP project is a important first step toward enabling richer user and task models than currently available with ad-hoc search systems. ACKNOWLEDGMENTS We thank our collaborators Maxine Eskenazi, Charles Perfetti and Jonathan Brown; John Cho and Alexandros Ntoulas of UCLA for their crawler code; and the anonymous reviewers. This work was supported by U.S. Dept. of Education grant R305G03123. Any opinions, findings, conclusions, or recommendations expressed in this material are the authors' and do not necessarily reflect those of the sponsors. REFERENCES [1] Bikel, D. M., Miller, S., Schwartz, R., Weischedel, R. M., Nymbol: A high-performance learning name-finder. In Proceedings of the 5th Conference on Applied Natural Language Processing, 194 - 201, 1997. [2] Broder, A. A taxonomy of Web search. In SIGIR Forum, 36(2). 3 - 10, 2002. [3] Chakrabarti, S., van der Berg, M., & Dom, B. Focused crawling: a new approach to topic-specific web resource discovery. In Proc. of the 8th International World-Wide Web Conference (WWW8), 1999. [4] Collins-Thompson, K., & Callan, J. A language modeling approach to predicting reading difficulty. Proceedings of HLT/NAACL 2004, Boston, USA, 2004. [5] Finn, A., Kushmerick, N. & Smyth, B. Fact or fiction: Content classification for digital libraries. Joint DELOS-NSF Workshop on Personalisation and Recommender Systems in Digital Libraries (Dublin), 2001. [6] Foltz, P. W., Kintsch, W., Landauer, T. K. Analysis of text coherence using Latent Semantic Analysis. Discourse Processes 25(2-3), 285 - 307, 1998. [7] Google Kids Directory. http://directory.google.com/Top/Kids_and_Teens/ [8] Lavrenko, V., and Croft, B. Relevance-based language models. In Proc. of the 24th Annual International ACM SIGIR Conference, New Orleans, 120 - 127, 2001. [9] Ogilvie, P. and Callan, J. Experiments using the Lemur Toolkit. In Proc.of the 10th Text Retrieval Conference, TREC 2001. NIST Special Publication 500-250, 103-108, 2001. 545

computer-assisted learning;user model;searching;reading comprehension;Information retrieval;information retrieval

112

Categorizing Web Queries According to Geographical Locality

Web pages (and resources, in general) can be characterized according to their geographical locality. For example, a web page with general information about wildflowers could be considered a global page, likely to be of interest to a ge-ographically broad audience. In contrast, a web page with listings on houses for sale in a specific city could be regarded as a local page, likely to be of interest only to an audience in a relatively narrow region. Similarly, some search engine queries (implicitly) target global pages, while other queries are after local pages. For example, the best results for query [wildflowers] are probably global pages about wildflowers such as the one discussed above. However, local pages that are relevant to, say, San Francisco are likely to be good matches for a query [houses for sale] that was issued by a San Francisco resident or by somebody moving to that city. Unfortunately, search engines do not analyze the geographical locality of queries and users, and hence often produce sub-optimal results. Thus query [wildflowers ] might return pages that discuss wildflowers in specific U.S. states (and not general information about wildflowers), while query [houses for sale] might return pages with real estate listings for locations other than that of interest to the person who issued the query. Deciding whether an unseen query should produce mostly local or global pages--without placing this burden on the search engine users--is an important and challenging problem, because queries are often ambiguous or underspecify the information they are after. In this paper, we address this problem by first defining how to categorize queries according to their (often implicit) geographical locality. We then introduce several alternatives for automatically and efficiently categorizing queries in our scheme, using a variety of state-of-the-art machine learning tools. We report a thorough evaluation of our classifiers using a large sample of queries from a real web search engine, and conclude by discussing how our query categorization approach can help improve query result quality.

INTRODUCTION Web pages (and resources, in general) can be characterized according to their geographical locality. For example, a web page with general information about wildflowers could be considered a global page, likely to be of interest to a ge-ographically broad audience. In contrast, a web page with listings on houses for sale in a specific city could be regarded as a local page, likely to be of interest only to an audience in a relatively narrow region. Earlier research [9] has addressed the problem of automatically computing the "geographical scope" of web resources. Often search engine queries (implicitly) target global web pages, while other queries are after local pages. For example, the best results for query [wildflowers] are probably global pages about wildflowers discussing what types of climates wildflowers grow in, where wildflowers can be purchased, or what types of wildflower species exist. In contrast, local pages that are relevant to, say, San Francisco are likely to be good matches for a query [houses for sale] that was issued by a San Francisco resident, or by somebody moving to San Francisco, even if "San Francisco" is not mentioned in the query. The user's intent when submitting a query may not be always easy to determine, but if underspecified queries such as [houses for sale] can be detected, they can be subsequently modified by adding the most likely target geographical location or by getting further user input to cus-tomize the results. Unfortunately, search engines do not analyze the geographical locality of queries and users, and hence often produce sub-optimal results, even if these results are on-topic and reasonably "popular" or "authoritative." Thus query [wildflowers] might return pages that discuss wildflowers in specific U.S. states (and not general information about wildflowers ). In fact, as of the writing of this paper, the first 10 results that Google provides for this query include 5 pages each of which discusses wildflowers in only one U.S. 325 state (e.g., "Texas Wildflowers"). Similarly, the top 10 results that Google returns for query [houses for sale] include real estate pages for Tuscany, United Kingdom, and New Zealand. These pages are likely to be irrelevant to, say, somebody interested in San Francisco real estate who types such an underspecified query. Deciding whether a query posed by a regular search engine user should produce mostly local or global pages is an important and challenging problem, because queries are often ambiguous or underspecify the information they are after, as in the examples above. By identifying that, say, query [wildflowers] is likely after "global" information, a search engine could rank the results for this query so that state-specific pages do not appear among the top matches. By identifying that, say, query [houses for sale] is likely after "local" information, a search engine could filter out pages whose geographical locality is not appropriate for the user who issued the query. Note that deciding which location is of interest to a user who wrote an underspecified query such as [houses for sale] is an orthogonal, important issue that we do not address in this paper. Our focus is on identifying that such a query is after "local" pages in nature, and should therefore be treated differently by a search engine than queries that are after "global" pages. By knowing that a user query is after local information, a search engine might choose to privilege pages whose geographical locality coincides with that of the user's or, alternatively, attempt to obtain further input from the user on what location is of interest. In this paper, we first define how to categorize user queries according to their (often implicit) geographical locality. We then introduce several alternatives for automatically and efficiently classifying queries according to their locality, using a variety of state-of-the-art machine learning tools. We report a thorough evaluation of our classifiers using a large sample of queries from a real web search engine query log. Finally, we discuss how our query categorization approach can help improve query result quality. The specific contributions of this paper are as follows: A discussion on how to categorize user queries according to their geographical locality, based on a careful analysis of a large query log from the Excite web site (Section 3). A feature representation for queries; we derive the feature representation of a query from the results produced for the query by a web search engine such as Google (Section 4.1). A variety of automatic query classification strategies that use our feature representation for queries (Section 4.2). A large-scale experimental evaluation of our strategies over real search engine queries (Section 5). Preliminary query reformulation and page re-ranking strategies that exploit our query classification techniques to improve query result quality (Section 6). RELATED WORK Traditional information-retrieval research has studied how to best answer keyword-based queries over collections of text documents [18]. These collections are typically assumed to be relatively uniform in terms of, say, their quality and scope. With the advent of the web, researchers are studying other "dimensions" to the data that help separate useful resources from less-useful ones in an extremely heterogeneous environment like the web. Notably, the Google search engine [4] and the HITS algorithm [7, 13] estimate the "impor-tance" of web pages by analyzing the hyperlinks that point to them, thus capturing an additional dimension to the web data, namely how important or authoritative the pages are. Ding et al. [9] extract yet another crucial dimension of the web data, namely the geographical scope of web pages. For example, the Stanford Daily newspaper has a geographical scope that consists of the city of Palo Alto (where Stanford University is located), while the New York Times newspaper has a geographical scope that includes the entire U.S. To compute the geographical scope of a web page, Ding et al. propose two complementary strategies: a technique based on the geographical distribution of HTML links to the page, and a technique based on the distribution of geographical references in the text of the page. Ding et al. report on a search-engine prototype that simply filters out from the results for a user query any pages not in the geographical scope of the user. This technique does not attempt to determine whether a query is best answered with "global" or "local" pages, which is the focus of our paper. Ding et al. built on the work by Buyukkokten et al. [6], who discussed how to map a web site (e.g., http://www-db.stanford.edu) to a geographical location (e.g., Palo Alto) and presented a tool to display the geographical origin of the HTML links to a given web page. This tool then helps visualize the geographical scope of web pages [6]. A few commercial web sites manually classify web resources by their location, or keep directory information that lists where each company or web site is located. The North-ernLight search engine 1 extracts addresses from web pages, letting users narrow their searches to specific geographical regions (e.g., to pages "originated" within a five-mile radius of a given zip code). Users benefit from this information because they can further filter their query results. McCurley [14] presented a variety of approaches for recognizing geographical references on web pages, together with a nav-igational tool to browse pages by geographical proximity and their spatial context. (Please refer to [16] for additional references.) None of these techniques addresses our focus problem in this paper: automatically determining the geographical locality associated with a given, unmodified search engine query. DEFINING GEOGRAPHICAL LOCALITY As discussed above, queries posed to a web search engine can be regarded as local, if their best matches are likely to be "local" pages, or as global, if their best matches are likely to be "global" pages. In an attempt to make this distinction more concrete, we now discuss several examples of local and global queries. Global queries often do not include a location name, as is the case for query [Perl scripting]. A user issuing this query is probably after tutorials about the Perl language, and hence pages on the topic with a restricted geographi-1 http://www.northernlight.com/ 326 cal scope are less desirable than global pages. Other global queries do not mention a location explicitly either, but are topically associated with one particular location. An example of such a query is [Elgin marbles], which is topically associated with the city of Athens. We consider these queries as global, since their best matches are broad, global pages, not localized pages with a limited geographical scope. In-terestingly , global queries sometimes do include a location name. For example, a query might be just a location name (e.g., [Galapagos Islands]) or a request for concrete information about a location (e.g., [Boston area codes]). General resources about the location (e.g., tourist guides) are arguably to be preferred for such queries, which are hence regarded as global. Other global queries include locations that are strongly associated topic-wise with the rest of the query. Query [Woody Allen NYC] is an example of such a query. The location mentioned in this query (i.e., "NYC," for "New York City") is not used to restrict query results to pages of interest to New York residents, but rather expresses a topic specification. Query [Ansel Adams Yosemite] is another example: photographer Ansel Adams took a famous series of photographs in Yosemite. Local queries often include a location name, as is the case for query [Wisconsin Christmas tree producers association]. The location mentioned in this query (i.e., "Wisconsin") is used to "localize" the query results. Query [houses for sale New York City] is a related example. Other local queries do not include a location name, but still implicitly request "localized" results. Query [houses for sale] is an example of such a query. These queries tend to be underspecified, but are still asked by (presumably naive) search engine users. We conducted a thorough examination of a large number (over 1,200) of real search engine queries. Most queries that we encountered can be cleanly categorized as being either global or local. However, other queries are inherently ambiguous, and their correct category is impossible to determine without further information on the user intent behind them. For example, query [New York pizza] could be con-strued as a local query if it is, say, after pizza delivery web sites for the New York area. In contrast, the same query could be regarded as a global query if the user who issues it wants to learn about the characteristics of New York-style pizza. USING CLASSIFIERS TO DETERMINE LOCALITY We earlier established that queries are associated with local or global status, which influences the kind of results that are desirable. Since current search engines do not directly take into account geographical information, for certain types of queries they produce a large number of on-topic but un-wanted results, as in the [houses for sale] example discussed earlier. In this section, we discuss automatic methods that can determine, given a query, whether the query is a local or global one. To build the two-class classifier, we experimented with several state-of-the-art classification techniques , using widely available implementations for each. We describe below the features used in the classification, how we extract them from web pages, and the classifiers with which we experimented. 4.1 Classification Features Web queries, which we treat in this paper as ordered bags of words with no other structure, are typically fairly short. In the collection of 2,477,283 real queries that we used in our experiments (Section 5.1), 84.9% were five words long or shorter. Because few words are available per query, basing the classification directly on the words in the query may lead to severe sparse data problems. Even more importantly, some of the characteristics that make a query local or global are not directly observable in the query itself, but rather in the results returned. For example, a query that returns results that contain few references to geographical locations is likely to be global, while a query that returns results spread uniformly over many locations without including a significant percentage of results with no locations is likely to be local. For these reasons, we base our classification on a sample of results actually returned for a given query rather than the words in the query itself. By observing distributional characteristics in the unmodified results, the classifier can infer the type of the query (global or local) so that the results can be appropriately filtered or re-ordered, or the query modified . In a way the approach is similar in spirit to query expansion techniques that rely on pseudo-relevance feedback [5]. In our experiments, we use Google (via the Google API 2 ) to obtain the top 50 web pages that match the query. For simplicity, we limited our search to HTML pages, skipping over non-HTML documents. We chose Google because it represents state-of-the-art web search technology and offers a published interface for submitting large numbers of queries. We represent the web pages returned by Google as text documents. This conversion is achieved by using the lynx HTML browser with the -dump option. We base our classification features on measures of frequency and dispersion of location names in these text files. For this purpose, we have constructed a database of 1,605 location names by concatenating lists of all country names 3 , of the capitals of these countries 4 , of the fifty U.S. states, and of all cities in the United States with more than 25,000 people 5 . We then compare the words in each text document with the database of location names, and output any matching entries and their count per document. This matching is case insensitive, because we found capitalization of location names in web pages to be erratic. Note that we do not attempt to disambiguate words that match location names but also have other senses (e.g., "China"), as this is a hard problem in natural language analysis; instead, we count such words as locations. An alternative approach that would detect and disambiguate location names would be to use a named-entity tagger. We experimented with a well-known third-party named-entity tagger, but we encountered a very high error rate because of the noise often introduced in web pages. Our classification features combine these location counts in various ways. For each query, we measure the average 2 http://www.google.com/apis 3 Obtained from the United Nations, http://www.un.org/ Overview/unmember.html. 4 Obtained from the CIA World Factbook, http://www. capitals.com/. 5 Obtained from the U.S. Census Bureau (2000 census figures), http://www.census.gov/prod/2002pubs/00ccdb/ cc00_tabC1.pdf. 327 (per returned web page) number of location words in the retrieved results. We count the average frequency of location words at different levels of detail (country, state, city), as well as the average of the aggregate total for all locations. We obtain these frequencies for both the total count (tokens) and the unique location words in each page (types), as it is possible that a few locations would be repeated many times across the results, indicating a global query, or that many locations would be repeated few times each, indicating a local query. We also consider the total number of unique locations across all the returned documents taken together, divided by the number of retrieved documents. For the average token frequencies of city, state, and country locations we also calculate the minimum and maximum across the set of returned web pages. To account for the hierarchical nature of location information, we calculate an alternative frequency for states where we include in the count for each state the counts for all cities from that state that were found in that text; this allows us to group together location information for cities in the same state. We also include some distributional measures , namely the fraction of the pages that include at least one location of any kind, the percentage of words that match locations across all pages, and the standard deviation of the total per page location count. Finally, we add to our list of features the total number of words in all of the returned documents, to explore any effect the local/global distinction may have on the size of the returned documents. These calculations provide for 20 distinct features that are passed on to the classifier. 6 The core data needed to produce these 20 query features (i.e., the locations mentioned in each web page) could be efficiently computed by a search engine such as Google at page-indexing time. Then, the final feature computation could be quickly performed at query time using this core data. 4.2 Classification Methods We initially trained a classifier using Ripper [8], which constructs a rule-based classifier in an incremental manner. The algorithm creates an initial set of very specific rules based on the data, similar to the way in which decision trees are generated. The rules are then pruned iteratively to eliminate the ones that do not seem to be valid for a large enough subset of the training data, so as to prevent overfitting. Although Ripper provides a robust classifier with high accuracy and transparency (a human can easily examine the produced rules), it outputs binary "local""global" decisions . In many cases, it is preferable to obtain a measure of confidence in the result or an estimate of the probability that the assigned class is the correct one. To add this capability to our classifier, we experimented with logistic regression [19]. Logistic, or log-linear, regression models a binary output variable (the class) as a function of a weighted sum of several input variables (the classification features). Con-ceptually , a linear predictor is first fitted over the training data in a manner identical to regular regression analysis, i.e., = w 0 + k i=1 w i F i where F i is the i-th feature and w i is the weight assigned to 6 Studying the effect on classification accuracy of a richer feature set (e.g., including as well all words on the result pages) is the subject of interesting future work. that feature during training. Subsequently, is transformed to the final response, C, via the logistic transformation C = e 1 + e which guarantees that C is between 0 and 1. Each of the endpoints of the interval (0 , 1) is associated with one of the classes, and C gives the probability that the correct class is the one associated with "1". In practice, the calculations are not performed as a separate regression and transformation, but rather as a series of successive regressions of transformed variables via the iterative reweighted least squares algorithm [1]. 7 We used the implementation of log-linear regression provided in the R statistical package. 8 Another desideratum for our classifier is its ability to support different costs for the two possible kinds of errors (misclassifying local queries versus misclassifying global queries). Which kind of error is the most important may vary for different settings; for our search modification application, we consider the misclassification of global queries as local ones a more serious error. This is because during our subsequent modification of the returned results (Section 6), we reorder the results for some of the queries that we consider global, but we modify the original queries for some of the queries classified as local, returning potentially very different results. Consequently, the results can change more significantly for a query classified as local, and the potential for error is higher when a global query is labeled local than the other way around. Both Ripper and log-linear regression can incorporate different costs for each type of error. We experimented with a third classification approach that also supports this feature , Support Vector Machines (SVMs) [2], which have been found quite effective for text matching problems [11]. SVM classifiers conceptually convert the original measurements of the features in the data to points in a high-dimensional space that facilitates the separation between the two classes more than the original representation. While the transformation between the original and the high-dimensional space may be complex, it needs not to be carried out explicitly. Instead, it is sufficient to calculate a kernel function that only involves dot products between the transformed data points, and can be calculated directly in the original feature space. We report experiments with two of the most common kernel functions: a linear kernel, K(x, y) = x y + 1 and a Gaussian (radial basis function) kernel, K(x, y) = e - x-y 2 /2 2 where is a parameter (representing the standard deviation of the underlying distribution). This latter kernel has been recommended for text matching tasks [10]. Regardless of the choice of kernel, determining the optimal classifier is equivalent to determining the hyperplane that maximizes the total distance between itself and representative transformed data points (the support vectors). Finding the optimal classifier therefore becomes a constrained quadratic optimization 7 This is because the modeled distribution is binomial rather than normal, and hence the variance depends on the mean-see [19] for the technical details. 8 http://www.r-project.org/ 328 Set Original number of queries Number of appropriate queries Global Local Training 595 439 368 (83.8%) 71 (16.2%) Development 199 148 125 (84.5%) 23 (15.5%) Test 497 379 334 (88.1%) 45 (11.9%) Table 1: Distribution of global and local queries in our training, development, and test sets. problem. In our experiments, we use the SVM-Light implementation 9 [12]. In many binary classification tasks, one of the two classes predominates, and thus trained classifiers tend to favor that class in the absence of strong evidence to the contrary. This certainly applies to our task; as we show in Section 5.1, 8389% of web queries are global. Weiss and Provost [21] showed that this imbalance can lead to inferior classifier performance on the test set, and that the problem can be addressed through oversampling of the rarer class in the training data. Their method examines different oversampling rates by constructing artificial training sets where the smaller class is randomly oversampled to achieve a specific ratio between samples from the two classes. For each such sampling ratio, a classifier is trained, which assigns a score to each object indicating strength of evidence for one of the classes. By fixing a specific strength threshold, we divide the classifier output into the two classes. Further, by varying this threshold 10 we can obtain an error-rate curve for each class as a function of the threshold. The entire process results in a Receiver-Operator Characteristic (ROC) curve [3] for each sampling ratio. Specific points on the curve that optimize the desired combination of error rates can then be selected, and the performance of the classification method across the different thresholds can be measured from the area between the curve and the x-axis. Weiss and Provost use the C4.5 classifier [17], a decision tree classifier with additional pruning of nodes to avoid overfitting. We use a software package provided by them (and consequently also the C4.5 algorithm) to explore the effect that different ratios of global to local queries during training have on classifier performance. EXPERIMENTAL RESULTS We now describe the data (Section 5.1) and metrics (Section 5.2) that we use for the experimental evaluation of the query classifiers (Section 5.3). 5.1 Data For the experiments reported in this paper, we used a sample of real queries submitted to the Excite search engine. 11 We had access to a portion of the December 1999 query log of Excite, containing 2,477,283 queries. We randomly selected initial sets of queries for training, development (tuning the parameters of the classifiers we train), and testing purposes by selecting each of these queries for inclusion in each set with a constant (very small) probability. These probabilities were set to 400/2,477,283, 400/2,477,283, and 500/2,477,283 9 Available from http://svmlight.joachims.org/. 10 Setting the threshold to each extreme assigns all or none of the data points to that category. 11 http://www.excite.com/ for the three sets, respectively. Subsequently we combined the training and development set, and reassigned the queries in the combined set so that three-fourths were placed in the training set and one-fourth in the development; we kept the test set separate. This process generated 595, 199, and 497 queries in the initial versions of the training, development, and test sets. We further eliminated queries that passed any of the following tests: Upon examination, they appeared likely to produce results with explicit sexual content. When supplied to Google--and after filtering out any non-HTML results and any broken links--the queries produced fewer than 40 files. This constraint is meant to ensure that we are not including in our experimental data queries that contain misspellings or deal with extremely esoteric subjects, for which not enough material for determining locality would be available. They had already been included in an earlier set (we constructed first the training set, then the development set, and finally the test set). Since multiple people may issue the same query, duplicates can be found in the log. Although our algorithms take no special advantage of duplicates, we eliminated them to avoid any bias. Taking into account variations of upper/lower case and spacing between queries (but not word order ), this constraint removed 6 queries from the test set. These filtering steps left us with 439 queries in the training set, 148 queries in the development set, and 379 queries in the test set. We then classified the queries using the criteria laid out in Section 3. Table 1 shows the size of the three sets before and after filtering, and the distribution of local and global queries in each set. We observe that, in general, most queries (8389%) tend to be global. 5.2 Evaluation Metrics We consider a number of evaluation metrics to rate the performance of the various classifiers and their configurations . Since a large majority of the queries are global (85.6% in the training, development, and test sets combined), overall classification accuracy (i.e., the percentage of correct classification decisions) may not be the most appropriate measure. This is because a baseline method that always suggests the most populous class ("global") will have an accuracy equal to the proportion of global queries in the evaluated set. Yet such a classifier will offer no improvement during search since it provides no new information. The situation is analogous to applications in information retrieval or medicine where very few of the samples should be labeled positive (e.g., in a test for a disease that affects only 0.1% 329 of patients). While we do not want overall accuracy to decrease from the baseline (at least not significantly), we will utilize measures that capture the classifier's improved ability to detect the rarer class relative to the baseline method. Two standard such metrics are precision and recall for the local queries. Precision is the ratio of the number of items correctly assigned to the class divided by the total number of items assigned to the class. Recall is the ratio of the number of items correctly assigned to a class as compared with the total number of items in the class. Note that the baseline method achieves precision of 100% but recall of 0%. For a given classifier with adjustable parameters, often precision can be increased at the expense of recall, and vice versa; therefore we also compute the F-measure [20] (with equal weights) to combine precision and recall into a single number, F-measure = 2 Precision Recall Precision + Recall Finally, we argued earlier that one kind of misclassification errors may be assigned a higher cost. We can then calculate the average cost [15], Average cost = X{ Global , Local } Cost( X) Rate(X) where Cost( X) is the cost of wrong X classifications and Rate( X) is the rate of wrong X classifications. Average cost is the measure to minimize from a decision theory perspective . The rate of wrong classifications for a class is equal to the number of data points that have been misclassified into that class divided by the total number of classification decisions, and the costs for each misclassification error are predetermined parameters. If both costs are set to 1, then the average cost becomes equal to the total error rate, i.e., one minus accuracy. In our experiments, we report the average cost considering the mislabeling of global queries as local twice as important as the mislabeling of local queries, for the reasons explained in the previous section. 5.3 Results We trained the classifiers of Section 4.2 on the 439 queries in our training set. Ripper and the regression model were trained on that training set without modification. For C4.5 and SVMs, we explored the effect that different proportions of local queries in the training set have on overall performance . For that purpose, we used our development set to evaluate the performance effects of different local query proportions , and select the optimal classifier within each family. For the C4.5-based classifier, we used the C4.4 software provided by Foster Provost and Claudia Perlich to explore the effect of different proportions of local and global queries. We created training sets by randomly oversampling or un-dersampling the minority (local) class as needed, in increments of 10%. For any given proportion of local queries between 10% and 90%, we started from our training set, modified it according to the above sampling method to have the desired proportion of local queries, trained the corresponding C4.5 classifier, and evaluated its performance on our development set. The natural proportion of the local class in the unmodified training data is also included as one of the proportions used to build and evaluate a classifier. In this manner, we obtain curves for the various metrics as the proportion of local queries varies (Figure 1). We observe Figure 1: Evaluation metrics for C4.5 classifiers trained on different proportions of local queries. that the highest value for precision and F-measure, and the lowest value for the average cost, are obtained when the classifier is trained with a significantly amplified proportion of local queries (80%). Further, running C4.5 with 80% local queries also produced the largest area under the ROC curve obtained when different precision/recall tradeoffs in the development set are explored. On the basis of this information , we selected the proportion of 80% local queries as the optimal configuration for C4.5. We refer to that configuration as C4.5(80), and this is the version of C4.5 that we evaluated on the test set. Using our own implementation for constructing extended training sets with a given proportion of local queries, we performed similar experiments for Support Vector Machines with linear and Gaussian kernels. For these classifiers, we also experimented with versions trained with equal error costs for the two kinds of classification errors, and with versions where, during training, a false local assignment counts for twice as much as a false global assignment. We found that the optimal proportion of local queries is closer to the natural proportion with SVMs compared to C4.5 classifiers; the proportions chosen from our development set were 50% for the linear SVM classifier with equal error costs, 30% for the linear SVM classifier with unequal error costs, 30% for the Gaussian SVM classifier with equal error costs, and 20% for the Gaussian SVM classifier with unequal error costs. We denote the optimal classifiers from these four families as SVM-Linear-E(50), SVM-Linear-U(30), SVM-Gaussian-E (30), and SVM-Gaussian-U(20), respectively. Figure 2 shows the curve obtained for the SVM-Gaussian-U family of classifiers. Having determined the best value for the proportion of local queries for C4.5 and SVM-based classifiers, we evaluate these classifiers, as well as the classifiers obtained from Ripper and log-linear regression, on our test set. 12 Table 2 shows the values of the evaluation metrics obtained on the unseen test set. The classifier using a linear kernel SVM with unequal error costs achieves the highest F-measure, 12 We also experimented with variable error costs for the Ripper classifier, using the same 2:1 error cost correspondence, but the resulting classifier was identical to the Ripper classifier obtained with equal error costs. 330 Classifier Recall Precision F-Measure Average Cost Accuracy Ripper 53.33% 47.06% 50.00% 0.1979 87.34% Log-linear Regression 37.78% 58.62% 45.95% 0.1372 89.45% C4.5(80) 40.00% 32.73% 36.00% 0.2665 83.11% SVM-Linear-E(50) 48.89% 48.89% 48.89% 0.1821 87.86% SVM-Linear-U(30) 48.89% 53.66% 51.16% 0.1609 88.92% SVM-Gaussian-E(30) 37.78% 53.13% 44.16% 0.1530 88.65% SVM-Gaussian-U(20) 37.78% 53.13% 44.16% 0.1530 88.65% Baseline (always global) 0.00% 100.00% 0.00% 0.1187 88.13% Table 2: Evaluation metrics on the test set of selected classifiers optimized over the development set. Figure 2: Evaluation metrics for Support Vector Machines with Gaussian kernel and false local assignments weighted twice as much as false global assignments , trained on different proportions of local queries. while the log-linear classifier achieves the lowest average classification cost. As expected, the SVM classifiers that were trained with unequal error costs achieve the same or lower average cost (which also utilizes the same unequal error costs) compared to their counterparts trained with equal error costs. Overall, Ripper, log-linear regression, and the two SVM classifiers with linear kernels achieve the highest performance, with small differences between them. They are followed by the two SVM classifiers with a Gaussian kernel function, while C4.5 trails significantly behind the other classifiers. The features used for classification vary considerably from classifier to classifier. 13 Ripper achieves one of the best classification performances using only one simple rule, based only on the average number of city locations per returned web page: if that number exceeds a threshold, the query is classified as local, otherwise as global. On the other hand, the C4.5 and SVM classifiers utilize all or almost all the features. The log-linear regression classifier falls in-between these two extremes, and primarily utilizes the average numbers of unique city, state, and country names per retrieved page, as well as the total number of unique locations per page (4 features). For concreteness, and to conclude our discussion, Table 3 13 Most classifiers automatically ignore some of the provided features, to avoid overfitting. shows the performance of our classifiers on a few representative examples of local and global queries. IMPROVING SEARCH RESULTS The core of this paper is on classifying queries as either local or global. In this section, we present preliminary ideas on how to exploit this classification to improve the quality of the query results. Further exploration of these and other directions is the subject of interesting future work. Consider a query that has been classified as local using the techniques of Section 4. By definition, this query is best answered with "localized" pages. We can easily determine if the query includes any location name by using the dictionary-based approach of Section 4.1. If no locations are present in the query (e.g., as in query [houses for sale]), in the absence of further information we can attempt to "localize" the query results to the geographical area of the user issuing the query, for which we can rely on registration information provided by the user, for example. Consequently, we can simply expand the query by appending the user's location to it, to turn, say, the query [houses for sale] into [houses for sale San Francisco] for a San Francisco resident. Alternatively, a search engine might attempt to obtain additional information from the user to further localize the query as appropriate. For example, the query [houses for sale] can then be transformed into [houses for sale New York City] for a San Francisco resident who is moving to New York City. In either case, the expanded query will tend to produce much more focused and localized results than the original query does. As of the writing of this paper, all of the top-10 results returned by Google for query [houses for sale San Francisco] are results of relevance to a person interested in Bay Area real estate. In contrast, most of the results for the original query, [houses for sale], are irrelevant to such a person, as discussed in the Introduction. An alternative, more expensive strategy for handling these queries is to compute and exploit the geographical scope of web pages as defined in [9]. Then, pages with a geographical scope that includes the location of the user issuing the query would be preferred over other pages. In contrast, a local query in which locations are mentioned is likely to return pages with the right locality, making any further modification of the query or reranking of the results unnecessary. Consider now a query that has been classified as global using the techniques of Section 4. By definition, this query is best answered with "broad" pages. Rather than attempting to modify a global query so that it returns "broad" pages, we can follow a result reranking strategy to privilege these pages over more localized ones. One possible reranking strategy is to reorder the results from, say, Google for 331 Class Query Classifier Ripper Regression C4.5(80) SVM-LE SVM-LU SVM-GE SVM-GU Global [Perl scripting] Global -0.9381 Global -1.9163 -1.7882 -1.0627 -1.0590 [world news] Global -0.8306 Local -0.5183 -0.3114 -0.4166 -0.1440 [wildflowers] Global -0.5421 Global -0.7267 -0.8082 -0.8931 -0.8144 [Elgin marbles] Local 0.4690 Local 0.6426 0.6654 0.0378 0.1016 [Galapagos Islands] Global -0.7941 Global -1.2834 -1.1998 -0.9826 -0.8575 [Boston zip code] Local -0.0243 Local 0.6874 0.6152 0.0408 0.0797 [Woody Allen NYC] Global -0.2226 Global -0.3253 -0.3541 -0.6182 -0.5272 Local [houses for sale] Global -0.6759 Global -1.0769 -1.0962 -0.9242 -0.8516 [Volkswagen clubs] Local -0.0933 Global 1.0844 0.7917 0.0562 0.0837 [Wisconsin Christmas tree producers association] Global 0.1927 Local -0.1667 -0.4421 -0.4461 -0.3582 [New York style pizza delivery] Global -0.0938 Global -0.5945 -0.6809 -0.5857 -0.4824 Table 3: Classification assignments made by different classifiers on several example queries. SVM-LE, SVM-LU , SVM-GE, and SVM-GU stand for classifiers SVM-Linear-E(50), SVM-Linear-U(30), SVM-Gaussian-E (30), and SVM-Gaussian-U(20), respectively. For regression and SVM classifiers, positive numbers indicate assignment to the local class, and negative numbers indicate assignment to the global class; the absolute magnitude of the numbers increases as the classifier's confidence in its decision increases. (We linearly transformed the regression output from the (0, 1) to the ( -1, 1) range.) The scale of the numbers is consistent across queries and between all SVM classifiers, but not directly comparable between regression classifiers (bound between -1 and 1) and SVM classifiers (unbounded). the unmodified query based on the geographical scope of the pages as defined in [9]. Thus pages with a broad geographical scope (e.g., covering the entire United States) would prevail over other pages with a narrower scope. A less expensive alternative is to classify the result pages as local or global following a procedure similar to that of Section 4 for queries. Specifically, we implemented this alternative by training C4.5 Rules, a rule-based version of the C4.5 decision-tree classifier, with a collection of 140 web pages categorized in the Yahoo! directory. Pages classified under individual states in the "Regional" portion of the directory were regarded as local, while pages under general categories were regarded as global. The feature representation for the pages was analogous to that for the queries in Section 4.1 but restricted to features that are meaningful over individual pages (e.g., total number of locations on a page), as opposed to over a collection of pages (e.g., minimum number of locations per page in the top-50 result pages for a query). At query time, we reorder the results so as to privilege global pages over local ones. This is based on the locality classification of the pages, which can be precomputed off-line since it is query-independent or performed on the fly as we do in our prototype implementation. This procedure is efficient, and produced promising initial results for a handful of global queries (e.g., [wildflowers]) that we tried. Our preliminary approach to query modification is therefore as follows: Given a query specified by the user, we supply first the unmodified query to the search engine and collect the top 50 results. We extract location names from these results 14 , and calculate the features of Section 4.1. Using one of the best performing classifiers of Section 4, we determine if the query is global or local. If it is local and 14 As noted earlier, these names could be cached along with each web page at the time of indexing, to increase efficiency. contains at least one location name, nothing is done--the results returned from the unmodified query are presented to the user. If the query is local and contains no location, we add the user's location (or, alternatively, request further information from the user, as discussed), reissue the query and present the results. Finally, if the query is global, we calculate the scope of each retrieved web page using part of the location features computed earlier and the C4.5 Rules classifier, and rerank the results so that more global pages are higher in the list shown to the user. We have built a prototype implementation of this algorithm, using the classifier obtained from Ripper (because of the relative simplicity of its rules) for query classification, and Google as the search engine. CONCLUSION We have described an attribute of queries, locality, that-to the best of our knowledge--has not been explored before either in theoretical work or in practical search engines but can significantly affect the appropriateness of the results returned to the user. We defined a categorization scheme for queries based on their geographical locality, and showed how queries can be represented for purposes of determining locality by features based on location names found in the results they produce. Using these features, automatic classifiers for determining locality can be built. We explored several state-of-the-art classification approaches, and evaluated their performance on a large set of actual queries. The empirical results indicated that for many queries locality can be determined effectively. The bulk of the paper discussed methods for classifying queries according to locality, and empirically established that this is desirable and feasible for many queries. We also presented some first thoughts on possible query refor-332 mulation and result reranking strategies that utilize locality information to actually improve the results the user sees. Although our strategies for query modification and result reranking are preliminary, they illustrate a promising family of approaches that we plan to investigate in the future so that we can exploit the classification of queries based on their geographical locality in order to improve search result quality. Acknowledgments This material is based upon work supported in part by the National Science Foundation under Grants No. IIS-97-33880 and IIS-98-17434. We are grateful to Claudia Perlich and Foster Provost for providing us with their adaptation of the C4.5 classifier that we used in our experiments. Also, we would like to thank Thorsten Joachims for answering our questions on SVM-Light, and David Parkes for his helpful comments and insight. REFERENCES [1] D. M. Bates and D. G. Watts. Nonlinear Regression Analysis and its Applications. Wiley, New York, 1988. [2] B. E. Boser, I. M. Guyon, and V. Vapnik. A training algorithm for optimal margin classifiers. In Proceedings of the Fifth Annual Workshop on Computational Learning Theory, Pittsburgh, 1992. [3] A. Bradley. The use of the area under the ROC curve in the evaluation of machine learning algorithms. Pattern Recognition, 30(7):11451159, 1998. [4] S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. In Proceedings of the Seventh International World Wide Web Conference (WWW7), Apr. 1998. [5] C. Buckley, J. Allan, G. Salton, and A. Singhal. Automatic query expansion using SMART: TREC 3. In Proceedings of the Third Text REtrieval Conference (TREC-3), pages 6980, April 1995. NIST Special Publication 500-225. [6] O. Buyukkokten, J. Cho, H. Garcia-Molina, L. Gravano, and N. Shivakumar. Exploiting geographical location information of web pages. In Proceedings of the ACM SIGMOD Workshop on the Web and Databases (WebDB'99), June 1999. [7] S. Chakrabarti, B. Dom, P. Raghavan, S. Rajagopalan, D. Gibson, and J. Kleinberg. Automatic resource compilation by analyzing hyperlink structure and associated text. In Proceedings of the Seventh International World Wide Web Conference (WWW7), Apr. 1998. [8] W. W. Cohen. Learning trees and rules with set-valued functions. In Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence, 1996. [9] J. Ding, L. Gravano, and N. Shivakumar. Computing geographical scopes of web resources. In Proceedings of the Twenty-sixth International Conference on Very Large Databases (VLDB'00), 2000. [10] G. W. Flake, E. J. Glover, S. Lawrence, and C. L. Giles. Extracting query modifications from nonlinear SVMs. In Proceedings of the Eleventh International World-Wide Web Conference, Dec. 2002. [11] M. A. Hearst. Trends and controversies: Support vector machines. IEEE Intelligent Systems, 13(4):1828, July 1998. [12] T. Joachims. Estimating the generalization of performance of an SVM efficiently. In Proceedings of the Fourteenth International Conference on Machine Learning, 2000. [13] J. Kleinberg. Authoritative sources in a hyperlinked environment. In Proceedings of the Ninth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 668677, Jan. 1998. [14] K. S. McCurley. Geospatial mapping and navigation of the web. In Proceedings of the Tenth International World Wide Web Conference (WWW10), May 2001. [15] M. Pazzani, C. Merz, P. Murphy, K. Ali, T. Hume, and C. Brunk. Reducing misclassification costs. In Proceedings of the Eleventh International Conference on Machine Learning, Sept. 1997. [16] R. Purves, A. Ruas, M. Sanderson, M. Sester, M. van Kreveld, and R. Weibel. Spatial information retrieval and geographical ontologies: An overview of the SPIRIT project. In Proceedings of the 25th ACM International Conference on Research and Development in Information Retrieval (SIGIR'02), 2002. [17] R. J. Quinlan. C4.5: Programs for Machine Learning. Morgan Kaufman, 1993. [18] G. Salton. Automatic Text Processing: The transformation, analysis, and retrieval of information by computer. Addison-Wesley, 1989. [19] T. J. Santner and D. E. Duffy. The Statistical Analysis of Discrete Data. Springer-Verlag, New York, 1989. [20] C. J. van Rijsbergen. Information Retrieval. Butterworths, London, 2nd edition, 1979. [21] G. M. Weiss and F. Provost. The effect of class distribution on classifier learning: An empirical study. Technical Report ML-TR-44, Computer Science Department, Rutgers University, Aug. 2001. 333

geographical locality;categorization scheme;query modification;web search;query categorization / query classification;web queries;search engines;global page;local page;information retrieval;search engine;query classification

113

Information Revelation and Privacy in Online Social Networks

Participation in social networking sites has dramatically increased in recent years. Services such as Friendster, Tribe, or the Facebook allow millions of individuals to create online profiles and share personal information with vast networks of friends - and, often, unknown numbers of strangers. In this paper we study patterns of information revelation in online social networks and their privacy implications. We analyze the online behavior of more than 4,000 Carnegie Mellon University students who have joined a popular social networking site catered to colleges. We evaluate the amount of information they disclose and study their usage of the site's privacy settings. We highlight potential attacks on various aspects of their privacy, and we show that only a minimal percentage of users changes the highly permeable privacy preferences.

EVOLUTION OF ONLINE NETWORKING In recent years online social networking has moved from niche phenomenon to mass adoption. Although the concept dates back to the 1960s (with University of Illinois Plato computer-based education tool, see [16]), viral growth and Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. WPES'05, November 7, 2005, Alexandria, Virginia, USA. Copyright 2005 ACM 1-59593-228-3/05/0011 ... $ 5.00. commercial interest only arose well after the advent of the Internet. 1 The rapid increase in participation in very recent years has been accompanied by a progressive diversification and sophistication of purposes and usage patterns across a multitude of different sites. The Social Software Weblog 2 now groups hundreds of social networking sites in nine categories , including business, common interests, dating, face-to -face facilitation, friends, pets, and photos. While boundaries are blurred, most online networking sites share a core of features: through the site an individual offers a "profile" - a representation of their sel[ves] (and, often, of their own social networks) - to others to peruse, with the intention of contacting or being contacted by others , to meet new friends or dates (Friendster, 3 Orkut 4 ), find new jobs (LinkedIn 5 ), receive or provide recommendations (Tribe 6 ), and much more. It is not unusual for successful social networking sites to experience periods of viral growth with participation expanding at rates topping 20% a month. Liu and Maes estimate in [18] that "well over a million self-descriptive personal profiles are available across different web-based social networks" in the United States, and Leonard, already in 2004, reported in [16] that world-wide "[s]even million people have accounts on Friendster. [...] Two million are registered to MySpace. A whopping 16 million are supposed to have registered on Tickle for a chance to take a personality test." The success of these sites has attracted the attention of the media (e.g., [23], [3], [16], [4], [26]) and researchers. The latter have often built upon the existing literature on social network theory (e.g., [20], [21], [11], [12], [32]) to discuss its online incarnations. In particular, [7] discusses issues of trust and intimacy in online networking; [9] and [8] focus on participants' strategic representation of their selves to others; and [18] focus on harvesting online social network profiles to obtain a distributed recommender system. In this paper, we focus on patterns of personal information revelation and privacy implications associated with online networking. Not only are the participation rates to online 1 One of the first networking sites, SixDegrees.com, was launched in 1997 but shut down in 2000 after "struggling to find a purpose for [its] concept" [5]. 2 Http://www.socialsoftware.weblogsinc.com/ . 3 Http://www.friendster.com/ . 4 Http://www.orkut.com/ . 5 Http://www.linkedin.com/ . 6 Http://www.tribe.net/ . 71 social networking staggering among certain demographics; so, also, are the amount and type of information participants freely reveal. Category-based representations of a person's broad interests are a recurrent feature across most networking sites [18]. Such categories may include indications of a person's literary or entertainment interests, as well as political and sexual ones. In addition, personally identified or identifiable data (as well as contact information) are often provided, together with intimate portraits of a person's social or inner life. Such apparent openness to reveal personal information to vast networks of loosely defined acquaintances and complete strangers calls for attention. We investigate information revelation behavior in online networking using actual field data about the usage and the inferred privacy preferences of more than 4,000 users of a site catered to college students, the Facebook. 7 Our results provide a preliminary but detailed picture of personal information revelation and privacy concerns (or lack thereof) in the wild, rather than as discerned through surveys and laboratory experiments. The remainder of this paper is organized as follows. We first elaborate on information revelation issues in online social networking in Section 2. Next, we present the results of our data gathering in Section 3. Then, we discuss their implications in terms of users attitudes and privacy risks in Section 4. Finally, we summarize our findings and conclude in Section 5. INFORMATION REVELATION AND ONLINE SOCIAL NETWORKING While social networking sites share the basic purpose of online interaction and communication, specific goals and patterns of usage vary significantly across different services. The most common model is based on the presentation of the participant's profile and the visualization of her network of relations to others - such is the case of Friendster. This model can stretch towards different directions. In match-making sites, like Match.com 8 or Nerve 9 and Salon 10 Personals , the profile is critical and the network of relations is absent. In diary/online journal sites like LiveJournal, 11 profiles become secondary, networks may or may not be visible , while participants' online journal entries take a central role. Online social networking thus can morph into online classified in one direction and blogging in another. Patterns of personal information revelation are, therefore, quite variable. First, the pretense of identifiability changes across different types of sites. The use of real names to (re)present an account profile to the rest of the online community may be encouraged (through technical specifications, registration requirements, or social norms) in college websites like the Facebook, that aspire to connect participants' profiles to their public identities. The use of real names may be toler-ated but filtered in dating/connecting sites like Friendster, that create a thin shield of weak pseudonymity between the public identity of a person and her online persona by making only the first name of a participant visible to others, 7 Http://www.facebook.com/ . 8 Http://www.match.com/ . 9 Http://personals.nerve.com/ . 10 Http://personals.salon.com/ . 11 Http://www.livejournal.com/ . and not her last name. Or, the use of real names and personal contact information could be openly discouraged, as in pseudonymous-based dating websites like Match.com, that attempt to protect the public identity of a person by making its linkage to the online persona more difficult. However, notwithstanding the different approaches to identifiability, most sites encourage the publication of personal and identifiable personal photos (such as clear shots of a person's face). Second, the type of information revealed or elicited often orbits around hobbies and interests, but can stride from there in different directions. These include: semi-public information such as current and previous schools and employers (as in Friendster); private information such as drinking and drug habits and sexual preferences and orientation (as in Nerve Personals); and open-ended entries (as in LiveJournal ). Third, visibility of information is highly variable. In certain sites (especially the ostensibly pseudonymous ones) any member may view any other member's profile. On weaker-pseudonym sites, access to personal information may be limited to participants that are part of the direct or extended network of the profile owner. Such visibility tuning controls become even more refined on sites which make no pretense of pseudonymity, like the Facebook. And yet, across different sites, anecdotal evidence suggests that participants are happy to disclose as much information as possible to as many people as possible. It is not unusual to find profiles on sites like Friendster or Salon Personals that list their owners' personal email addresses (or link to their personal websites), in violation of the recommendation or requirements of the hosting service itself. In the next sub-section , we resort to the theory of social networks to frame the analysis of such behavior, which we then investigate em-pirically in Section 3. 2.1 Social Network Theory and Privacy The relation between privacy and a person's social network is multi-faceted. In certain occasions we want information about ourselves to be known only by a small circle of close friends, and not by strangers. In other instances, we are willing to reveal personal information to anonymous strangers, but not to those who know us better. Social network theorists have discussed the relevance of relations of different depth and strength in a person's social network (see [11], [12]) and the importance of so-called weak ties in the flow of information across different nodes in a network. Network theory has also been used to explore how distant nodes can get interconnected through relatively few random ties (e.g., [20], [21], [32]). The privacy relevance of these arguments has recently been highlighted by Strahilevitz in [27]. Strahilevitz has proposed applying formal social network theory as a tool for aiding interpretation of privacy in legal cases. He suggests basing conclusions regarding privacy "on what the parties should have expected to follow the initial disclosure of information by someone other than the defen-dant" (op cit, p. 57). In other words, the consideration of how information is expected to flow from node to node in somebody's social network should also inform that person's expectations for privacy of information revealed in the network. However, the application of social network theory to the 72 study of information revelation (and, implicitly, privacy choices) in online social networks highlights significant differences between the offline and the online scenarios. First, offline social networks are made of ties that can only be loosely categorized as weak or strong ties, but in reality are extremely diverse in terms of how close and intimate a subject perceives a relation to be. Online social networks, on the other side, often reduce these nuanced connections to simplistic binary relations: "Friend or not" [8]. Observing online social networks, Danah Boyd notes that "there is no way to determine what metric was used or what the role or weight of the relationship is. While some people are willing to indicate anyone as Friends, and others stick to a conservative definition, most users tend to list anyone who they know and do not actively dislike. This often means that people are indicated as Friends even though the user does not particularly know or trust the person" [8] (p. 2). Second, while the number of strong ties that a person may maintain on a social networking site may not be significantly increased by online networking technology, Donath and Boyd note that "the number of weak ties one can form and maintain may be able to increase substantially, because the type of communication that can be done more cheaply and easily with new technology is well suited for these ties" [9] (p. 80). Third, while an offline social network may include up to a dozen of intimate or significant ties and 1000 to 1700 "ac-quaintances" or "interactions" (see [9] and [27]), an online social networks can list hundreds of direct "friends" and include hundreds of thousands of additional friends within just three degrees of separation from a subject. This implies online social networks are both vaster and have more weaker ties, on average, than offline social networks . In other words, thousands of users may be classified as friends of friends of an individual and become able to access her personal information, while, at the same time, the threshold to qualify as friend on somebody's network is low. This may make the online social network only an imaginary (or, to borrow Anderson's terminology, an imagined ) community (see [2]). Hence, trust in and within online social networks may be assigned differently and have a different meaning than in their offline counterparts. Online social networks are also more levelled, in that the same information is provided to larger amounts of friends connected to the subject through ties of different strength. And here lies a paradox. While privacy may be considered conducive to and necessary for intimacy (for [10], intimacy resides in selectively revealing private information to certain individuals , but not to others), trust may decrease within an online social network. At the same time, a new form of intimacy becomes widespread: the sharing of personal information with large and potential unknown numbers of friends and strangers altogether. The ability to meaningfully interact with others is mildly augmented, while the ability of others to access the person is significantly enlarged. It remains to be investigated how similar or different are the mental models people apply to personal information revelation within a traditional network of friends compared to those that are applied in an online network. 2.2 Privacy Implications Privacy implications associated with online social networking depend on the level of identifiability of the information provided, its possible recipients, and its possible uses. Even social networking websites that do not openly expose their users' identities may provide enough information to identify the profile's owner. This may happen, for example, through face re-identification [13]. Liu and Maes estimate in [18] a 15% overlap in 2 of the major social networking sites they studied. Since users often re-use the same or similar photos across different sites, an identified face can be used to identify a pseudonym profile with the same or similar face on another site. Similar re-identifications are possible through demographic data, but also through category-based representations of interests that reveal unique or rare overlaps of hobbies or tastes. We note that information revelation can work in two ways: by allowing another party to identify a pseudonymous profile through previous knowledge of a sub-ject's characteristics or traits; or by allowing another party to infer previously unknown characteristics or traits about a subject identified on a certain site. We present evaluations of the probabilities of success of these attacks on users of a specific networking site in Section 4. To whom may identifiable information be made available? First of all, of course, the hosting site, that may use and extend the information (both knowingly and unknowingly revealed by the participant) in different ways (below we discuss extracts from the privacy policy of a social networking site that are relevant to this discussion). Obviously, the information is available within the network itself, whose extension in time (that is, data durability) and space (that is, membership extension) may not be fully known or knowable by the participant. Finally, the easiness of joining and extending one's network, and the lack of basic security measures (such as SSL logins) at most networking sites make it easy for third parties (from hackers to government agencies) to access participants data without the site's direct collaboration (already in 2003, LiveJournal used to receive at least five reports of ID hijacking per day, [23]). How can that information be used? It depends on the information actually provided - which may, in certain cases, be very extensive and intimate. Risks range from identity theft to online and physical stalking; from embarrassment to price discrimination and blackmailing. Yet, there are some who believe that social networking sites can also offer the solution to online privacy problems. In an interview, Tribe.net CEO Mark Pincus noted that "[s]ocial networking has the potential to create an intelligent order in the current chaos by letting you manage how public you make yourself and why and who can contact you." [4]. We test this position in Section 4. While privacy may be at risk in social networking sites, information is willingly provided. Different factors are likely to drive information revelation in online social networks. The list includes signalling (as discussed in [9]), because the perceived benefit of selectively revealing data to strangers may appear larger than the perceived costs of possible privacy invasions; peer pressure and herding behavior; relaxed attitudes towards (or lack of interest in) personal privacy; incomplete information (about the possible privacy implications of information revelation); faith in the networking service or trust in its members; myopic evaluation of privacy risks (see [1]); or also the service's own user interface, that may drive the unchallenged acceptance of permeable default privacy settings. We do not attempt to ascertain the relative impact of 73 different drivers in this paper. However, in the following sections we present data on actual behavioral patterns of information revelation and inferred privacy attitudes in a college-targeted networking site. This investigation offers a starting point for subsequent analysis of the motivations behind observed behaviors. THE FACEBOOK.COM Many users of social networking sites are of college age [8], and recent ventures have started explicitly catering to the college crowd and, in some cases, to specific colleges (e.g., the Facebook.com, but also Universitysingles.ca, quad5.com, CampusNetwork.com, iVentster.com, and others). College-oriented social networking sites provide opportunities to combine online and face-to-face interactions within an ostensibly bounded domain. This makes them different from traditional networking sites: they are communities based "on a shared real space" [26]. This combination may explain the explosive growth of some of these services (according to [26], the Facebook has spread "to 573 campuses and 2.4 million users. [...] [I]t typically attracts 80 percent of a school's undergraduate population as well as a smattering of graduate students, faculty members, and recent alumni.") Also because of this, college-oriented networks offer a wealth of personal data of potentially great value to external observers (as reported by [6], for example, the Pentagon manages a database of 16-to-25-year-old US youth data, containing around 30 million records, and continuously merged with other data for focused marketing). Since many of these sites require a college's email account for a participant to be admitted to the online social network of that college, expectations of validity of certain personal information provided by others on the network may increase. Together with the apparent sharing of a physical environment with other members of the network, that expectation may increase the sense of trust and intimacy across the online community. And yet, since these services can be easily accessed by outsiders (see Section 4) and since members can hardly control the expansion of their own network (often, a member's network increases also through the activity of other members), such communities turn out to be more imagined than real, and privacy expectations may not be matched by privacy reality. The characteristics mentioned above make college-oriented networking sites intriguing candidates for our study of information revelation and privacy preferences. In the rest of this paper we analyze data gathered from the network of Carnegie Mellon University (CMU) students enlisted on one of such sites, the Facebook. The Facebook has gained huge adoption within the CMU student community but is present with similar success at many other colleges nationwide. It validates CMU-specific network accounts by requiring the use of CMU email addresses for registration and login. Its interface grants participants very granular control on the searchability and visibility of their personal information (by friend or location, by type of user, and by type of data). The default settings, however, are set to make the participants profile searchable by anybody else in any school in the Facebook network, and make its actual content visible to any other user at the same college or at another college in the same physical location. 12 12 At the time of writing, the geography feature which gen-17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 0 2 4 6 8 10 12 14 16 18 20 Age Percentage of Profiles Male Female Figure 1: Age distribution of Facebook profiles at CMU. The majority of users (95.6%) falls into the 18-24 age bracket. The Facebook is straightforward about the usage it plans for the participants' personal information: at the time of this writing, its privacy policy [30] reports that the site will collect additional information about its users (for instance, from instant messaging), not originated from the use of the service itself. The policy also reports that participants' information may include information that the participant has not knowingly provided (for example, her IP address), and that personal data may be shared with third parties. 3.1 Access Tools In June 2005, we separately searched for all "female" and all "male" profiles for CMU Facebook members using the website's advanced search feature and extracted their profile IDs. Using these IDs we then downloaded a total of 4540 profiles - virtually the entire CMU Facebook population at the time of the study. 3.2 Demographics The majority of users of the Facebook at CMU are undergraduate students (3345 or 73.7% of all profiles; see Table 1). This corresponds to 62.1% of the total undergraduate population at CMU [31]. Graduate students, staff and faculty are represented to a much lesser extent (6.3%, 1.3%, and 1.5% of the CMU population, respectively). The majority of users is male (60.4% vs. 39.2%). Table 2 shows the gender distribution for the different user categories. The strong dominance of undergraduate users is also reflected in the user age distribution shown in Figure 1. The vast majority of users (95.6%) falls in the 18-24 age bracket. Overall the average age is 21.04 years. erates networks based on physical location is by default not available to undergraduate students. However, the status of a profile can easily be changed to e.g. "graduate student" for which the feature is accessible. 74 Table 1: Distribution of CMU Facebook profiles for different user categories. The majority of users are undergraduate students. The table lists the percentage of the CMU population (for each category) that are users of the Facebook (if available). # Profiles % of Facebook Profiles % of CMU Population Undergraduate Students 3345 74.6 62.1 Alumni 853 18.8 Graduate Students 270 5.9 6.3 Staff 35 0.8 1.3 Faculty 17 0.4 1.5 Table 2: Gender distribution for different user categories. # Profiles % of Category % of CMU Population Male 2742 60.4 Overall Female 1781 39.2 Male 2025 60.5 62.0 Undergraduate Students Female 1320 39.5 62.3 Male 484 56.7 Alumni Female 369 43.3 Male 191 70.7 6.3 Graduate Students Female 79 29.3 6.3 Male 23 65.7 Staff Female 12 34.3 Male 17 100 3.4 Faculty Female 0 0.0 0.0 3.3 Types and Amount of Information Disclosed The Facebook offers users the ability to disclose a large and varied amount of personal information. We evaluated to which extent users at CMU provide personal information. Figure 2 shows the percentages of CMU profiles that disclose different categories of information. In general, CMU users of the Facebook provide an astonishing amount of information: 90.8% of profiles contain an image, 87.8% of users reveal their birth date, 39.9% list a phone number (including 28.8% of profiles that contain a cellphone number), and 50.8% list their current residence. The majority of users also disclose their dating preferences (male or female), current relationship status (single, married , or in a relationship), political views (from "very liberal" to "very conservative"), and various interests (including music , books, and movies). A large percentage of users (62.9%) that list a relationship status other than single even identify their partner by name and/or link to their Facebook profile. Note that, as further discussed below in Section 3.4, Facebook profiles tend to be fully identified with each participant's real first and last names, both of which are used as the profile's name. In other words, whoever views a profile is also able to connect the real first and last name of a person to the personal information provided - that may include birthday or current residence. Across most categories, the amount of information revealed by female and male users is very similar. A notable exception is the phone number, disclosed by substantially more male than female users (47.1% vs. 28.9%). Single male users tend to report their phone numbers in even higher frequencies , thereby possibly signalling their elevated interest 0 10 20 30 40 50 60 70 80 90 100 Summer Job Favorite Movies Favorite Books Favorite Music Interests Political Preference Relationship Partner Relationship Status Dating Interests Highschool AIM Screenname Phone Address Home Town Birthday Profile Image Percentage of Profiles Figure 2: Percentages of CMU profiles revealing various types of personal information. in making a maximum amount of contact information easily available. Additional types of information disclosed by Facebook users (such as the membership of one's own network of friends at the home college or elsewhere, last login information , class schedule, and others) are discussed in the rest of this paper. 3.4 Data Validity and Data Identifiability The terms of service of the site encourage users to only publish profiles that directly relate to them and not to other 75 entities, people or fictional characters. In addition, in order to sign up with the Facebook a valid email address of one of the more than 500 academic institutions that the site covers has to be provided. This requirement, along with the site's mission of organizing the real life social networks of their members, provides incentives for users to only publish accurate information. We tested how valid the published data appears to be. In addition, we studied how identifiable or granular the provided data is. In general, determining the accuracy of the information provided by users on the Facebook (or any other social networking website) is nontrivial for all but selected individual cases. We therefore restrict our validity evaluation to the measurement of the manually determined perceived accuracy of information on a randomly selected subset of 100 profiles. 3.4.1 Profile Names We manually categorized the names given on Facebook profiles as being one of the the following: 1. Real Name Name appears to be real. 2. Partial Name Only a first name is given. 3. Fake Name Obviously fake name. Table 3 shows the results of the evaluation. We found 89% of all names to be realistic and likely the true names for the users (for example, can be matched to the visible CMU email address provided as login), with only 8% of names obviously fake. The percentage of people that choose to only disclose their first name was very small: 3%. Table 3: Categorization of name quality of a random subset of 100 profile names from the Facebook. The vast majority of names appear to be real names with only a very small percentage of partial or obviously fake names. Category Percentage Facebook Profiles Real Name 89% Partial Name 3% Fake Name 8% In other words, the vast majority of Facebook users seem to provide their fully identifiable names, although they are not forced to do so by the site itself. As comparison, 98.5% of the profiles that include a birthday actually report the fully identified birth date (day, month, and year), although, again, users are not forced to provide the complete information (the remaining 1.5% of users reported only the month or the month and day but not the year of birth). Assessing the validity of birth dates is not trivial. However, in certain instances we observed friends posting birthday wishes in the comments section of the profile of a user on the day that had been reported by the user as her birthday. In addition, the incentives to provide a fake birth date (rather than not providing one at all, which is permitted by the system) would be unclear. 3.4.2 Identifiability of Images on Profile The vast majority of profiles contain an image (90.8%, see Section 3.3). While there is no explicit requirement to provide a facial image, the majority of users do so. In order to assess the quality of the images provided we manually labelled them into one of four categories: 1. Identifiable Image quality is good enough to enable person recognition . 2. Semi-Identifiable The profile image shows a person, but due to the image composition or face pose the person is not directly recognizable. Other aspects however (e.g. hair color, body shape, etc.) are visible. 3. Group Image The image contains more than one face and no other profile information (e.g. gender) can be used to identify the user in the image. 4. Joke Image Images clearly not related to a person (e.g. cartoon or celebrity image). Table 4 shows the results of labelling the profile images into the four categories. In the majority of profiles the images are suitable for direct identification (61%). Overall, 80% of images contain at least some information useful for identification . Only a small subset of 12% of all images are clearly not related to the profile user. We repeated the same evaluation using 100 randomly chosen images from Friendster, where the profile name is only the first name of the member (which makes Friendster profiles not as identifiable as Facebook ones). Here the percentage of "joke images" is much higher (23%) and the percentage of images suitable for direct identification lower (55%). 13 3.4.3 Friends Networks The Facebook helps in organizing a real-life social network online. Since Facebook users interact with many of the other users directly in real-life, often on a daily basis, the network of friends may function as profile fact checker, potentially triggering questions about obviously erroneous information. Facebook users typically maintain a very large network of friends. On average, CMU Facebook users list 78.2 friends at CMU and 54.9 friends at other schools. 76.6% of users have 25 or more CMU friends, whereas 68.6% of profiles show 25 or more non-CMU friends. See Figure 3 for histogram plots of the distribution of sizes of the networks for friends at CMU and elsewhere. This represents some effort, since adding a friend requires explicit confirmation. 3.5 Data Visibility and Privacy Preferences For any user of the Facebook, other users fall into four different categories: friends, friends of friends, non-friend users 13 We note that Friendster's profiles used to be populated by numerous fake and/or humorous profiles, also called "Fakesters" (see [8]). Friendster management tried to eliminate fake profiles and succeeded in significantly reducing their number, but not completely extirpating them from the network. Based on our manual calculations, the share of fake Friendster profiles is currently comparable to the share of fake Facebook profiles reported above. 76 Table 4: Categorization of user identifiability based on manual evaluation of a randomly selected subset of 100 images from both Facebook and Friendster profiles. Images provided on Facebook profiles are in the majority of cases suitable for direct identification (61%). The percentage of images obviously unrelated to a person ("joke image") is much lower for Facebook images in comparison to images on Friendster profiles (12% vs. 23%). Category Percentage Facebook Profiles Percentage Friendster Profiles Identifiable 61% 55% Semi-Identifiable 19% 15% Group Image 8% 6% Joke Image 12% 23% 0 20 40 60 80 100 120 140 160 0 0.05 0.1 Percentage of Profiles Number of CMU Friends Listed 0 20 40 60 80 100 120 140 160 0 0.05 0.1 Percentage of Profiles Number of Non-CMU Friends Listed (a) Network of CMU friends (b) Network of Non-CMU friends Figure 3: Histogram of the size of networks for both CMU friends (a) and non-CMU friends (b). Users maintain large networks of friends with the average user having 78.2 friends at CMU and 54.9 friends elsewhere. at the same institution and non-friend users at a different institution. 14 By default, everyone on the Facebook appears in searches of everyone else, independent of the searchers institutional affiliation. In search results the users' full names (partial searches for e.g. first names are possible) appear along with the profile image, the academic institution that the user is attending, and the users' status there. The Facebook reinforces this default settings by labelling it "recom-mended" on the privacy preference page. Also by default the full profile (including contact information) is visible to everyone else at the same institution. Prior research in HCI has shown that users tend to not change default settings [19]. This makes the choice of default settings by website operators very important. On the other hand, the site provides users a very granular and relatively sophisticated interface to control the searchability and visibility of their profiles. Undergrad users, for example, can make their profiles searchable only to other undergrad users, or only users who are friends, or users who are friends of friends, or users at the same institution - or combinations of the above constraints. In addition, visibility of the entire profile can be similarly controlled. Granular control on contact information is also provided. Sociological theories of privacy have noted how an individual may selectively disclose personal information to others in order to establish different degrees of trust and intimacy with them (see [10]). In light of these theories, we tested 14 The Facebook recently introduced a new relationship category based on user location, e.g. Pittsburgh, which we did not consider in this study. how much CMU Facebook users take advantage of the ability the site provides to manage their presentation of sel[ves]. By creating accounts at different institutions, and by using accounts with varying degree of interconnectedness with the rest of the CMU network, we were able to infer how individual users within the CMU network were selecting their own privacy preference. 3.5.1 Profile Searchability We first measured the percentage of users that changed the search default setting away from being searchable to everyone on the Facebook to only being searchable to CMU users. We generated a list of profile IDs currently in use at CMU and compared it with a list of profile IDs visible from a different academic institution. We found that only 1.2% of users (18 female, 45 male) made use of this privacy setting. 3.5.2 Profile Visibility We then evaluated the number of CMU users that changed profile visibility by restricting access to CMU users. We used the list of profile IDs currently in use at CMU and evaluated which percentage of profiles were fully accessible to an unconnected user (not friend or friend of friend of any profile). Only 3 profiles (0.06%) in total did not fall into this category. 3.5.3 Facebook Data Access We can conclude that only a vanishingly small number of users change the (permissive) default privacy preferences. In general, fully identifiable information such as personal 77 image and first and last name is available to anybody registered at any Facebook member network. Since the Facebook boasts a 80% average participation rate among undergraduate students at the hundreds of US institutions it covers, and since around 61% of our CMU subset provides identifiable face images, it is relatively easy for anybody to gain access to these data, and cheap to store a nation-wide database of fully identified students and their IDs. In other words, information suitable for creating a brief digital dossier consisting of name, college affiliation, status and a profile image can be accessed for the vast majority of Facebook users by anyone on the website. (To demonstrate this we downloaded and identified the same information for a total of 9673 users at Harvard University.) Additional personal data - such as political and sexual orientation , residence address, telephone number, class schedule , etc. - are made available by the majority of users to anybody else at the same institution, leaving such data accessible to any subject able to obtain even temporary control of an institution's single email address. PRIVACY IMPLICATIONS It would appear that the population of Facebook users we have studied is, by large, quite oblivious, unconcerned, or just pragmatic about their personal privacy. Personal data is generously provided and limiting privacy preferences are sparingly used. Due to the variety and richness of personal information disclosed in Facebook profiles, their visibility, their public linkages to the members' real identities, and the scope of the network, users may put themselves at risk for a variety of attacks on their physical and online persona. Some of these risks are common also in other online social networks, while some are specific to the Facebook. In this section we outline a number of different attacks and quantify the number of users susceptible based on the data we extracted. See Table 5 for an overview. 4.1 Stalking Using the information available on profiles on the Facebook a potential adversary (with an account at the same academic institution) can determine the likely physical location of the user for large portions of the day. Facebook profiles include information about residence location, class schedule, and location of last login. A students' life during college is mostly dominated by class attendance. Therefore, knowledge of both the residence and a few classes that the student is currently attending would help a potential stalker to determine the users whereabouts. In the CMU population 860 profiles fall into our definition of this category (280 female, 580 male), in that they disclose both their current residence and at least 2 classes they are attending. Since our study was conducted outside of the semester (when many students might have deleted class information from their profiles) we speculate this number to be even higher during the semester. A much larger percentage of users is susceptible to a form of cyber-stalking using the AOL instant messenger (AIM). Unlike other messengers, AIM allows users to add "buddies" to their list without knowledge of or confirmation from the buddy being added. Once on the buddy list the adversary can track when the user is online. In the CMU population 77.7% of all profiles list an AIM screen name for a total of more than 3400 users. 4.2 Re-identification Data re-identification typically deals with the linkage of datasets without explicit identifiers such as name and address to datasets with explicit identifiers through common attributes [25]. Examples include the linkage of hospital discharge data to voter registration lists, that allows to re-identify sensitive medical information [28]. 4.2.1 Demographics re-identification It has been shown previously that a large portion of the US population can be re-identified using a combination of 5-digit ZIP code, gender, and date of birth [29]. The vast majority of CMU users disclose both their full birthdate (day and year) and gender on their profiles (88.8%). For 44.3% of users (total of 1676) the combination of birthdate and gender is unique within CMU. In addition, 50.8% list their current residence, for which ZIP codes can be easily obtained. Overall, 45.8% of users list birthday, gender, and current residence. An adversary with access to the CMU section of the Facebook could therefore link a comparatively large number of users to outside, de-identified data sources such as e.g. hospital discharge data. 4.2.2 Face Re-Identification In a related study we were able to correctly link facial images from Friendster profiles without explicit identifiers with images obtained from fully identified CMU web pages using a commercial face recognizer [13]. The field of automatic face recognition has advanced tremendously over the last decade and is now offering a number of commercial solutions which have been shown to perform well across a wide range of imaging conditions [14, 17, 24]. As shown in Section 3.4 a large number of profiles contain high quality images. At CMU more than 2500 profiles fall in this category 15 . Potential de-identified data sources include other social networking sites (e.g. Friendster) or dating sites (e.g. Match.com) that typically host anonymous profiles. 4.2.3 Social Security Numbers and Identity Theft An additional re-identification risk lies in making birthdate , hometown, current residence, and current phone number publicly available at the same time. This information can be used to estimate a person's social security number and exposes her to identity theft. The first three digits of a social security number reveal where that number was created (specifically, the digits are determined by the ZIP code of the mailing address shown on the application for a social security number). The next two digits are group identifiers, which are assigned according to a peculiar but predictable temporal order. The last four digits are progressive serial numbers. 16 When a person's hometown is known, the window of the first three digits of her SNN can be identified with probability decreasing with the home state's populousness. When that person's birthday is also known, and an attacker has access to SSNs of other people with the same birthdate in the same state as the target (for example obtained from the SSN death index or from stolen SSNs), it is possible to pin down a window of values in which the two middle digits 15 In fact, 90.8% of profiles have images, out of which 61% are estimated to be of sufficient quality for re-identification. 16 See http://www.ssa.gov/foia/stateweb.html and http://policy.ssa.gov/poms.nsf/lnx/0100201030. 78 Table 5: Overview of the privacy risks and number of CMU profiles susceptible to it. Risk # CMU Facebook Profiles % CMU Facebook Profiles 280 (Female) 15.7 (Female) Real-World Stalking 580 (Male) 21.2 (Male) Online Stalking 3528 77.7 Demographics Re-Identification 1676 44.3 Face Re-Identification 2515 (estimated) 55.4 are likely to fall. The last four digits (often used in unpro-tected logins and as passwords) can be retrieved through social engineering. Since the vast majority of the Facebook profiles we studied not only include birthday and hometown information, but also current phone number and residence (often used for verification purposes by financial institutions and other credit agencies), users are exposing themselves to substantial risks of identity theft. 4.3 Building a Digital Dossier The privacy implications of revealing personal and sensitive information (such as sexual orientation and political views) may extend beyond their immediate impact, which can be limited. Given the low and decreasing costs of storing digital information, it is possible to continuously monitor the evolution of the network and its users' profiles, thereby building a digital dossier for its participants. College students , even if currently not concerned about the visibility of their personal information, may become so as they enter sensitive and delicate jobs a few years from now - when the data currently mined could still be available. 4.4 Fragile Privacy Protection One might speculate that the perceived privacy protection of making personal information available only to members of a campus community may increase Facebook users' willingness to reveal personal information. However, the mechanisms protecting this social network can be circumvented. Adding to this the recognition that users have little control on the composition of their own networks (because often a member's friend can introduce strangers into that member's network), one may conclude that the personal information users are revealing even on sites with access control and managed search capabilities effectively becomes public data. 4.4.1 Fake Email Address The Facebook verifies users as legitimate members of a campus community by sending a confirmation email containing a link with a seemingly randomly generated nine digit code to the (campus) email address provided during registration. Since the process of signing up and receiving the confirmation email only takes minutes, an adversary simply needs to gain access to the campus network for a very short period of time. This can be achieved in a number of well-known ways, e.g. by attempting to remotely access a hacked or virus-infected machine on the network or physi-cally accessing a networked machine in e.g. the library, etc. 4.4.2 Manipulating Users Social engineering is a well-known practice in computer security to obtain confidential information by manipulating legitimate users [22]. Implementation of this practice on the Facebook is very simple: just ask to be added as someone's friend. The surprisingly high success rate of this practice was recently demonstrated by a Facebook user who, using an automatic script, contacted 250,000 users of the Facebook across the country and asked to be added as their friend. According to [15], 75,000 users accepted: thirty percent of Facebook users are willing to make all of their profile information available to a random stranger and his network of friends. 4.4.3 Advanced Search Features While not directly linked to from the site, the Facebook makes the advanced search page of any college available to anyone in the network. Using this page various profile information can be searched for, e.g. relationship status, phone number, sexual preferences, political views and (college) residence . By keeping track of the profile IDs returned in the different searches a significant portion of the previously inaccessible information can be reconstructed. CONCLUSIONS Online social networks are both vaster and looser than their offline counterparts. It is possible for somebody's profile to be connected to hundreds of peers directly, and thousands of others through the network's ties. Many individuals in a person's online extended network would hardly be defined as actual friends by that person; in fact many may be complete strangers. And yet, personal and often sensitive information is freely and publicly provided. In our study of more than 4,000 CMU users of the Facebook we have quantified individuals' willingness to provide large amounts of personal information in an online social network, and we have shown how unconcerned its users appear to privacy risks: while personal data is generously provided , limiting privacy preferences are hardly used; only a small number of members change the default privacy preferences , which are set to maximize the visibility of users profiles . Based on the information they provide online, users expose themselves to various physical and cyber risks, and make it extremely easy for third parties to create digital dossiers of their behavior. These risks are not unique to the Facebook. However, the Facebook's public linkages between an individual profile and the real identity of its owner, and the Facebook's perceived connection to a physical and ostensibly bounded community (the campus), make Facebook users a particularly interesting population for our research. Our study quantifies patterns of information revelation and infers usage of privacy settings from actual field data, rather than from surveys or laboratory experiments. Still, the relative importance of the different drivers influencing Facebook users' information revelation behavior has to be quantified. Our evidence is compatible with a number of 79 different hypotheses. In fact, many simultaneous factors are likely to play a role. Some evidence is compatible with a signalling hypothesis (see Section 3.3): users may be prag-matically publishing personal information because the benefits they expect from public disclosure surpass its perceived costs. Yet, our evidence is also compatible with an interface design explanation, such as the acceptance (and possibly ignorance ) of the default, permeable settings (see Section 3.5). Peer pressure and herding behavior may also be influencing factors, and so also myopic privacy attitudes (see [1]) and the sense of protection offered by the (perceived) bounds of a campus community. Clarifying the role of these different factors is part of our continuing research agenda. Acknowledgements We would like to thank Anne Zimmerman and Bradley Ma-lin for first bringing the Facebook to our attention. We would also like to thank danah boyd, Lorrie Cranor, Julie Downs, Steven Frank, Julia Gideon, Charis Kaskiris, Bart Nabbe, Mike Shamos, Irina Shklovski, and four anonymous referees for comments. This work was supported in part by the Data Privacy Lab in the School of Computer Science and by the Berkman Fund at Carnegie Mellon University. REFERENCES [1] A. Acquisti. Privacy in electronic commerce and the economics of immediate gratification. In Proceedings of the ACM Conference on Electronic Commerce (EC '04), pages 2129, 2004. [2] B. Anderson. Imagined Communities: Reflections on the Origin and Spread of Nationalism. Verso, London and New York, revised edition, 1991. [3] S. Arrison. Is Friendster the new TIA? TechCentralStation, January 7, 2004. [4] J. Black. The perils and promise of online schmoozing. BusinessWeek Online, February 20, 2004. [5] J. Brown. Six degrees to nowhere. Salon.com, September 21, 1998. [6] D. Cave. 16 to 25? Pentagon has your number, and more. The New York Times, June 24, 2005. [7] d. boyd. Reflections on friendster, trust and intimacy. In Intimate (Ubiquitous) Computing Workshop Ubicomp 2003, October 12-15, Seattle, Washington, USA, 2003. [8] d. boyd. Friendster and publicly articulated social networking. In Conference on Human Factors and Computing Systems (CHI 2004), April 24-29, Vienna, Austria, 2004. [9] J. Donath and d. boyd. Public displays of connection. BT Technology Journal, 22:7182, 2004. [10] S. Gerstein. Intimacy and privacy. In F. D. Schoeman, editor, Philosophical Dimensions of Privacy: An Anthology. Cambridge University Press, Cambridge, UK, 1984. [11] M. Granovetter. The strength of weak ties. American Journal of Sociology, 78:13601380, 1973. [12] M. Granovetter. The strength of weak ties: A network theory revisited. Sociological Theory, 1:201233, 1983. [13] R. Gross. Re-identifying facial images. Technical report, Carnegie Mellon University, Institute for Software Research International, 2005. In preparation. [14] R. Gross, J. Shi, and J. Cohn. Quo vadis face recognition? In Third Workshop on Empirical Evaluation Methods in Computer Vision, 2001. [15] K. Jump. A new kind of fame. The Columbian Missourian, September 1, 2005. [16] A. Leonard. You are who you know. Salon.com, June 15, 2004. [17] S. Li and A. Jain, editors. Handbook of Face Recognition. Springer Verlag, 2005. [18] H. Liu and P. Maes. Interestmap: Harvesting social network profiles for recommendations. In Beyond Personalization - IUI 2005, January 9, San Diego, California, USA, 2005. [19] W. Mackay. Triggers and barriers to customizing software. In Proceedings of CHI'91, pages 153160. ACM Press, 1991. [20] S. Milgram. The small world problem. Psychology Today, 6:6267, 1967. [21] S. Milgram. The familiar stranger: An aspect of urban anonymity. In S. Milgram, J. Sabini, and M. Silver, editors, The Individual in a Social World: Essays and Experiments. Addison-Wesley, Reading, MA, 1977. [22] K. Mitnick, W. Simon, and S. Wozniak. The art of deception: controlling the human element of security. John Wiley & Sons, 2002. [23] A. Newitz. Defenses lacking at social network sites. SecurityFocus, December 31, 2003. [24] P. Phillips, P. Flynn, T. Scruggs, K. Bowyer, J. Chang, K. Hoffman, J. Marques, J. Min, and J. Worek. Overview of the face recognition grand challenge. In IEEE Conference on Computer Vision and Pattern Recognition, June 20-25, San Diego, California, USA, 2005. [25] P. Samarati and L. Sweeney. Protecting privacy when disclosing information: k-anonymity and its enforcement through generalization and cell suppression. Technical report, SRI International, 1998. [26] I. Sege. Where everybody knows your name. Boston.com, April 27, 2005. [27] L. J. Strahilevitz. A social networks theory of privacy. The Law School, University of Chicago, John M. Olin Law & Economics Working Paper No. 230 (2D Series), December 2004. [28] L. Sweeney. k-Anonymity: a model for protecting privacy. International Journal on Uncertainty, Fuzziness and Knowledge-based Systems, 10(5):557570, 2002. [29] L. Sweeney. Uniqueness of simple demographics in the U.S. population. Technical report, Carnegie Mellon University, Laboratory for International Data Privacy, 2004. [30] The Facebook. Privacy policy. http://facebook.com/policy.php, August 2005. [31] University Planning. Carnegie Mellon Factbook 2005. Carnegie Mellon University, February 2005. [32] D. Watts. Six Degrees: The Science of a Connected Age. W.W.Norton & Company, 2003. 80

information relevation;privacy;social networking sites;information revelation;privacy risk;Online privacy;online social networking;online behavior;college;social network theory;facebook;stalking;re-identification;data visibility;privacy perference

114

Integrating the Document Object Model with Hyperlinks for Enhanced Topic Distillation and Information Extraction

Topic distillation is the process of finding authoritative Web pages and comprehensive "hubs" which reciprocally endorse each other and are relevant to a given query. Hyperlink-based topic distillation has been traditionally applied to a macroscopic Web model where documents are nodes in a directed graph and hyperlinks are edges. Macroscopic models miss valuable clues such as banners, navigation panels , and template-based inclusions, which are embedded in HTML pages using markup tags. Consequently, results of macroscopic distillation algorithms have been deteriorating in quality as Web pages are becoming more complex. We propose a uniform fine-grained model for the Web in which pages are represented by their tag trees (also called their Document Object Models or DOMs) and these DOM trees are interconnected by ordinary hyperlinks. Surprisingly, macroscopic distillation algorithms do not work in the fine-grained scenario. We present a new algorithm suitable for the fine-grained model. It can dis-aggregate hubs into coherent regions by segmenting their DOMtrees. Mutual endorsement between hubs and authorities involve these regions , rather than single nodes representing complete hubs. Anecdotes and measurements using a 28-query, 366000-document benchmark suite, used in earlier topic distillation research, reveal two benefits from the new algorithm: distillation quality improves, and a by-product of distillation is the ability to extract relevant snippets from hubs which are only partially relevant to the query.

Introduction Kleinberg's Hyperlink Induced Topic Search (HITS) [14] and the PageRank algorithm [3] underlying Google have revolutionized ranking technology for Web search engines. PageRank evaluates the "prestige score" of a page as roughly proportional to the sum of prestige scores of pages citing it (Note: To view the HTML version using Netscape, add the following line to your ~/.Xdefaults or ~/.Xresources file: Netscape*documentFonts.charset*adobe-fontspecific: iso-8859-1 For printing use the PDF version, as browsers may not print the mathematics properly.) Copyright is held by author/owner. WWW10, May 15, 2001, Hong Kong. ACM1-58113-348-0/01/0005. using hyperlinks. HITS also identifies collections of resource links or "hubs" densely coupled to authoritative pages on a topic. The model of the Web underlying these and related systems is a directed graph with pages (HTML files) as nodes and hyperlinks as edges. Since those papers were published, the Web has been evolving in fascinating ways, apart from just getting larger. Web pages are changing from static files to dynamic views generated from complex templates and backing semi-structured databases. A variety of hypertext-specific idioms such as navigation panels, advertisement banners, link exchanges, and Web-rings, have been emerging. There is also a migration of Web content from syntac-tic HTML markups towards richly tagged, semi-structured XML documents (http://www.w3.org/XML/) interconnected at the XML element level by semantically rich links (see, e.g., the XLink proposal at http://www.w3.org/TR/xlink/). These refinements are welcome steps to implementing what Berners-Lee and others call the semantic Web (http://www. w3.org/1999/04/13-tbl.html), but result in document, file, and site boundaries losing their traditional significance. Continual experiments performed by several researchers [2, 15] reveal a steady deterioration of distillation quality through the last few years. In our experience, poor results are frequently traced to the following causes: Links have become more frequent and "noisy" from the perspective of the query, such as in banners, navigation panels, and advertisements. Noisy links do not carry human editorial endorsement, a basic assumption in topic distillation. Hubs may be "mixed", meaning only a portion of the hub may be relevant to the query. Macroscopic distillation algorithms treat whole pages as atomic, indivisible nodes with no internal structure. This leads to false reinforcements and resulting contamination of the query responses. Thanks in part to the visibility of Google, content creators are well aware of hyperlink-based ranking technology. One reaction has been the proliferation of nepotistic "clique attacks"--a collection of sites linking to each other without semantic reason, e.g. http://www.411fun.com, http:// www.411fashion.com and http://www.411-loans.com. (Figures 8 and 9 provide some examples.) Some examples look suspiciously like a conscious attempt to spam search engines that use link analysis. Interestingly, in most cases, the visual presentation clearly marks noisy links which surfers rarely follow, but macroscopic algorithms are unable to exploit it. 211 &lt;html&gt; &lt;head&gt; &lt;title&gt;Portals&lt;/title&gt; &lt;/head&gt; &lt;body&gt; &lt;ul&gt; &lt;li&gt; &lt;a href="..."&gt;Yahoo&lt;/a&gt; &lt;/li&gt; &lt;li&gt; &lt;a href="..."&gt;Lycos&lt;/a&gt; &lt;/li&gt; &lt;/ul&gt; &lt;/body&gt; &lt;/html&gt; html head body title ul li li a a Figure 1: In the fine-grained model, DOMs for individual pages are trees interconnected by ordinary hyperlinks.Each triangle is the DOM tree corresponding to one HTML page.Green boxes represent text. Many had hoped that HITS-like algorithms would put an end to spamming, but clearly the situation is more like an ongoing arms-race. Google combines link-based ranking with page text and anchor text in undisclosed ways, and keeps tweaking the combination, but suffers an occasional embarrassment 1 . Distillation has always been observed to work well for "broad" topics (for which there exist well-connected relevant Web subgraphs and "pure" hubs) and not too well for "narrow" topics, because w.r.t. narrow topics most hubs are mixed and have too many irrelevant links. Mixed hubs and the arbitrariness of page boundaries have been known to produce glitches in the Clever system [6]: there has been no reliable way to classify hubs as mixed or pure. If a fine-grained model can suitably dis-aggregate mixed hubs, distillation should become applicable to narrow queries too. Yet another motivation for the fine-grained model comes from the proliferation of mobile clients such as cell-phones and PDAs with small or no screens. Even on a conventional Web browser, scrolling through search results for promising responses, then scrolling through those responses to satisfy a specific information need are tedious steps. The tedium is worse on mobile clients. Search engines that need to serve mobile clients must be able to pinpoint narrow sections of pages and sites that address a specific information need, and limit the amount of extra matter sent back to the client [4]. 1.1 Our contributions We initiate a study of topic distillation with a fine-grained model of the Web, built using the Document Object Model (DOM) of HTML pages. The DOM can model reasonably clean HTML, support XML documents that adhere to rigid schema definitions, and embed free text in a natural way. In our model, HTML pages are represented by their DOMs and these DOMtrees are interconnected by ordinary hyperlinks (figure 1). The sometimes artificial distinction between Web-level, site-level, page-level, and intra-page structures is thereby blurred. Surprisingly, macroscopic distillation algorithms perform poorly in the fine-grained setting; we demonstrate this using analysis and anecdotes. Our main technical contribution is a new fine-grained distillation al-1 http://searchenginewatch.com/sereport/99/11-google.html (local copy GoogleDrEvil.html) and http://searchenginewatch.com/ sereport/01/02-bush.html (local copy GoogleBush.html) provide some samples. Bibliometry, Graph theory PageRank/ Google HITS Clever@IBM Exploiting anchor text Topic distillation @Compaq Outlier elimination DOM structure This paper Figure 2: This work in the context of HITS and related research. gorithm which can identify mixed hubs and segment their corresponding DOMtrees into maximal subtrees which are "coherent" w.r.t. the query, i.e., each is almost completely relevant or completely irrelevant. The segmentation algorithm uses the Minimum Description Length (MDL) principle [16] from Information Theory [9]. Rather than collapse these diverse hub subtrees into one node, the new algorithm allocates a node for each subtree. This intermediate level of detail, between the macroscopic and the fine-grained model, is essential to the success of our algorithm. We report on experiments with 28 queries involving over 366000 Web pages. This benchmark has been used in previous research on resource compilation and topic distillation [5, 2, 6]. Our experience is that the fine-grained model and algorithm significantly improve the quality of distillation, and are capable of extracting DOMsubtrees from mixed hubs that are relevant to the query. We note that in this study we have carefully and deliberately isolated the model from possible influences of text analysis. By controlling our experimental environment to not use text, we push HITS-like ideas to the limit, evaluating exactly the value added by information present in DOM structures. In ongoing work, we have added textual support to our framework and obtained even better results [7]. 1.2 Benefits and applications Apart from offering a more faithful model of Web content, our approach enables solutions to the following problems. Better topic distillation: We show less tendency for topic drift and contamination when the fine-grained model is used. Web search using devices with small or no screen: The ability to identify page snippets relevant to a query is attractive to search services suitable for mobile clients. Focused crawling: Identification of relevant DOMsubtrees can be used to better guide a focused crawler's link expansion [8]. Annotation extraction: Experiments with a previous macroscopic distillation algorithm (Clever [6]) revealed that volunteers preferred Clever to Yahoo! only when Yahoo!'s manual site annotations were removed in a blind test. Our work may improve on current techniques for automatic annotation extraction [1] by first collecting candidate hub page fragments and then subjecting the text therein to further segmentation techniques. Data preparation for linguistic analysis: Information extraction is a natural next step after resource discovery. It is easier to build extractors based on statistical and linguistic models if the domain or subject matter of the input documents is suitably segmented [12], as is effected by our hub subtree extraction technique, which is a natural successor to resource discovery, and a precursor to linguistic analysis. 212 1.3 Outline of the paper In 2.1 we review HITS and related algorithms. This section can be skipped by a reader who is familiar with HITS-related literature. In 2.2 we illustrate some recent and growing threats to the continued success of macroscopic distillation algorithms. We show why the fine-grained model does not work with traditional HITS-like approaches in 3, and then propose our framework in 4. We report on experimental results in 5 and conclude in 6 with some comments on ongoing and future work. Preliminaries We review the HITS family of algorithms and discuss how they were continually enhanced to address evolving Web content. 2.1 Review of HITS and related systems The HITS algorithm [14] started with a query q which was sent to a text search engine. The returned set of pages R q was fetched from the Web, together with any pages having a link to any page in R q , as well as any page cited in some page of R q using a hyperlink. Links that connected pages on the same Web server (based on canonical host name match) were dropped from consideration because they were often seen to serve only a navigational purpose, or were "nepotistic" in nature. Suppose the resulting graph is G q = ( V q , E q ). We will drop the subscript q where clear from context. Each node v in V is assigned two scores: the hub score h(v) and the authority score a(v), initialized to any positive number. Next the HITS algorithm alternately updates a and h as follows: a(v) = (u,v)E h(u) and h(u) = (u,v)E a(v), making sure after each iteration to scale a and h so that v h(v) = v a(v) = 1, until the ranking of nodes by a and h stabilize (see figure 3). If E is represented in the adjacency matrix format (i.e., E[i, j] = 1 if there is an edge (i, j) and 0 otherwise) then the above operation can be written simply as a = E T h and h = Ea, interspersed with scaling to set |h| 1 = |a| 1 = 1. The HITS algorithm effectively uses power iterations [11] to find a, the principal eigenvector of E T E; and h, the principal eigenvector of EE T . Pages with large a are popular or authoritative sources of information; pages with large h are good collections of links. A key feature of HITS is how endorsement or popularity diffuses to siblings. If ( u, v) and (u, w) are edges and somehow a(v) becomes large, then in the next iteration h(u) will increase, and in the following iteration, a(w) will increase. We will describe this as " v's authority diffuses to w through the hub u." This is how sibling nodes reinforce each other's authority scores. We will revisit this property later in 3. Google has no notion of hubs. Roughly speaking, each page v has a single "prestige" score p(v) called its PageRank [3] which is defined as proportional to (u,v)E p(u), the sum of prestige scores of pages u that cite v. Some conjecture that the prestige model is adequate for the living Web, because good hubs readily acquire high prestige as well. Our work establishes the value of a bipartite model like HITS, and indeed, the value of an asymmetric model where hubs Expanded graph Rootset Keyword Search engine Query a = Eh h = E T a h a h h h a a a hello world stdio Centroid of rootset Similarity cone Distant vectors pruned Figure 3: (a) HITS, a macroscopic topic distillation algorithm with uniform edge weights; (b) The B&H algorithm, apart from using non-uniform edge weights, discards pages in the expanded set which are too dissimilar to the rootset pages to prevent topic drift.Documents are represented as vectors with each component representing one token or word [17]. are analyzed quite differently from authorities. Therefore we will not discuss prestige-based models any further. 2.2 The impact of the evolving Web on hyperlink analysis Elegant as the HITS model is, it does not adequately capture various idioms of Web content. We discuss here a slew of follow-up work that sought to address these issues. Kleinberg dropped links within the same Web-site from consideration because these were often found to be navigational , "nepotistic" and noisy. Shortly after HITS was published , Bharat and Henzinger (B&H [2]) found that nepotism was not limited to same-site links. In many trials with HITS, they found two distinct sites s 1 and s 2 , where s 1 hosted a number of pages u linking to a page v on s 2 , driving up a(v) beyond what may be considered fair. B&H proposed a simple and effective fix for such "site-pair" nepotism: if k pages on s 1 point to v, let the weight of each of these links be 1 /k, so that they add up to one, assuming a site (not a page) is worth one unit of voting power. Later work in the Clever system [6] used a small edge weight for same-site links and a larger weight for other links, but these weights were tuned empirically by evaluating the results on specific queries. Another issue with HITS were "mixed hubs" or pages u that included a collection of links of which only a subset was relevant to a query. Because HITS modeled u as a single node with a single h score, high authority scores could diffuse from relevant links to less relevant links. E.g., responses to the query movie awards sometimes drifted into the neighboring , more densely linked domain of movie companies. Later versions of Clever tried to address the issue in two ways. First, links within a fixed number of tokens of query terms were assigned a large edge weight (the width of the "activation window" was tuned by trial-and-error). Second, hubs which were "too long" were segmented at a few prominent boundaries (such as &lt;UL&gt; or &lt;HR&gt;) into "pagelets" with their own scores. The boundaries were chosen using a static set of rules depending on the markup tags on those pages alone. To avoid drift, B&H also computed a vector space representation [17] of documents in the response set (shown in Figure 3) and then dropped pages that were judged to be "outliers" using a suitable threshold of (cosine) similarity to the vector space centroid. B&H is effective for improving precision, but may reduce recall if mixed hubs are pruned because of small similarity to the root set centroid. This 213 Query term Activation window Figure 4: Clever uses a slightly more detailed page model than HITS.Hyperlinks near query terms are given heavier weights. Such links are shown as thicker lines. may in turn distort hub and authority scores and hence the desired ranking. Losing a few hubs may not be a problem for broad queries but could be serious for narrower queries. As resource discovery and topic distillation become more commonplace, we believe the quest will be for every additional resource than can possibly be harvested, not merely the ones that "leap out at the surfer." Our goal should therefore be to extract relevant links and annotations even from pages which are partially or largely irrelevant. Generalizing hyperlinks to interconnected DOMs HTML documents have always embedded many sources of information (other that text) which have been largely ignored in previous distillation research. Markups are one such source. From a well-formed HTML document, it ought to be possible to extract a tree structure called the Document Object Model (DOM). In real life HTML is rarely well formed, but using a few simple patches, it is possible to generate reasonably accurate DOMs. For XML sources adhering to a published DTD, a DOMis precise and well defined. For simplicity, we shall work with a greatly pared-down version of the DOMfor HTML pages. We will discard all text, and only retain those paths in the DOMtree that lead from the root to a leaf which is an &lt;A...&gt; element with an HREF leading to another page. Hyperlinks always originate from leaf DOMelements, typically deep in the DOMtree of the source document. If same-site links are ignored, very few macro-level hyperlinks target an internal node in a DOMtree (using the "#" modifier in the URL). To simplify our model (and experiments) we will assume that the target of a hyperlink is always the root node of a DOMtree. In our experiments we found very few URLs to be otherwise. A first-cut approach (which one may call MicroHITS ) would be to use the fine-grained graph directly in the HITS algorithm. One may even generalize "same-site" to "same-DOM" and use B&H-like edge-weights. This approach turns out to work rather poorly. To appreciate why, consider two simple example graphs shown in Figure 5 and their associated eigenvectors. The first graph is for the macro setting. Expanding out a E T Ea we get a(2) a(2) + a(3) and Bipar tite c o r e DOM Tree 1 2 3 = = 1 1 0 1 1 0 0 0 0 ; 0 0 0 0 0 0 1 1 0 E E E T 3 2 4 1 5 = = 1 0 0 0 0 0 1 0 1 0 0 0 0 0 0 0 1 0 2 0 0 0 0 0 0 ; 0 1 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 E E E T Figure 5: A straight-forward application of HITS-like algorithms to a DOM graph may result in some internal DOM nodes blocking the diffusion of authority across siblings. a(3) a(2) + a(3), which demonstrates the mutual reinforcement. In the second example nodes numbered 3 and 4 are part of one DOM tree. This time, we get a(2) 2a(2) + a(4) and a(4) a(2) + a(4), but there is no coupling between a(2) and a(5), which we would expect at the macroscopic level. Node 4 (marked red) effectively blocks the authority from diffusing between nodes 2 and 5. One may hope that bigger DOMtrees and multiple paths to authorities might alleviate the problem, but the above example really depicts a basic problem. The success of HITS depends critically on reinforcement among bipartite cores (see figure 5) which may be destroyed by the introduction of fine-grained nodes. Proposed model and algorithm At this point the dilemma is clear: by collapsing hubs into one node, macroscopic distillation algorithms lose valuable detail, but the more faithful fine-grained model prevents bipartite reinforcement. In this section we present our new model and distillation algorithm that resolves the dilemma. Informally, our model of hub generation enables our algorithm to find a cut or frontier across each hub's DOMtree. Subtrees attached to these cuts are made individual nodes in the distillation graph. Thus the hub score of the entire page is dis-aggregated at this intermediate level. The frontiers are not computed one time as a function of the page alone, neither do they remain unchanged during the HITS iterations. The frontiers are determined by the current estimates of the hub scores of the leaf HREF nodes. We will first describe the hub segmentation technique and then use it in a modified iterative distillation algorithm. 4.1 Scoring internal micro-hub nodes Macroscopic distillation algorithms rank and report complete hub pages, even if they are only partially relevant. In this section we address the problem of estimating the hub score of each DOMnode in the fine-grained graph, given an estimate of authority scores. Because inter-page hyperlinks originate in leaf DOMnodes and target root nodes of DOM trees, we will also assume that only those DOMnodes that are document roots can have an authority score. 214 At the end of the h Ea substep of MicroHITS, leaf DOMnodes get a hub score. Because leaf nodes point to exactly one page via an HREF, the hub score is exactly the authority score of the target page. Macroscopic distillation algorithms in effect aggregate all the leaf hub scores for a page into one hub score for the entire page. Reporting leaf hub scores in descending order would be useless, because they would simply follow the authority ranking and fail to identify good hub aggregates. Instead of the total hub score, one may consider the density of hub scores in a subtree, which may be defined as the total hub score in the subtree divided by the number of HREF leaves. The maximum density will be achieved by the leaf node that links to the best authority. In our experience small subtrees with small number of leaves dominate the top ranks, again under-aggregating hub scores and pitting ancestor scores against descendant scores. 4.1.1 A generative model for hubs To help us find suitable frontiers along which we can aggregate hub scores, we propose the following generative model for hubs. Imagine that the Web has stopped changing and with respect to a fixed query, all Web pages have been manually rated for their worth as hubs. From these hub scores, one may estimate that the hub scores have been generated from a distribution 0 . (E.g., 0 may represent an exponential distribution with mean 0 .005.) If the author of a hub page sampled URLs at random to link to, the distribution of hub scores at the leaves of the page would approach the global distribution provided enough samples were taken. However, authors differ in their choice of URLs. Hub authors are not aware of all URLs relevant to a given query or their relative authority; otherwise all hubs authored on a topic would be complete and identical, and therefore all but one would be pointless to author. (Here we momentarily ignore the value added by annotations and commentaries on hub pages.) Therefore, the distribution of hub scores for pages composed by a specific author will be different from 0 . (E.g., the author's personal average of hub scores may be 0 .002, distributed exponentially.) Moreover, the authors of mixed hubs deliberately choose to dedicate not the entire page, but only a fragment or subtree of it, to URLs that are relevant to the given query. (As an extreme case a subtree could be a single HREF.) We can regard the hub generation process as a progressive specialization of the hub score distribution starting from the global distribution. For simplicity, assume all document roots are attached to a "super-root" which corresponds to the global distribution 0 . As the author works down the DOMtree, "corrections" are applied to the score distribution at nodes on the path. At some suitable depth, the author fixes the score distribution and generates links to pages so that hub scores follow that distribution. This does not mean that there are no interesting DOMsubtrees below this depth. The model merely posits that up to some depth, DOMstructure is indicative of systematic choices of score distributions, whereas beyond that depth variation is statistical. 0 Global distribution Progressive `distortion' Model frontier Other pages v u Cumulative distortion cost = KL( 0 ; u ) + ... + KL( u ; v ) Data encoding cost is roughly v H h v h ) | Pr( log `Hot' subtree `Cold' subtree Figure 6: Our fine-grained model of Web linkage which unifies hyperlinks and DOM structure. 4.1.2 Discovering DOM frontiers from generated hubs During topic distillation we observe pages which are the outcome of the generative process described above, and our goal is to discover the "best" frontier at which the score distributions were likely to have been fixed. A balancing act is involved here: one may choose a large and redundant frontier near the leaves and model the many small, homogeneous subtrees (each with a different distribution w ) attached to that frontier accurately, or one may choose a short frontier near the root with a few subtrees which are harder to model because they contain diverse hub scores. The balancing act requires a common currency to compare the cost of the frontier with the cost of modeling hub score data beneath the frontier. This is a standard problem in segmentation, clustering, and model estimation. A particularly successful approach to optimizing the trade-off is to use the Minimum Description Length (MDL) principle [16]. MDL provides a recipe for bringing the cost of model corrections to the same units as the cost for representing data w.r.t a model, and postulates that "learning" is equivalent to minimizing the sum total of model and data encoding costs. Data encoding cost: First we consider the cost of encoding all the h-values at the leaves of a subtree rooted at node w. Specifically, let the distribution associated with w be w . The set of HREF leaf nodes in the subtree rooted at node w is denoted L w , and the set of hub scores at these leaves is denoted H w . As part of the solution we will need to evaluate the number of bits needed to encode h-values in H w using the model w . There are efficient codes which can achieve a data encoding length close to Shannon's entropy-based lower bound [9] of hH w log Pr w ( h) bits , (1) where Pr w ( h) is the probability of hub score h w.r.t. a distribution represented by w . (E.g., w may include the mean and variance of a normal distribution.) We will use this lower bound as an approximation to our data encoding cost. (This would work if the h-values followed a discrete probability distribution, which is not the case with hub scores. We will come back to this issue in 4.2.) 215 Model encoding cost: Next we consider the model encoding cost. Consider node v in the DOMtree. We will assume that 0 is known to all, and use the path from the global root to v to inductively encode each node w.r.t its parent. Suppose we want to specialize the distribution v of some v away from u , the distribution of its parent u. The cost for specifying this change is given by the well-known Kullback-Leibler (KL) distance [9] KL( u ; v ), expressed as KL( u ; v ) = x Pr u ( x) log Pr u ( x) Pr v ( x) . (2) Intuitively, this is the cost of encoding the distribution v w.r.t. a reference distribution u . E.g., if X is a binary random variable and its probabilities of being zero and one are ( .2, .8) under 1 and ( .4, .6) under 2 , then KL( 2 ; 1 ) = .4 log .4 .2 + .6 log .6 .8 . Unlike in the case of entropy, the sum can be taken to an integral in the limit for a continuous variable x. Clearly for u = v , the KL distance is zero; it can also be shown that this is a necessary condition, and that the KL distance is asymmetric in general but always non-negative. If u is specialized to v and v is specialized to w , the cost is additive, i.e., KL( u ; v ) + KL( v ; w ). We will denote the cost of such a path as KL( u ; v ; w ). Moreover, the model encoding cost of v starting from the global root model will be denoted KL( 0 ; . . . ; v ). Combined optimization problem: Given the model at the parent node u and the observed data H v , we should choose v so as to minimize the sum of the KL distance and data encoding cost: KL( v ; u ) hH v log Pr v ( h). (3) If v is expressed parametrically, this will involve an optimization over those parameters. With the above set-up, we are looking for a cut or frontier F across the tree, and for each v F , a v , such that vF KL( 0 ; . . . ; v ) hH v log Pr v ( h) (4) is minimized. The first part expresses the total model encoding cost of all nodes v on the frontier F starting from the global root distribution. The second part corresponds to the data encoding cost for the set of hub scores H v at the leaves of the subtrees rooted at the nodes v. Figure 6 illustrates the two costs. 4.2 Practical considerations The formulation above is impractical for a number of reasons . There is a reduction from the knapsack problem to the frontier-finding problem. Dynamic programming can be used to give close approximations [13, 18], but with tens of thousands of macro-level pages, each with hundreds of DOM nodes, something even simpler is needed. We describe the simplifications we had to make to control the complexity of our algorithm. We use the obvious greedy expansion strategy. We initialize our frontier with the global root and keep picking a node u from the frontier to see if expanding it to its immediate children {v} will result in a reduction in code length, if so we replace u by its children, and continue until no further improvement is possible. We compare two costs locally at each u: The cost of encoding all the data in H u with respect to model u . The cost of expanding u to its children, i.e., v KL( u ; v ), plus the cost of encoding the subtrees H v with respect to v . If the latter cost is less, we expand u, otherwise, we prune it, meaning that u becomes a frontier node. Another issue is with optimizing the model v . Usually, closed form solutions are rare and numerical optimization must be resorted to; again impractical in our setting. In practice, if H v is moderately large, the data encoding cost tends to be larger than the model cost. In such cases, a simple approximation which works quite well is to first minimize the data encoding cost for H v by picking parameter values for v that maximize the probability of the observed data (the "maximum likelihood" or ML parameters), thus fix v , then evaluate KL( u ; v ). (As an example, if a coin tossed n times turns up heads k times, the ML parameter for bias is simply k/n, but if a uniform u = U(0, 1) is chosen, the mean of v shifts slightly to ( k + 1)/(n + 2) which is a negligible change for moderately large k and n.) Non-parametric evaluation of the KL distance is complicated , and often entails density estimates. We experimented with two parametric distributions: the Gaussian and exponential distributions for which the KL distance has closed form expressions. We finally picked the exponential distribution because it fit the observed hub score distribution more closely. If represents an exponential distribution with mean and probability density f(x) = (1/) exp(-x/), then KL( 1 ; 2 ) = log 2 1 + 1 2 - 1 , (5) where i corresponds to i ( i = 1, 2). The next issue is how to measure data encoding cost for continuous variables. There is a notion of the relative entropy of a continuous distribution which generalizes discrete entropy, but the relative entropy can be negative and is useful primarily for comparing the information content in two signal sources. Therefore we need to discretize the hub scores. A common approach to discretizing real values is to scale the smallest value to one, in effect allocating log( h max /h min ) bits per value. This poses a problem in our case. Consider the larger graph in figure 5. If h is initialized to (1 , 1, 1, 1, 1) T , after the first few multiplications by EE T which represents the linear transformation ( h(1), . . . , h(5)) T (h(1) + h(3), 0, h(1) + 2h(3), h(4), 0) T , we get (2 , 0, 3, 1, 0) T , (5 , 0, 8, 1, 0) T , (13 , 0, 21, 1, 0) T , and (34 , 0, 55, 1, 0) T . Even if we disregard the zeroes, the ratio of the largest to the smallest positive component of h grows without bound. As scaling is employed to prevent overflow, h(4) decays towards zero. This makes the log(h max /h min ) strategy useless. 216 A reasonable compromise is possible by noting that the user is not interested in the precision of all the hub scores. E.g., reporting the top fraction of positive hub scores to within a small multiplicative error of is quite enough. We used = 0.8 and = 0.05 in our experiments. 4.3 Distillation using segmented hubs In this section we will embed the segmentation algorithm discussed in the previous section into the edge-weighted B&H algorithm. (Unlike the full B&H algorithm, we do no text analysis at this stage. We continue to call the edge-weighted version of HITS as "B&H" for simplicity.) The main modification will be the insertion of a call to the segmentation algorithm after the h Ea step and before the complementary step a E T h. It is also a reasonable assumption that the best frontier will segment each hub non-trivially, i.e., below its DOMroot. Therefore we can invoke the segmentation routine separately on each page. Let the segmentation algorithm described previously be invoked as F segment(u) where u is the DOMtree root of a page and F is the returned frontier for that page. Here is the pseudo-code for one iteration: h Ea for each document DOMroot u F segment(u) for each frontier node v F h(v) wL v h(w) for each w L v h(w) h(v) reset h(v) 0 a E T h normalize a so that u a(u) = 1. For convenience we can skip the hub normalization and only normalize authorities every complete cycle; this does not affect ranking. The reader will observe that this is not a linear relaxation as was the case with HITS, Clever, or B&H, because segment may lead us to aggregate and redistribute different sets of hub scores in different iterations, based on the current leaf hub scores. (Also note that if F were fixed for each page for all time, the system would still be linear and therefore guaranteed to converge.) Although convergence results for non-linear dynamical systems are rare [10], in our experiments we never found convergence to be a problem (see 5). However, we do have to take care with the initial values of a and h, unlike in the linear relaxation situation where any positive value will do. Assume that the first iteration step transfers weights from authorities to hubs, and consider how we can initialize the authority scores. In contrast to HITS, we cannot start with all a(v) = 1. Why not? Because both good and bad authorities will get this score, resulting in many hub DOMsubtrees looking more uniformly promising than they should. This will lead the segment algorithm to prune the frontier too eagerly, resulting in potentially excessive authority diffusion, as in HITS. We propose a more conservative initialization policy. Similar to B&H, we assume that the textual content of the rootset documents returned by the text search engine is more reliably relevant than the radius-1 neighbors included for distillation. Therefore we start our algorithm by assigning only root-set authority scores to one. Of course, once the iterations start, this does not prevent authority from diffusing over to siblings, but the diffusion is controlled by hub segmentation. There is one other way in which we bias our algorithm to be conservative w.r.t. authority diffusion. If a DOMnode has only one child with a positive hub score, or if there is a tie in the cost of expanding vs. pruning, we expand the node, thereby pushing the frontier down and preventing the leaf hub score from spreading out to possibly irrelevant outlinks. Taken together, these two policies may be a little too conservative, sometimes preventing desirable authority diffusion and bringing our algorithm closer to MicroHITS than we would like. For example, the graph being distilled may be such that page u has one DOMsubtree clearly (to a human reading the text) dedicated to motorcycles, but only one link target v is in the expanded set. In ongoing work we are integrating text analysis into our fine-grained model to avoid such pitfalls [7]. Experiments and results We used the 28 queries used in the Clever studies [5, 6] and by B&H [2] (shown in Figure 7). For each, RagingSearch returned at most 500 responses in the root set. These 500 28 pages were fetched and all their outlinks included in our database as well. RagingSearch and HotBot were used to get as many inlinks to the root set as possible; these were also included in our database. This resulted in about 488000 raw URLs. After normalizing URLs and eliminating duplicates, ap-proximately 366000 page fetches succeeded. We used the w3c command-line page fetching tool from http://www.w3c. org for its reliable timeout mechanism. We then scanned all these pages and filled a global (macro-)link table with 2105271 non-local links, i.e., links between pages not on the same hostname (as a lowercase string without port number). We then proceeded to parse the documents into their DOMs in order to populate a different set of tables that represented the DOMnodes and the micro-links between them. We used the javax.swing.text.html.parser package and built a custom pared-down DOMgenerator on top of the SAX scanner provided. The total number of micro-links was 9838653, and the total number of micro-nodes likewise increased. Out of the two million non-local links, less than 1% had targets that were not the root of the DOMtree of a page. Thus our introduction of the asymmetry in handling hubs and authorities seems to be not a great distortion of reality. Even though our experiments were performed on a 700 MHz Pentium Xeon processor with 512 MB RAM and 60 GB of disk, handling this scale of operation required some care and compromise. In particular, to cut down the micro-graph to only about 10 million edges, we deleted all DOMpaths that did not lead to an &lt;A...&gt;...&lt;/A&gt; element. Otherwise, we estimated that the number of micro-links would be at least two orders of magnitude larger 2 . 2 In our ongoing work we are having to address this issue as we are also analyzing text. 217 # Query Drift Mixed 1 ``affirmative action'' large 2 alcoholism 3 ``amusement park*'' small 4 architecture 5 bicycling 6 blues 7 ``classical guitar'' small 8 cheese 9 cruises 10 ``computer vision'' 11 ``field hockey'' 12 gardening 13 ``graphic design'' large 14 ``Gulf war'' large 15 HIV 16 ``lyme disease'' small 17 ``mutual fund*'' small 18 ``parallel architecture'' 19 ``rock climbing'' large 20 +recycling +can* 21 +stamp +collecting 22 Shakespeare 23 sushi small 24 telecommuting large 25 +Thailand +tourism large 26 ``table tennis'' small 27 ``vintage cars'' small 28 +Zen +buddhism large Figure 7: The set of 28 broad queries used for comparing B&H (without text analysis) and our system.The second column shows the extent of drift in the B&H response.The third column shows if mixed hubs were found within the top 50 hubs reported. Figure 7 shows the 28 queries used by the Clever study and by B&H. As indicated before, our baseline was B&H with edge-weighting but without text-based outlier elimination , which we will simply call "B&H". We did not have any arbitrary cut-off for the number of in-links used as we did not know which to discard and which to keep. As B&H noted, edge-weighting improved results significantly, but without text analysis is not adequate to prevent drift. Of the 28 queries, half show drift to some extent. We discuss a few cases. "Affirmative action" is understandably dominated by lists of US universities because they publicize their support for the same. Less intuitive was the drift into the world of software, until we found http://206.243.171.145/7927. html in the root set which presents a dialup networking software called Affirmative Action, and links to many popular freeware sites (figure 8). By itself, this page would not survive the link-based ranking, but the clique of software sites leads B&H astray. Another example was "amusement parks" where B&H fell prey to multi-host nepotism in spite of edge-weighting. A densely connected conglomerate including the relevant starting point http://www.411fun.com/THEMEPARKS/ (figure 9) formed a multi-site nepotistic cluster and misled macroscopic algorithms. In both these cases there were ample clues in the DOM structure alone (leave alone text) that authority diffusion should be suppressed. We obtained several cases of reduced drift using our technique. (In ongoing work we are getting the improvement evaluated by volunteers.) One striking example was for the query "amusement parks" where our algorithm prevented http://www.411... from taking over the show (see figure 10; complete results are in AP-macro. html and AP-micro.html). Figure 8: The part of this HTML page that contains the query affirmative action is not very popular, but adjoining DOM subtrees (upper right corner) create a dense network of software sites and mislead macroscopic distillation algorithms.Dotted red lines are drawn by hand. Figure 9: The 411 "clique attack" comprises a set of sibling sites with different hostnames and a wide variety of topics linking to each other.A human can easily avoid paying attention to the sibling sites but macroscopic distillation will get misled.Dotted red lines are drawn by hand. Figure 7 also shows that for almost half the queries, we found excellent examples of mixed hubs within the top 50 hubs reported. Given the abundance of hubs on these topics, we had anticipated that the best hubs would be "pure". While this was to some extent true, we found quite a few mixed hubs too. Our system automatically highlighted the most relevant DOMsubtree; we present some examples in figure 11 and urge the reader to sample the annotated hubs packaged with the HTML version of this paper. Macroscopic Fine-grained http://www.411boating.com http://www.411jobs.com http://www.411insure.com http://www.411hitech.com http://www.411freestuff.com http://www.411commerce.com http://www.411-realestate.com http://www.411worldtravel.com http://www.411worldsports.com http://www.411photography.com http://www.kennywood.com http://www.beachboardwalk.com http://www.sixflags.com http://www.cedarpoint.com http://www.pgathrills.com http://www.pki.com http://www.valleyfair.com http://www.silverwood4fun.com http://www.knotts.com http://www.thegreatescape.com http://www.dutchwonderland.com Figure 10: The fine-grained algorithm is less susceptible to clique attacks.The query here is amusement parks. 218 Figure 11: Two samples of mixed hub annotations: amusement parks amidst roller-coaster manufacturers and sushi amidst international cuisine. Query Annotated file alcoholism AL1.html Amusement parks AP1.html Architecture AR1.html Classical guitar CG1.html HIV HI1.html Shakespeare SH1.html Sushi SU1.html We verified that our smoothing algorithm was performing non-trivial work: it was not merely locating top-scoring authorities and highlighting them. Within the highlighted regions, we typically found as many unvisited links as links already rated as authorities. In ongoing work we are using these new links for enhanced focused crawling. A key concern for us was whether the smoothing iterations will converge or not. Because the sites of hub aggregation are data-dependent, the transform was non-linear , and we could not give a proof of convergence. In practice we faced no problems with convergence; figure 12 is typical of all queries. This raised another concern: was the smoothing subroutine doing anything dynamic and useful, or was convergence due to its picking the same sites for hub aggregation every time? In figures 13 and 14 we plot relative numbers of nodes pruned vs. expanded against the number of iterations . Queries which do not have a tendency to drift look like figure 13. Initially, both numbers are small. As the system bootstraps into controlled authority diffusion, more candidate hubs are pruned, i.e., accepted in their entirety. Diffused authority scores in turn lead to fewer nodes get-1 .00E-07 1.00E-06 1.00E-05 1.00E-04 1.00E-03 1.00E-02 1.00E-01 0 2 4 6 8 10 Iterations M e an aut h s c o r e c h ang e Figure 12: In spite of the non-linear nature of our relaxation algorithm, convergence is quick in practice.A typical chart of average change to authority scores is shown against successive iterations. ting expanded. For queries with a strong tendency to drift (figure 14), the number of nodes expanded does not drop as low as in low-drift situations. For all the 28 queries, the respective counts stabilize within 1020 iterations. 0 500 1000 1500 2000 2500 3000 3500 4000 0 1 2 3 4 5 6 7 8 9 10 #Prune #Expand Data Figure 13: Our micro-hub smoothing technique is highly adaptive: the number of nodes pruned vs.expanded changes dramatically across iterations, but stabilizes within 1020 iterations .There is also a controlled induction of new nodes into the response set owing to authority diffusion via relevant DOM subtrees (query: bicycling). 0 200 400 600 800 1000 1200 0 1 2 3 4 5 6 7 8 9 10 #Prune #Expand Data Figure 14: For some queries for which B&H showed high drift, our algorithm continues to expand a relatively larger number of nodes in an attempt to suppress drift (query: affirmative action). Finally, we checked how close we were to B&H ranking. We expected our ranking to be correlated with theirs, but verified that there are meaningful exceptions. Figure 15 show a scatter plot of authority scores. It illustrates that we systematically under-rate authorities compared to B&H (the axes have incomparable scale; the leading straight line should be interpreted as y = x). This is a natural outcome of eliminating pseudo-authorities that gain prominence in B&H via mixed hubs. 219 0 0.005 0.01 0.015 0.02 0.025 0 0.002 0.004 0.006 0.008 0.01 0.012 Authority score B&H O u r a ut h o ri t y s c ore Figure 15: Our ranking is correlated to B&H, but not identical; we tend to systematically under-rate authorities compared to B&H. Conclusion and future work We have presented a fine-grained approach to topic distillation that integrates document substructure (in the form of the Document Object Model) with regular hyperlinks. Plugging in the fine-grained graph in place of the usual coarse-grained graph does not work because the fine-grained graph may not have the bipartite cores so vital to the success of macroscopic distillation algorithms. We propose a new technique for aggregating and propagating micro-hub scores at a level determined by the Minimum Description Length principle applied to the DOMtree with hub scores at the leaves. We show that the resulting procedure still converges in practice, reduces drift, and is moreover capable of identifying and extracting regions (DOMsubtrees) relevant to the query out of a broader hub or a hub with additional less-relevant contents and links. In ongoing work, apart from completing a detailed user study (as in the Clever project), we are exploring three more ideas. First, our algorithm depends on DOMbranch points to be able to separate relevant hub fragments from irrelevant ones. We have seen some pages with a long sequence of URLs without any helpful DOMstructure such as &lt;LI&gt; providing natural segment candidates. Second, we need to bring back some of the text analysis techniques that have improved HITS and integrate them with our model. Third, we are measuring if the link localization done by our system can help in faster resource discovery. Acknowledgment: Thanks to Vivek Tawde and Hrishikesh Gupta for helpful discussions, to S. Sudarshan for stimulating discussions and generous support from the Informatics Lab, IIT Bombay, and the anonymous reviewers for helping to improve the presentation. References [1] E.Amitay and C.Paris. Automatically summarising web sites: Is there a way around it? In 9th International Conference on Information and Knowledge Management (CIKM 2000), Washington, DC, USA, 2000.ACM. Online at http://www.mri.mq.edu.au/~einat/publications/ cikm2000.pdf. [2] K.Bharat and M.Henzinger. Improved algorithms for topic distillation in a hyperlinked environment.In 21st International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 104111, Aug. 1998.Online at ftp://ftp.digital.com/pub/DEC/SRC/ publications/monika/sigir98.pdf. [3] S.Brin and L.Page. The anatomy of a large-scale hypertextual web search engine.In Proceedings of the 7th World-Wide Web Conference (WWW7), 1998.Online at http://decweb.ethz.ch/WWW7/1921/com1921.htm. [4] O.Buyukkokten, H.Garcia-Molina, and A.Paepcke. Focused web searching with PDAs.In World Wide Web Conference, Amsterdam, May 2000.Online at http://www9. org/w9cdrom/195/195.html. [5] S.Chakrabarti, B.Dom, D.Gibson, J.Kleinberg, P.Raghavan , and S.Rajagopalan. Automatic resource compilation by analyzing hyperlink structure and associated text.In 7th World-wide web conference (WWW7), 1998.Online at http://www7.scu.edu.au/programme/fullpapers/1898/ com1898.html. [6] S.Chakrabarti, B.E.Dom, S.Ravi Kumar, P.Raghavan, S.Rajagopalan, A.Tomkins, D.Gibson, and J.Kleinberg. Mining the Web's link structure. IEEE Computer, 32(8):60 67, Aug.1999. [7] S.Chakrabarti, M.Joshi, and V.Tawde. Enhanced topic distillation using text, markup tags, and hyperlinks. Submitted for publication, Jan.2001. [8] S.Chakrabarti, M.van den Berg, and B.Dom. Focused crawling: a new approach to topic-specific web resource discovery. Computer Networks, 31:16231640, 1999.First appeared in the 8th International World Wide Web Conference , Toronto, May 1999.Available online at http://www8. org/w8-papers/5a-search-query/crawling/index.html. [9] T.M.Cover and J.A.Thomas. Elements of Information Theory.John Wiley and Sons, Inc., 1991. [10] D.A.Gibson, J.M.Kleinberg, and P.Raghavan.Clustering categorical data: An approach based on dynamical systems. In VLDB, volume 24, pages 311322, New York, Aug.1998. [11] G.H.Golub and C.F.van Loan. Matrix Computations. Johns Hopkins University Press, London, 1989. [12] M.Hearst. Multi-paragraph segmentation of expository text.In Proceedings of the 32nd Annual Meeting of the Association for Computational Linguistics, Las Cruces, NM, June 1994.Online at http://www.sims.berkeley.edu/ ~hearst/publications.shtml. [13] D.S.Johnson and K.A.Niemi. On knapsacks, partitions, and a new dynamic programming technique for trees. Mathematics of Operations Research, 8(1):114, 1983. [14] J.Kleinberg. Authoritative sources in a hyperlinked environment.In ACM-SIAM Symposium on Discrete Algorithms, 1998.Online at http://www.cs.cornell.edu/ home/kleinber/auth.ps. [15] R.Lempel and S.Moran. The stochastic approach for link-structure analysis (SALSA) and the TKC effect.In WWW9, pages 387401, Amsterdam, May 2000.Online at http:// www9.org/w9cdrom/175/175.html. [16] J.Rissanen. Stochastic complexity in statistical inquiry. In World Scientific Series in Computer Science, volume 15. World Scientific, Singapore, 1989. [17] G.Salton and M.J.McGill. Introduction to Modern Information Retrieval.McGraw-Hill, 1983. [18] S.Sarawagi. Explaining differences in multidimensional aggregates.In International Conference on Very Large Databases (VLDB), volume 25, 1999.Online at http: //www.it.iitb.ernet.in/~sunita/papers/vldb99.pdf. 220

PageRank algorithm;segmentation;HITS;link localization;Topic distillation;DOM;Document Object Model;XML;microscopic distillation;text analysis;Minimum Description Length principle;Google;hub fragmentation;hyperlink;topic distillation

115

Integration of Information Assurance and Security into the IT2005 Model Curriculum

In this paper we present the context of the work of the Curriculum Committee on IT2005, the IT curriculum volume described in the Overview Draft document of the Joint Task Force for Computing Curriculum 2004. We also provide a brief introduction to the history and work of the Information Assurance Education community. These two perspectives provide the foundation for the main thrust of the paper, which is a description of the Information Assurance and Security (IAS) component of the IT2005 document. Finally, we end the paper with an example of how IAS is being implemented at BYU as a "pervasive theme" that is woven throughout the curriculum and conclude with some observations about the first year's experience.

INTRODUCTION In December 2001 a meeting (CITC-1) of interested parties from fifteen four-year IT programs from the US along with representatives from IEEE, ACM, and ABET began work on the formalization of Information Technology as an accredited academic discipline. The effort has evolved into SIGITE, the ACM SIG for Information Technology Education. During this evolution three main efforts have proceeded in parallel: 1) Definition of accreditation standards for IT programs, 2) Creation of a model curriculum for four-year IT programs, and 3) Description of the characteristics that distinguish IT programs from the sister disciplines in computing. One of the biggest challenges during the creation of the model curriculum was understanding and presenting the knowledge area that was originally called "security". Some of us were uncomfortable with the term because it was not broad enough to cover the range of concepts that we felt needed to be covered. We became aware of a community that had resolved many of the issues associated with the broader context we were seeking, Information Assurance. Information assurance has been defined as "a set of measures intended to protect and defend information and information systems by ensuring their availability, integrity, authentication, confidentiality, and non-repudiation. This includes providing for restoration of information systems by incorporating protection, detection, and reaction capabilities." The IA community and work done by IA educators became useful in defining requisite security knowledge for information technology education programs. We believe that the Information Technology and the Information Assurance Education communities have much to share. At the 9 th Colloquium for Information System Security Education in Atlanta we introduced CC2005 and IT2005 to the IA Education community[1]. In the current paper we introduce the history and current state of IA education to the SIGITE community. In addition, we demonstrate how significant concepts from the Information Assurance community have been integrated into IT2005. 1.1 CC2005 and IT2005 In the first week of December of 2001 representatives from 15 undergraduate information technology (IT) programs from across the country gathered together near Provo, Utah, to develop a community and begin to establish academic standards for this rapidly growing discipline. This first Conference on Information Technology Curriculum (CITC-1) was also attended by representatives from two professional societies, the Association for Computing Machinery (ACM) and the Institute of Electrical and Electronics Engineers, Inc. (IEEE), and also the Accreditation Board for Engineering and Technology, Inc. (ABET). This invitational conference was the culmination of an effort begun several months earlier by five of these universities who had formed a steering committee to organize a response from existing IT programs to several initiatives to define the academic discipline of IT. The steering committee wanted to ensure that the input of existing programs played a significant role in the definition of the field. A formal society and three main committees were formed by the attendees of CITC-1. The society was the Society for Information Technology Education (SITE); one of the committees formed was Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SIGITE'05, October 20-22, 2005, Newark, New Jersey, USA. Copyright 2005 ACM 1-59593-252-6/05/0010...$5.00. 7 the executive board for SITE, composed of a president, vice-president , secretary, treasurer, regional representatives, and an activities chairperson. The other two committees formed were the IT Curriculum Committee, including subcommittees for 4-year and 2-year programs, and the IT Accreditation Committee, also including subcommittees for 4-year and 2-year programs. The development of IT as an academic discipline is similar to the process that computer science (CS) went through in the 70's and 80's. In fact, looking at the placement of CS programs in academic institutions around the U.S. illustrates the debate that swirled around the discipline as its core was being defined. Some CS programs are in departments of mathematics, others are in engineering schools, and many others have become mainstay programs within newly emerging colleges of computing. Information technology, as it is practiced at this moment in its evolution, reflects similar growing pains. IT programs exist in colleges of computing, in CS departments, in schools of technology, and in business schools. Professors of information technology possess degrees in information systems, electronics, communications, graphics arts, economics, mathematics, computer science, and other disciplines. Few to none of them have a degree in information technology. It should be acknowledged here that IT has two substantially different interpretations, and that these should be clarified. Information Technology (IT) in its broadest sense encompasses all aspects of computing technology. IT, as an academic discipline, focuses on meeting the needs of users within an organizational and societal context through the selection, creation, application, integration and administration of computing technologies. A more detailed history of SIGITE is available in [2]. SIGITE is directly involved with the Joint Task Force for Computing Curriculum 2004 and has 2 representatives on the task force. This task force is a continuation of the effort that created CC2001 [3] the current computer science curriculum standard. CC2001 has been relabeled CS2001 and the current draft of the CC2004 Overview document [4] presents the structure being used to describe computing and its sub-disciplines (See Figure 1). The SIGITE Curriculum Committee is responsible for IT2005, the Information Technology Curriculum Volume. IT2005 was made available for comment in mid 2005. Figure 1 1.2 Information Assurance Education Information assurance has been defined as "a set of measures intended to protect and defend information and information systems by ensuring their availability, integrity, authentication, confidentiality, and non-repudiation. This includes providing for restoration of information systems by incorporating protection, detection, and reaction capabilities." (National Security Agency, http://www.nsa.gov/ia/iaFAQ.cfm?MenuID=10#1).[5] Information assurance education, then, includes all efforts to prepare a workforce with the needed knowledge, skills, and abilities to assure our information systems, especially critical national security systems. Information assurance education has been growing in importance and activity for the past two decades. A brief look at the involved entities and history will shed light on the growth. The National Information Assurance Education and Training Partnership (NIETP) program is a partnership among government, academia and industry focused on advancing information assurance education, training, and awareness. The NIETP was initiated in 1990 under National Security Directive 42 and has since been reauthorized several times. The NIETP serves in the capacity of national manager for information assurance education and training related to national security systems and coordinates this effort with the Committee on National Security Systems (CNSS). "The CNSS provides a forum for the discussion of policy issues, sets national policy, and promulgates direction, operational procedures, and guidance for the security of national security systems. National security systems are information systems operated by the U.S. Government, its contractors or agents that contain classified information or that: 1. involve intelligence activities; 2. involve cryptographic activities related to national security; 3. involve command and control of military forces; 4. involve equipment that is an integral part of a weapon or weapons system(s); or 5. are critical to the direct fulfillment of military or intelligence missions (not including routine administrative and business applications)." http://www.cnss.gov/history.html[6] CNSS is responsible for the development of principles, policies, guidelines, and standards that concern systems holding or related to national security information. Education and training standards are among the many standards and guidelines that CNSS issues. The training/education standards issued to date include: a) NSTISSI 1 4011 The National Training Standard for Information Systems Security Professionals, b) CNSSI 4012 The National Information Assurance Training Standard for Senior Systems Managers, c) CNSSI 4013 The National Information Assurance Training Standard for System Administrators, d) CNSSI 4014 - Information Assurance Training Standard for Information Systems Security Officers, and e) NSTISSI 4015 The National Training Standard for Systems Certifiers. CNSSI 4016 The National 1 Under Executive Order (E.O.) 13231 of October 16, 2001, Critical Infrastructure Protection in the Information Age, the President redesigned the National Security Telecommunications and Information Systems Security Committee (NSTISSC) as the Committee on National Security Systems (CNSS) 8 Training Standard for Information Security Risk Analysts will be released soon. The NSTISSI-CNSSI standards referenced above have been used to develop in-service training and education opportunities for enlisted and civilian employees in an effort to assure quality preparation of professionals entrusted with securing our critical information. In addition to providing a basis for in-service education and training, the NSTISSI-CNSSI standards have also been deployed to colleges and universities in an effort to also prepare qualified individuals preservice. The most significant effort to involve colleges and universities has been through the National Centers of Academic Excellence in Information Assurance Education (CAEIAE) Program. The CAEIAE program was started in 1998 by the National Security Agency (NSA) and is now jointly sponsored by the NSA and the Department of Homeland Security (DHS) in support of the President's National Strategy to Secure Cyberspace, February 2003. The purpose of the program is to recognize colleges and universities for their efforts in information assurance education and also to encourage more colleges and universities to develop courses and programs of study in information assurance. In order to be eligible to apply for CAEIAE certification, an institution must first demonstrate that it teaches the content covered in NSTISSI 4011 - The National Training Standard for Information Systems Security Professionals. Once an institution has been 4011 certified, it is eligible to apply for CAEIAE status. Criteria for becoming a CAEIAE include the following: a) evidence of partnerships in IA education, b) IA must be treated as a multidisciplinary science, c) evidence that the university encourages the practice of information assurance in its operations, d) demonstration of information assurance research, e) demonstration that the IA curriculum reaches beyond physical geographic borders, f) evidence of faculty productivity in information assurance research and scholarship, g) demonstration of state of the art information assurance resources, h) a declared concentration(s) in information assurance, i) a university recognized center in information assurance, and j) dedicated information assurance faculty (http://www.nsa.gov/ia/academia/caeCriteria.cfm?MenuID=10.1.1 .2).[7] In 1999, there were seven institutions designated as the inaugural CAEIAE schools. The certification is good for three years at which time institutions can reapply. Annually, an additional 6-10 institutions are awarded the certification; today, there are more than 60 CAEIAE institutions. The types of institutions and programs that are applying and being certified are growing not just in number, but also in diversity. In the first round of certification, the institutions were largely research institutions and their respective programs were at the graduate level in computer science. Today, institutions are certifying courses at the undergraduate level in computer science, management information systems, and information technology. The work being done by SIGITE is important to the further expansion of information assurance education as information assurance expands beyond the development of information systems to include the entire system life cycle including deployment, operation, maintenance, a retirement of such systems. INFORMATION ASSURANCE IN IT2005 The IT2005 volume is modeled on CS2001. It consists of 12 chapters and 2 appendices. The current draft resides at http://sigite.acm.org/activities/curriculum/[8] Chapter 1. Introduction Chapter 2. Lessons from Past Reports Chapter 3. Changes in the Information Technology Discipline Chapter 4. Principles Chapter 5. Overview of the IT Body of Knowledge Chapter 6. Overview of the Curricular Models Chapter 7. The Core of the Curriculum Chapter 8. Completing the Curriculum Chapter 9. Professional Practice Chapter 10. Characteristics of IT Graduates Chapter 11. Computing across the Curriculum Chapter 12. Institutional Challenges Acknowledgements Bibliography Appendix A. The IT Body of Knowledge Appendix B. IT Course Descriptions Chapters 5 and 7 are of particular interest for this discussion. Chapter 5 is an overview of the IT body of knowledge. A summary is included as Appendix A. Chapter 7 discusses the relationship of the core topics described in the body of knowledge to IT curriculum. IAS is explicitly mentioned in three contexts: Section 7.2 as part of the IT Fundamentals Knowledge Area (KA) Section 7.2 as a "pervasive theme" Section 7.4 as a KA that integrates the IAS concepts for students ready to graduate. IAS is the only area that is an IT Fundamental, a "pervasive theme" and also a complete KA with a recommended senior level course for integrating all of the concepts. Clearly, IT2005 presents Information Assurance and Security as a core competency required by every graduate of an IT program. During the early analysis of IT as an academic discipline, Delphi studies were performed that ranked "Security" as a central area for IT. [1] As we studied the issues several members of the committees involved were uncomfortable with "security" as the name for the knowledge area. The name seemed too restrictive. At the annual SIGITE conference in 2003 two of the authors were introduced to the other author and the Center for Research and Education in Information Assurance and Security (Cerias) at Purdue. The BYU faculty was dissatisfied with the security component in the IT curriculum and the SIGITE curriculum committee was struggling with the Security KA for IT2005. Through flyers at the conference we became aware of the Information Assurance Education Graduate Certificate (IAEGC) program funded by the NSA. With encouragement from colleagues and the administration of the School of Technology, the primary author attended the 2004 program. The experience has had a significant impact on IT2005 and the BYU curriculum. 9 We discovered that NSA had begun to use the umbrella term Information Assurance [9] to cover what we were calling security. Even though this term is defined to cover exactly what the IT community meant by security, the use of the terminology elicited a lot of blank stares. We found that explicitly adding security to the name of the knowledge area eliminated much of the confusion. We are indebted to the Center for Education and Research for Information Assurance and Security (CERIAS)[10] at Purdue whose name provided the inspiration to use IAS as a name for the knowledge area. Once the naming issue was resolved, the SIGITE curriculum committee struggled to find a model for IAS that could be understood by freshman IT students provide a framework to integrate IAS concepts that are integrated into nearly all of the other KAs be rich enough to support a senior level course that ties everything together. When A Model for Information Assurance: An Integrative Approach [11] was discovered the writing committee achieved consensus on a model. The cube (see Figure 2) provides a simple visual representation that a freshman can understand, yet the 3 dimensional structure facilitates the detailed analysis required for use in technology specific contexts, and is comprehensive enough to encompass a capstone learning experience. Figure 2 IT2005 uses this model to structure IAS concepts throughout the document. RECOMMENDATIONS FOR "PERVASIVE THEMES" IN IT2005 During the deliberations of the SIGITE Curriculum Committee, several topics emerged that were considered essential, but that did not seem to belong in a single specific knowledge area or unit. These topics, referred to as pervasive themes, are: 1. user advocacy 2. information assurance and security 3. ethics and professional responsibility 4. the ability to manage complexity through: abstraction & modeling, best practices, patterns, standards, and the use of appropriate tools 5. a deep understanding of information and communication technologies and their associated tools 6. adaptability 7. life-long learning and professional development 8. interpersonal skills The committee states "that these topics are best addressed multiple times in multiple classes, beginning in the IT fundamentals class and woven like threads throughout the tapestry of the IT curriculum"[12]. These themes need to be made explicit in the minds of the students and the faculty. The themes touch many of the topics throughout the curriculum. Every time a new technology is announced in the media, an instructor has the opportunity to drive home the importance of "life-long learning". Every time there is a cyber-crime in the media we have the opportunity to discuss the ethical and professional ramifications. It is recommended that an IT Fundamentals course be taught early in the curriculum where all of these themes are introduced and discussed as concepts that touch everything an IT professional does. Each of these topics deserves a full treatment; however, for the purposes of this paper we will focus on IAS, possibly the most pervasive theme. We will address one approach to achieve addressing IAS "multiple times in multiple classes" in section 6 below. THE INFORMATION ASSURANCE AND SECURITY KNOWLEDGE AREA In early 2003, the SIGITE curriculum committee divided into working groups around the knowledge areas defined by [3] to make an initial cut at the list of topics for each KA. A significant revision was accomplished and reviewed by the participants at the 2004 IAEGC program at Purdue in August 2004. The list of areas for the IAS KA was finalized in late 2004 at a full IT Curriculum Committee meeting. The draft of the completed IAS KA was completed in early Feb 2005 by the IAS working group, edited by the writing committee in late Feb 2005 and was presented to the full committee in April 2005. Figure 3 is a list of the IAS KA and its areas. The basic structure and vocabulary is derived directly from work done in the IA community, specifically Maconachy, et. al.[11]. The number in parenthesis is the number of lecture hours the committee thought would be required to give an IT student minimum exposure to the unit. It should be noted that the ordering of units in all of the KAs, is first "Fundamentals", if there is one, and then the units are sorted in order of the number of core hours. This ordering should not be considered as any indication of the order the units would be covered pedagogically in an implemented curriculum. 10 Figure 3 A summary of the IAS KA is in Appendix A, and a complete treatment is found in IT2005 [4], including topics, core learning outcomes, and example elective learning outcomes. In reviewing this model curriculum for IAS in Information Technology, it should be remembered that the core topics and associated lecture hours are the minimum coverage that every IT student in every program should receive. We would expect that most institutions would provide additional instruction in Information Assurance and Security according to the strengths/areas of specialization in their programs of study. IT AT BRIGHAM YOUNG UNIVERSITY The Information Technology program at BYU began officially in Fall 2001 with a faculty consisting of: 1. Two electronics engineering technology (EET) professors who were instrumental in the evolution of the existing EET program at BYU into an IT program, 2. One electrical engineering, Ph.D. newly arrived from the aerospace industry. 3. One computer science instructor who had done part time teaching and had been part of the department for 1 year with several years in system development in health care. 4. One computer science Ph.D. with recent executive management responsibilities in network hardware and service provider businesses. 5. The former department chair of the technology education program for secondary schools joined in 2002. 6. One computer science Ph. D. with extensive industry experience in data privacy and IT management joined in 2004. This is obviously a diverse group of people, each of whom joined the department because they thought that the existing computing programs at BYU did not offer students preparation for the practical aspects of system delivery to customers. We are evenly divided between long-term academics and recent `retreads' from industry. However, the academics have also each had significant industrial experience, which provided the motivation for them to accept positions in the new IT program. The BYU curriculum began as a traditional "stovepipe" approach of courses oriented around topics like networking, databases, and operating systems borrowed from CS, EET, CE and IS, and evolved to a more integrated approach starting at the introductory levels so that advanced topic oriented courses are more easily sequenced. We have also discovered that the integrative nature of IT forces a focus on the seams between technologies rather than implementation of components. This fundamental difference in focus is one of the primary differences that distinguishes IT from other computing disciplines that focus on the design and implementation of components[12] [13]. Over the last 4 years, BYU faculty has participated actively in SIGITE and attempted to share what has been learned with the emerging IT community. [14] [15] [16] The BYU curriculum has evolved into what IT2005 calls a "core/integration first" approach [17]. Significant portions of the introductory material in operating systems, databases, web systems, networking had been moved to lower division courses by early 2004. Much of the shift occurred when the introduction to web systems was moved from the junior to the sophomore year and introductory material sufficient to understand web systems was included for networking, databases, operating system administration and OS process models. The improvements in flow and reduced redundancy have been noticeable in the upper division core courses. Appendix B graphs the current BYU course structure. In late 2004 and early 2005 we began implementing the "pervasive theme" of IAS in earnest. INTEGRATING IAS INTO THE EXISTING BYU CURRICULUM A senior level IAS class had been introduced into the curriculum in early 2004 and was made a requirement in 2005. However, we recognized that simply adding a required course at the end of a student's college experience would not be adequate. SIGITE discussions had placed security in the pervasive theme category at the very beginning, though the name of the KA wasn't chosen until 2004. We were faced with the challenge of integrating the IAS fundamentals into the introductory courses, morphing the security modules in the existing classes to use the MSRW [11] framework and bringing all of the students in the program up to speed on the new framework simultaneously. Our approach has been to prepare one hour modules on the MSRW framework that can be used in an existing course to bring students up to speed or taught in seminars as needed. We are in the process of integrating the IAS Fundamentals into our introductory courses. We successfully integrated the IAS modules into the sophomore introduction to web-based systems course, which was already introducing all of the major IT areas. The course was modified to replace a 3 week team project experience with a 2 week team oriented lab and then using the time for IAS topics. Much remains to be done, but the initial experience is positive. The faculty seems unified in their desire to implement IAS as a pervasive theme. For example, 2 lecture and 2 lab hours are now included in the computer communications course. 3 lecture hours and 3 lab hours were added to the web systems course. The IAS component of the database course was rearranged and strengthened with 1 lecture hour added. Similar adjustments have been made throughout the curriculum. IAS. Information Assurance and Security (23 core hours) IAS1. Fundamental Aspects (3) IAS2. Security Mechanisms (countermeasures) (5) IAS3. Operational Issues (3) IAS4. Policy (3) IAS5. Attacks (2) IAS6. Security Domains (2) IAS7. Forensics (1) IAS8. Information States (1) IAS9. Security Services (1) IAS10. Threat Analysis Model (1) IAS11. Vulnerabilities (1) 11 In addition to improving the IAS component of the BYU curriculum, we have done an analysis of our coverage of the proposed IT2005 core. We have several adjustments in other parts of our curriculum. Since we evolved from an EET program, the hardware coverage was extremely strong. We are weak in the coverage of systems and database administration. We will continue to adjust our curriculum as IT matures as an academic discipline. SUMMARY Information Technology is maturing rapidly as an academic discipline. A public draft of the IT volume described in the Computing Curriculum 2004 Overview is ready for review. The SIGITE Curriculum Committee is soliciting feedback on the document. This paper presents a brief history of SIGITE, the ACM SIG for Information Technology Education, and a brief introduction to the Information Assurance Education community. The authors believe that collaboration between these communities can be of benefit to all of the participants and the industry at large. SIGITE and the CC 2005 Joint Task Force solicit feedback on the documents at http://www.acm.org/education/ . ACKNOWLEDGMENTS The authors would like to thank the ACM Education committee for their support of the IT2005 effort, especially Russ Shackleford, without whose financial support and encouragement the document would be years away from completion. We would also like to express appreciation to the NSA for funding the IAEGC[18] program. Corey Schou's IAEGC lecture on helping students understand IAS in an hour was the genesis of the IAS approach in IT2005. The BYU authors would like to express appreciation to our colleagues and the administration of the School of Technology at Brigham Young University, who covered our classes and found the funding for the time and travel our participation in the SIGITE curriculum committee required. REFERENCES [1] Ekstrom, Joseph J., Lunt, Barry M., Integration of Information Assurance and Security into IT2005, 9 th Colloquium for Information Systems Security Education, June 6-9, 2005, Atlanta, Georgia. [2] Lunt, Barry M.; Ekstrom, Joseph J.; Lawson, Edith A.; Kamali, Reza; Miller, Jacob; Gorka, Sandra; Reichgelt, Han; "Defining the IT Curriculum: The Results of the Last 2 Years"; World Engineer's Convention 2004, Shanghai, China; Nov 2-6, 2004 [3] Joint Task Force for Computing Curricula (2001), Computing Curricula 2001, Computer Science Volume, December 15, 2001, Copyright 2001, ACM/IEEE [4] ]Joint Task Force for Computing Curricula (2004), Computing Curricula 2004: Overview Document, http://www.acm.org/education/Overview_Draft_11-22-04 .pdf retrieved Mar. 2, 2005. [5] http://www.nsa.gov/ia/iaFAQ.cfm?MenuID=10#1 [6] http://www.cnss.gov/history.html [7] http://www.nsa.gov/ia/academia/caeCriteria.cfm?MenuID=1 0.1.1.2 [8] SIGITE Curriculum Committee (2005), Computing Curriculum 2005, IT Volume, http://sigite.acm.org/activities/curriculum/ [9] NSA web site, Information Assurance Division; http://www.nsa.gov/ia/ verified Mar, 4, 2005. [10] Cerias web site, http://cerias.purdue.edu/; verified Mar 4, 2005 [11] Machonachy, W. Victor; Schou, Corey D.; Ragsdale, Daniel; Welch , Don; "A model for Information Assurance: An Integrated Approach", Proceedings of the 2001 IEEE Workshop on Information Assurance and Security, United States Military Academy, West Point , NY, 5-6 June 2001. [12] Ekstrom, Joseph, Renshaw, Stephen, Curriculum and Issues in a First Course of Computer Networking for Four-year Information Technology Programs, ASEE 2002 Session 2793 [13] Ekstrom, Joseph, Renshaw, Stephen, A Project-Based Introductory Curriculum in Networking, WEB and Database Systems for 4-year Information Technology Programs, CITC 3 Rochester NY, September, 2002 [14] Ekstrom, Joseph, Renshaw, Stephen, Database Curriculum Issues for Four-year IT Programs, CIEC 2003, Tucson, AZ, January, 2003. [15] Ekstrom, Joseph; Lunt, Barry; Education at the Seams: Preparing Students to Stitch Systems Together; Curriculum and Issues for 4-Year IT Programs, CITC IV Purdue University, West Lafayette, Indiana, October 2003. [16] Ekstrom, Joseph; Lunt, Barry M; Helps, C. Richard; Education at the Seams: Preliminary Evaluation of Teaching Integration as a Key to Education in Information Technology; ASEE 2004, Salt Lake City, Utah, June 2004. [17] Section 6.3 of ref [4]. [18] IAEGC, Information Assurance Education Graduate Certificate, http://www.cerias.purdue.edu/iae Validated April 13, 2005. 12 Appendix A From IT2005 Mar 2005 Draft The Information Technology Body of Knowledge ITF. Information Technology Fundamentals (33 core) ITF1. Pervasive Themes in IT (17) ITF2. Organizational Issues (6) ITF3. History of IT (3) ITF4. IT and Its Related and Informing Disciplines (3) ITF5. Application Domains (2) ITF6. Applications of Math and Statistics to IT (2) HCI. Human Computer Interaction (20 core hours) HCI1. Human Factors (6) HCI2. HCI Aspects of Application Domains (3) HCI3. Human-Centered Evaluation (3) HCI4. Developing Effective Interfaces (3) HCI5. Accessibility (2) HCI6. Emerging Technologies (2) HCI7. Human-Centered Software (1) IAS. Information Assurance and Security (23 core) IAS1. Fundamental Aspects (3) IAS2. Security Mechanisms (Countermeasures) (5) IAS3. Operational Issues (3) IAS4. Policy (3) IAS5. Attacks (2) IAS6. Security Domains (2) IAS7. Forensics (1) IAS8. Information States (1) IAS9. Security Services (1) IAS10. Threat Analysis Model (1) IAS11. Vulnerabilities (1) IM. Information Management (34 core hours) IM1. IM Concepts and Fundamentals (8) IM2. Database Query Languages (9) IM3. Data Organization Architecture (7) IM4. Data Modeling (6) IM5. Managing the Database Environment (3) IM6. Special-Purpose Databases (1) IPT. Integrative Programming & Technologies (23 core) IPT1. Intersystems Communications (5) IPT2. Data Mapping and Exchange (4) IPT3. Integrative Coding (4) IPT4. Scripting Techniques (4) IPT5. Software Security Practices (4) IPT6. Miscellaneous Issues (1) IPT7. Overview of programming languages (1) NET. Networking (20 core hours) NET1. Foundations of Networking (3). NET2. Routing and Switching (8) NET3. Physical Layer (6) NET4. Security (2) NET5. Application Areas (1) NET6. Network Management PF. Programming Fundamentals (38 core hours) PF1. Fundamental Data Structures (10) PF2. Fundamental Programming Constructs (9) PF3. Object-Oriented Programming (9) PF4. Algorithms and Problem-Solving (6) PF5. Event-Driven Programming (3) PF6. Recursion (1) PT. Platform Technologies (14 core hours) PT1. Operating Systems (10) PT2. Architecture and Organization (3) PT3. Computer Infrastructure (1) PT4. Enterprise Deployment Software PT5. Firmware PT6. Hardware SA. System Administration and Maintenance (11 core hours) SA1. Operating Systems (4) SA2. Applications (3) SA3. Administrative Activities (2) SA4. Administrative Domains (2) SIA. System Integration and Architecture (21 core hours) SIA1. Requirements (6) SIA2. Acquisition/Sourcing (4) SIA3. Integration (3) SIA4. Project Management (3) SIA5. Testing and QA (3) SIA6. Organizational Context (1) SIA7. Architecture (1) SP. Social and Professional Issues (23 core hours) SP1. Technical Writing for IT (5) SP2. History of Computing (3) SP3. Social Context of Computing (3) SP4. Teamwork Concepts and Issues (3) SP5. Intellectual Properties (2) SP6. Legal Issues in Computing (2) SP7. Organizational Context (2) SP8. Professional and Ethical Issues and Responsibilities (2) SP9. Privacy and Civil Liberties (1) WS. Web Systems and Technologies (21 core hours) WS1. Web Technologies (10) WS2. Information Architecture (4) WS3. Digital Media (3) WS4. Web Development (3) WS5. Vulnerabilities (1) WS6. Social Software Total Hours: 281 Notes: 1. Order of Knowledge Areas: Fundamentals first, then ordered alphabetically. 2. Order of Units under each Knowledge Area: Fundamentals first (if present), then ordered by number of core hours. 13 Appendix B 14

Information assurance;IT2005 volume;Pervasive Themes;BYU curriculum;NIETP Program;Training standards;In-service training development;Committee on National Security Systems;CITC-1;Information Technology;IT;CC2005;IA;SIGITE Curriculum committee;Education;IT2005;Security Knowledge;Information Assurance;IAS

116

Interactive Machine Learning

Perceptual user interfaces (PUIs) are an important part of ubiquitous computing. Creating such interfaces is difficult because of the image and signal processing knowledge required for creating classifiers. We propose an interactive machine-learning (IML) model that allows users to train, classify/view and correct the classifications. The concept and implementation details of IML are discussed and contrasted with classical machine learning models. Evaluations of two algorithms are also presented. We also briefly describe Image Processing with Crayons (Crayons), which is a tool for creating new camera-based interfaces using a simple painting metaphor. The Crayons tool embodies our notions of interactive machine learning.

INTRODUCTION Perceptual user interfaces (PUIs) are establishing the need for machine learning in interactive settings. PUIs like VideoPlace [8], Light Widgets [3], and Light Table [15,16] all use cameras as their perceptive medium. Other systems use sensors other than cameras such as depth scanners and infrared sensors [13,14,15]. All of these PUIs require machine learning and computer vision techniques to create some sort of a classifier. This classification component of the UI often demands great effort and expense. Because most developers have little knowledge on how to implement recognition in their UIs this becomes problematic. Even those who do have this knowledge would benefit if the classifier building expense were lessened. We suggest the way to decrease this expense is through the use of a visual image classifier generator, which would allow developers to add intelligence to interfaces without forcing additional programming. Similar to how Visual Basic allows simple and fast development, this tool would allow for fast integration of recognition or perception into a UI. Implementation of such a tool, however, poses many problems. First and foremost is the problem of rapidly creating a satisfactory classifier. The simple solution is to using behind-the-scenes machine learning and image processing. Machine learning allows automatic creation of classifiers, however, the classical models are generally slow to train, and not interactive. The classical machine-learning (CML) model is summarized in Figure 1. Prior to the training of the classifier, features need to be selected. Training is then performed "off-line" so that classification can be done quickly and efficiently. In this model classification is optimized at the expense of longer training time. Generally, the classifier will run quickly so it can be done real-time. The assumption is that training will be performed only once and need not be interactive. Many machine-learning algorithms are very sensitive to feature selection and suffer greatly if there are very many features. Feature Selection Train Classify Interactive Use Figure 1 Classical machine learning model With CML, it is infeasible to create an interactive tool to create classifiers. CML requires the user to choose the features and wait an extended amount of time for the algorithm to train. The selection of features is very problematic for most interface designers. If one is designing an interactive technique involving laser spot tracking, most designers understand that the spot is generally red. They are not prepared to deal with how to sort out this spot from red clothing, camera noise or a variety of other problems. There are well-known image processing features for handling these problems, but very few interface designers would know how to carefully select them in a way that the machine learning algorithms could handle. The current approach requires too much technical knowledge on the part of the interface designer. What we would like to do is replace the classical machine-learning model with the interactive model shown in Figure 2. This interactive training allows the classifier to be coached along until the desired results are met. In this model the designer is correcting and Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. IUI'03, January 1215, 2003, Miami, Florida, USA. Copyright 2003 ACM 1-58113-586-6/03/0001...$5.00. 39 teaching the classifier and the classifier must perform the appropriate feature selection. Feature Selection Train Classify Interactive Use Feedback To Designer Manual Correction Figure 2 Interactive machine learning (IML) model The pre-selection of features can be eliminated and transferred to the learning part of the IML if the learning algorithm used performs feature selection. This means that a large repository of features are initially calculated and fed to the learning algorithm so it can learn the best features for the classification problem at hand. The idea is to feed a very large number of features into the classifier training and let the classifier do the filtering rather than the human. The human designer then is focused on rapidly creating training data that will correct the errors of the classifier. In classical machine learning, algorithms are evaluated on their inductive power. That is, how well the algorithm will perform on new data based on the extrapolations made on the training data. Good inductive power requires careful analysis and a great deal of computing time. This time is frequently exponential in the number of features to be considered. We believe that using the IML model a simple visual tool can be designed to build classifiers quickly. We hypothesize that when using the IML, having a very fast training algorithm is more important than strong induction. In place of careful analysis of many feature combinations we provide much more human input to correct errors as they appear. This allows the interactive cycle to be iterated quickly so it can be done more frequently. The remainder of the paper is as follows. The next section briefly discusses the visual tool we created using the IML model, called Image Processing with Crayons (Crayons). This is done to show one application of the IML model's power and versatility. Following the explanation of Crayons, we explore the details of the IML model by examining its distinction from CML, the problems it must overcome, and its implementation details. Finally we present some results from some tests between two of the implemented machine learning algorithms. From these results we base some preliminary conclusions of IML as it relates to Crayons. IMAGE PROCESSING WITH CRAYONS Crayons is a system we created that uses IML to create image classifiers. Crayons is intended to aid UI designers who do not have detailed knowledge of image processing and machine learning. It is also intended to accelerate the efforts of more knowledgeable programmers. There are two primary goals for the Crayons tool: 1) to allow the user to create an image/pixel classifier quickly, and 2) to allow the user to focus on the classification problem rather than image processing or algorithms. Crayons is successful if it takes minutes rather than weeks or months to create an effective classifier. For simplicity sake, we will refer to this as the UI principle of fast and focused. This principle refers to enabling the designer to quickly accomplish his/her task while remaining focused solely on that task. Figure 3 shows the Crayons design process. Images are input into the Crayons system, which can then export the generated classifier. It is assumed the user has already taken digital pictures and saved them as files to import into the system, or that a camera is set up on the machine running Crayons, so it can capture images from it. Exporting the classifier is equally trivial, since our implementation is written in Java. The classifier object is simply serialized and output to a file using the standard Java mechanisms. Figure 3 Classifier Design Process An overview of the internal architecture of Crayons is shown in Figure 4. Crayons receives images upon which the user does some manual classification, a classifier is created, then feedback is displayed. The user can then refine the classifier by adding more manual classification or, if the classifier is satisfactory, the user can export the classifier. The internal loop shown in Figure 4 directly correlates to the aforementioned train, feedback, correct cycle of the IML (see Figure 2). To accomplish the fast and focused UI principle, this loop must be easy and quick to cycle through. To be interactive the training part of the loop must take less than five seconds and generally much faster. The cycle can be broken down into two components: the UI and the Classifier. The UI component needs to be simple so the user can remain focused on the classification problem at hand. The classifier creation needs to be fast and efficient so the user gets feedback as quickly as possible, so they are not distracted from the classification problem. Figure 4 The classification design loop 40 Although the IML and the machine-learning component of Crayons are the primary discussion of this paper it is notable to mention that Crayons has profited from work done by Viola and Jones [19] and Jaimes and Chang [5,6,7]. Also a brief example of how Crayons can be used is illustrative. The sequence of images in Figure 5 shows the process of creating a classifier using Crayons. Figure 5 Crayons interaction process Figure 5 illustrates how the user initially paints very little data, views the feedback provided by the resulting classifier, corrects by painting additional class pixels and then iterates through the cycle. As seen in the first image pair in Figure 5, only a little data can generate a classifier that roughly learns skin and background. The classifier, however, over-generalizes in favor of background; therefore, in the second image pair you can see skin has been painted where the classifier previously did poorly at classifying skin. The resulting classifier shown on the right of the second image pair shows the new classifier classifying most of the skin on the hand, but also classifying some of the background as skin. The classifier is corrected again, and the resulting classifier is shown as the third image pair in the sequence. Thus, in only a few iterations, a skin classifier is created. The simplicity of the example above shows the power that Crayons has due to the effectiveness of the IML model. The key issue in the creation of such a tool lies in quickly generating effective classifiers so the interactive design loop can be utilized. MACHINE LEARNING For the IML model to function, the classifier must be generated quickly and be able to generalize well. As such we will first discuss the distinctions between IML and CML, followed by the problems IML must overcome because of its interactive setting, and lastly its implementation details including specific algorithms. CML vs IML Classical machine learning generally has the following assumptions. There are relatively few carefully chosen features, There is limited training data, The classifier must amplify that limited training data into excellent performance on new training data, Time to train the classifier is relatively unimportant as long as it does not take too many days. None of these assumptions hold in our interactive situation. Our UI designers have no idea what features will be appropriate. In fact, we are trying to insulate them from knowing such things. In our current Crayons prototype there are more than 150 features per pixel. To reach the breadth of application that we desire for Crayons we project over 1,000 features will be necessary. The additional features will handle texture, shape and motion over time. For any given problem somewhere between three and fifteen of those features will actually be used, but the classifier algorithm must automatically make this selection. The classifier we choose must therefore be able to accommodate such a large number of features, and/or select only the best features. In Crayons, when a designer begins to paint classes on an image a very large number of training examples is quickly generated. With 77K pixels per image and 20 images one can rapidly generate over a million training examples. In practice, the number stays in the 100K examples range because designers only paint pixels that they need to correct rather than all pixels in the image. What this means, however, is that designers can generate a huge amount of training data very quickly. CML generally focuses on the ability of a classifier to predict correct behavior on new data. In IML, however, if the classifier's predictions for new data are wrong, the designer can rapidly make those corrections. By rapid feedback and correction the classifier is quickly (in a matter of minutes) focused onto the desired behavior. The goal of the classifier is not to predict the designer's intent into new situations but rapidly reflect intent as expressed in concrete examples. Because additional training examples can be added so readily, IML's bias differs greatly from that of CML. Because it extrapolates a little data to create a classifier that will be frequently used in the future, CML is very concerned about overfit. Overfit is where the trained classifier adheres 41 too closely to the training data rather than deducing general principles. Cross-validation and other measures are generally taken to minimize overfit. These measures add substantially to the training time for CML algorithms. IML's bias is to include the human in the loop by facilitating rapid correction of mistakes. Overfit can easily occur, but it is also readily perceived by the designer and instantly corrected by the addition of new training data in exactly the areas that are most problematic. This is shown clearly in Figure 5 where a designer rapidly provides new data in the edges of the hand where the generalization failed. Our interactive classification loop requires that the classifier training be very fast. To be effective, the classifier must be generated from the training examples in under five seconds. If the classifier takes minutes or hours, the process of `train-feedback -correct' is no longer interactive, and much less effective as a design tool. Training on 100,000 examples with 150 features each in less than five seconds is a serious challenge for most CML algorithms. Lastly, for this tool to be viable the final classifier will need to be able to classify 320 x 240 images in less than a fourth of a second. If the resulting classifier is much slower than this it becomes impossible to use it to track interactive behavior in a meaningful way. IML Throughout our discussion thus far, many requirements for the machine-learning algorithm in IML have been made. The machine-learning algorithm must: learn/train very quickly, accommodate 100s to 1000s of features, perform feature selection, allow for tens to hundreds of thousands of training examples. These requirements put firm bounds on what kind of a learning algorithm can be used in IML. They invoke the fundamental question of which machine-learning algorithm fits all of these criteria. We discuss several options and the reason why they are not viable before we settle on our algorithm of choice: decision trees (DT). Neural Networks [12] are a powerful and often used machine-learning algorithm. They can provably approximate any function in two layers. Their strength lies in their abilities to intelligently integrate a variety of features. Neural networks also produce relatively small and efficient classifiers, however, there are not feasible in IML. The number of features used in systems like Crayons along with the number of hidden nodes required to produce the kinds of classifications that are necessary completely overpowers this algorithm. Even more debilitating is the training time for neural networks. The time this algorithm takes to converge is far to long for interactive use. For 150 features this can take hours or days. The nearest-neighbor algorithm [1] is easy to train but not very effective. Besides not being able to discriminate amongst features, nearest-neighbor has serious problems in high dimensional feature spaces of the kind needed in IML and Crayons. Nearest-neighbor generally has a classification time that is linear in the number of training examples which also makes it unacceptably slow. There are yet other algorithms such as boosting that do well with feature selection, which is a desirable characteristic. While boosting has shown itself to be very effective on tasks such as face tracing [18], its lengthy training time is prohibitive for interactive use in Crayons. There are many more machine-learning algorithms, however, this discussion is sufficient to preface to our decision of the use of decision trees. All the algorithms discussed above suffer from the curse of dimensionality. When many features are used (100s to 1000s), their creation and execution times dramatically increase. In addition, the number of training examples required to adequately cover such high dimensional feature spaces would far exceed what designers can produce. With just one decision per feature the size of the example set must approach 2 100 , which is completely unacceptable. We need a classifier that rapidly discards features and focuses on the 1-10 features that characterize a particular problem. Decision trees [10] have many appealing properties that coincide with the requirements of IML. First and foremost is that the DT algorithm is fundamentally a process of feature selection. The algorithm operates by examining each feature and selecting a decision point for dividing the range of that feature. It then computes the "impurity" of the result of dividing the training examples at that decision point. One can think of impurity as measuring the amount of confusion in a given set. A set of examples that all belong to one class would be pure (zero impurity). There are a variety of possible impurity measures [2]. The feature whose partition yields the least impurity is the one chosen, the set is divided and the algorithm applied recursively to the divided subsets. Features that do not provide discrimination between classes are quickly discarded. The simplicity of DTs also provides many implementation advantages in terms of speed and space of the resulting classifier. Quinlan's original DT algorithm [10] worked only on features that were discrete (a small number of choices). Our image features do not have that property. Most of our features are continuous real values. Many extensions of the original DT algorithm, ID3, have been made to allow use of realvalued data [4,11]. All of these algorithms either discretize the data or by selecting a threshold T for a given feature F divide the training examples into two sets where F&lt;T and F&gt;=T. The trick is for each feature to select a value T that gives the lowest impurity (best classification improvement). The selection of T from a large number of features and a large number of training examples is very slow to do correctly. 42 We have implemented two algorithms, which employ different division techniques. These two algorithms also represent the two approaches of longer training time with better generalization vs. shorter training time with poorer generalization. The first strategy slightly reduces interactivity and relies more on learning performance. The second relies on speed and interactivity. The two strategies are Center Weighted (CW) and Mean Split (MS). Our first DT attempt was to order all of the training examples for each feature and step through all of the examples calculating the impurity as if the division was between each of the examples. This yielded a minimum impurity split, however, this generally provided a best split close to the beginning or end of the list of examples, still leaving a large number of examples in one of the divisions. Divisions of this nature yield deeper and more unbalanced trees, which correlate to slower classification times. To improve this algorithm, we developed Center Weighted (CW), which does the same as above, except that it more heavily weights central splits (more equal divisions). By insuring that the split threshold is generally in the middle of the feature range, the resulting tree tends to be more balanced and the sizes of the training sets to be examined at each level of the tree drops exponentially. CW DTs do, however, suffer from an initial sort of all training examples for each feature, resulting in a O(f * N log N) cost up front, where f is the number of features and N the number of training examples. Since in IML, we assume that both f and N are large, this can be extremely costly. Because of the extreme initial cost of sorting all N training examples f times, we have extended Center Weighted with CWSS. The `SS' stand for sub-sampled. Since the iteration through training examples is purely to find a good split, we can sample the examples to find a statistically sound split. For example, say N is 100,000, if we sample 1,000 of the original N, sort those and calculate the best split then our initial sort is 100 times faster. It is obvious that a better threshold could be computed using all of the training data, but this is mitigated by the fact that those data items will still be considered in lower levels of the tree. When a split decision is made, all of the training examples are split, not just the sub-sample . The sub-sampling means that each node's split decision is never greater than O(f*1000*5), but that eventually all training data will be considered. Quinlan used a sampling technique called "windowing". Windowing initially used a small sample of training examples and increased the number of training examples used to create the DT, until all of the original examples were classified correctly [11]. Our technique, although similar, differs in that the number of samples is fixed. At each node in the DT a new sample of fixed size is drawn, allowing misclassified examples in a higher level of the DT to be considered at a lower level. The use of sub-sampling in CWSS produced very slight differences in classification accuracy as compared to CW, but reduced training time by a factor of at least two (for training sets with N 5,000). This factor however will continue to grow as N increases. (For N = 40,000 CWSS is approximately 5 times faster than CW; 8 for N = 80,000.) The CW and CWSS algorithms spend considerable computing resources in trying to choose a threshold value for each feature. The Mean Split (MS) algorithm spends very little time on such decisions and relies on large amounts of training data to correct decisions at lower levels of the tree. The MS algorithm uses T=mean(F) as the threshold for dividing each feature F and compares the impurities of the divisions of all features. This is very efficient and produces relatively shallow decision trees by generally dividing the training set in half at each decision point. Mean split, however, does not ensure that the division will necessarily divide the examples at points that are meaningful to correct classification. Successive splits at lower levels of the tree will eventually correctly classify the training data, but may not generalize as well. The resulting MS decision trees are not as good as those produced by more careful means such as CW or CWSS. However, we hypothesized, that the speedup in classification would improve interactivity and thus reduce the time for designers to train a classifier. We believe designers make up for the lower quality of the decision tree with the ability to correct more rapidly. The key is in optimizing designer judgment rather than classifier predictions. MSSS is a sub-sampled version of MS in the same manner as CWSS. In MSSS, since we just evaluate the impurity at the mean, and since the mean is a simple statistical value, the resulting divisions are generally identical to those of straight MS. As a parenthetical note, another important bottleneck that is common to all of the classifiers is the necessity to calculate all features initially to create the classifier. We made the assumption in IML that all features are pre-calculated and that the learning part will find the distinguishing features. Although, this can be optimized so it is faster, all algorithms will suffer from this bottleneck. There are many differences between the performances of each of the algorithms. The most important is that the CW algorithms train slower than the MS algorithms, but tend to create better classifiers. Other differences are of note though. For example, the sub sampled versions, CWSS and MSSS, generally allowed the classifiers to be generated faster. More specifically, CWSS was usually twice as fast as CW, as was MSSS compared to MS. Because of the gains in speed and lack of loss of classification power, only CWSS and MSSS will be used for comparisons. The critical comparison is to see which algorithm allows the user to create a satisfactory classifier the fastest. User tests comparing these algorithms are outlined and presented in the next section. 43 EVALUATIONS User tests were conducted to evaluate the differences between CWSS and MSSS. When creating a new perceptual interface it is not classification time that is the real issue. The important issue is designer time. As stated before, classification creation time for CWSS is longer than MSSS, but the center-weighted algorithms tend to generalize better than the mean split algorithms. The CWSS generally takes 110 seconds to train on training sets of 10,000-60,000 examples, while MSSS is approximately twice as fast on the same training sets. These differences are important; as our hypothesis was that faster classifier creation times can overcome poorer inductive strength and thus reduce overall designer time. To test the difference between CWSS and MSSS we used three key measurements: wall clock time to create the classifier, number of classify/correct iterations, and structure of the resulting tree (depth and number of nodes). The latter of these three corresponds to the amount of time the classifier takes to classify an image in actual usage. In order to test the amount of time a designer takes to create a good classifier, we need a standard to define "good classifier". A "gold standard" was created for four different classification problems: skin-detection, paper card tracking, robot car tracking and laser tracking. These gold standards were created by carefully classifying pixels until, in human judgment, the best possible classification was being performed on the test images for each problem. The resulting classifier was then saved as a standard. Ten total test subjects were used and divided into two groups. The first five did each task using the CWSS followed by the MSSS and the remaining five MSSS followed by CWSS. The users were given each of the problems in turn and asked to build a classifier. Each time the subject requested a classifier to be built that classifier's performance was measured against the performance of the standard classifier for that task. When the subject's classifier agreed with the standard on more than 97.5% of the pixels, the test was declared complete. Table 1, shows the average times and iterations for the first group, Table 2, the second group. CWSS MSSS Problem Time Iterations Time Iterations Skin 03:06 4.4 10:35 12.6 Paper Cards 02:29 4.2 02:23 5.0 Robot Car 00:50 1.6 01:00 1.6 Laser 00:46 1.2 00:52 1.4 Table 1 CWSS followed by MSSS MSSS CWSS Problem Time Iterations Time Iterations Skin 10:26 11.4 03:51 3.6 Paper Cards 04:02 5.0 02:37 2.6 Robot Car 01:48 1.2 01:37 1.2 Laser 01:29 1.0 01:16 1.0 Table 2 MSSS followed by CWSS The laser tracker is a relatively simple classifier because of the uniqueness of bright red spots [9]. The robot car was contrasted with a uniform colored carpet and was similarly straightforward. Identifying colored paper cards against a cluttered background was more difficult because of the diversity of the background. The skin tracker is the hardest because of the diversity of skin color, camera over-saturation problems and cluttered background [20]. As can be seen in tables 1 and 2, MSSS takes substantially more designer effort on the hard problems than CWSS. All subjects specifically stated that CWSS was "faster" than MSSS especially in the Skin case. (Some did not notice a difference between the two algorithms while working on the other problems.) We did not test any of the slower algorithms such as neural nets or nearest-neighbor. Interactively these are so poor that the results are self-evident. We also did not test the full CW algorithm. Its classification times tend into minutes and clearly could not compete with the times shown in tables 1 and 2. It is clear from our evaluations that a classification algorithm must get under the 10-20 second barrier in producing a new classification, but that once under that barrier, the designer's time begins to dominate. Once the designer's time begins to dominate the total time, then the classifier with better generalization wins. We also mentioned the importance of the tree structure as it relates to the classification time of an image. Table 3 shows the average tree structures (tree depth and number of nodes) as well as the average classification time (ACT) in milliseconds over the set of test images. CWSS MSSS Problem Depth Nodes ACT Depth Nodes ACT Skin 16.20 577 243 25.60 12530 375 Paper Cards 15.10 1661 201 16.20 2389 329 Car 13.60 1689 235 15.70 2859 317 Laser 13.00 4860 110 8.20 513 171 Table 3 Tree structures and average classify time (ACT) As seen in Table 3, depth, number of nodes and ACT, were all lower in CWSS than in MSSS. This was predicted as CWSS provides better divisions between the training examples. 44 While testing we observed that those who used the MSSS which is fast but less accurate, first, ended up using more training data, even when they used the CWSS, which usually generalizes better and needs less data. Those who used the CWSS first, were pleased with the interactivity of CWSS and became very frustrated when they used MSSS, even though it could cycle faster through the interactive loop. In actuality, because of the poor generalization of the mean split algorithm, even though the classifier generation time for MSSS was quicker than CWSS, the users felt it necessary to paint more using the MSSS, so the overall time increased using MSSS. CONCLUSION When using machine learning in an interactive design setting, feature selection must be automatic rather than manual and classifier training-time must be relatively fast. Decision Trees using a sub-sampling technique to improve training times are very effective for both of these purposes. Once interactive speeds are achieved, however, the quality of the classifier's generalization becomes important. Using tools like Crayons, demonstrates that machine learning can form an appropriate basis for the design tools needed to create new perceptual user interfaces. REFERENCES 1. Cover, T., and Hart, P. "Nearest Neighbor Pattern Classification." IEEE Transactions on Information Theory, 13, (1967) 21-27. 2. Duda, R. O., Hart, P. E., and Stork, D. G., Pattern Classification. (2001). 3. Fails, J.A., Olsen, D.R. "LightWidgets: Interacting in Everyday Spaces." Proceedings of IUI '02 (San Francisco CA, January 2002). 4. Fayyad, U.M. and Irani, K. B. "On the Handling of Continuous-valued Attributes in Decision Tree Generation." Machine Learning, 8, 87-102,(1992). 5. Jaimes, A. and Chang, S.-F. "A Conceptual Framework for Indexing Visual Information at Multiple Levels." IS&T/SPIE Internet Imaging 2000, (San Jose CA, January 2000). 6. Jaimes, A. and Chang, S.-F. "Automatic Selection of Visual Features and Classifier." Storage and Retrieval for Image and Video Databases VIII, IS&T/SPIE (San Jose CA, January 2000). 7. Jaimes, A. and Chang, S.-F. "Integrating Multiple Classifiers in Visual Object Detectors Learned from User Input." Invited paper, session on Image and Video Databases, 4th Asian Conference on Computer Vision (ACCV 2000), Taipei, Taiwan, January 8-11, 2000. 8. Krueger, M. W., Gionfriddo. T., and Hinrichsen, K., "VIDEOPLACE -- an artificial reality". Human Factors in Computing Systems, CHI '85 Conference Proceedings, ACM Press, 1985, 35-40. 9. Olsen, D.R., Nielsen, T. "Laser Pointer Interaction." Proceedings of CHI '01 (Seattle WA, March 2001). 10. Quinlan, J. R. "Induction of Decision Trees." Machine Learning, 1(1); 81-106, (1986). 11. Quinlan, J. R. "C4.5: Programs for machine learning." Morgan Kaufmann, San Mateo, CA, 1993. 12. Rumelhart, D., Widrow, B., and Lehr, M. "The Basic Ideas in Neural Networks." Communications of the ACM, 37(3), (1994), pp 87-92. 13. Schmidt, A. "Implicit Human Computer Interaction Through Context." Personal Technologies, Vol 4(2), June 2000. 14. Starner, T., Auxier, J. and Ashbrook, D. "The Gesture Pendant: A Self-illuminating, Wearable, Infrared Computer Vision System for Home Automation Control and Medical Monitoring." International Symposium on Wearable Computing (Atlanta GA, October 2000). 15. Triggs, B. "Model-based Sonar Localisation for Mobile Robots." Intelligent Robotic Systems '93, Zakopane, Poland, 1993. 16. Underkoffler, J. and Ishii H. "Illuminating Light: An Optical Design Tool with a Luminous-Tangible Interface." Proceedings of CHI '98 (Los Angeles CA, April 1998). 17. Underkoffler, J., Ullmer, B. and Ishii, H. "Emancipated Pixels: Real-World Graphics in the Luminous Room." Proceedings of SIGGRAPH '99 (Los Angeles CA, 1999), ACM Press, 385-392. 18. Vailaya, A., Zhong, Y., and Jain, A. K. "A hierarchical system for efficient image retrieval." In Proc. Int. Conf. on Patt. Recog. (August 1996). 19. Viola, P. and Jones, M. "Robust real-time object detection." Technical Report 2001/01, Compaq CRL, February 2001. 20. Yang, M.H. and Ahuja, N. "Gaussian Mixture Model for Human Skin Color and Its Application in Image and Video Databases." Proceedings of SPIE '99 (San Jose CA, Jan 1999), 458-466. 45

classification;Perceptual interface;image processing;perceptive user interfaces;Perceptual user iinterfaces;Machine learning;image/pixel classifier;Predict correct behaviour;Classification design loop;Interactive machine learning;interaction;Crayons prototype;Image processing with crayons;Crayons design process;Classical machine learning

117

Interestingness of Frequent Itemsets Using Bayesian Networks as Background Knowledge

The paper presents a method for pruning frequent itemsets based on background knowledge represented by a Bayesian network. The interestingness of an itemset is defined as the absolute difference between its support estimated from data and from the Bayesian network. Efficient algorithms are presented for finding interestingness of a collection of frequent itemsets, and for finding all attribute sets with a given minimum interestingness. Practical usefulness of the algorithms and their efficiency have been verified experimentally.

INTRODUCTION Finding frequent itemsets and association rules in database tables has been an active research area in recent years. Unfortunately, the practical usefulness of the approach is limited by huge number of patterns usually discovered. For larger databases many thousands of association rules may be produced when minimum support is low. This creates a secondary data mining problem: after mining the data, we are now compelled to mine the discovered patterns. The problem has been addressed in literature mainly in the context of association rules, where the two main approaches are Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. KDD'04, August 2225, 2004, Seattle, Washington, USA. Copyright 2004 ACM 1-58113-888-1/04/0008 ... $ 5.00. sorting rules based on some interestingness measure, and pruning aiming at removing redundant rules. Full review of such methods is beyond the scope of this paper. Overviews of interestingness measures can be found for example in [3, 13, 11, 32], some of the papers on rule pruning are [30, 31, 7, 14, 28, 16, 17, 33]. Many interestingness measures are based on the divergence between true probability distributions and distributions obtained under the independence assumption. Pruning methods are usually based on comparing the confidence of a rule to the confidence of rules related to it. The main drawback of those methods is that they tend to generate rules that are either obvious or have already been known by the user. This is to be expected, since the most striking patterns which those methods select can also easily be discovered using traditional methods or are known directly from experience. We believe that the proper way to address the problem is to include users background knowledge in the process. The patterns which diverge the most from that background knowledge are deemed most interesting. Discovered patterns can later be applied to improve the background knowledge itself. Many approaches to using background knowledge in machine learning are focused on using background knowledge to speed up the hypothesis discovery process and not on discovering interesting patterns. Those methods often assume strict logical relationships, not probabilistic ones. Examples are knowledge based neural networks (KBANNs) and uses of background knowledge in Inductive Logic Programming. See Chapter 12 in [20] for an overview of those methods and a list of further references. Tuzhilin et. al. [23, 22, 29] worked on applying background knowledge to finding interesting rules. In [29, 22] interestingness measures are presented, which take into account prior beliefs; in another paper [23], the authors present an algorithm for selecting a minimum set of interesting rules given background knowledge. The methods used in those papers are local, that is, they don't use a full joint probability of the data. Instead, interestingness of a rule is evaluated using rules in the background knowledge with the same consequent . If no such knowledge is present for a given rule, the rule is considered uninteresting. This makes it impossible to take into account transitivity. Indeed, in the presence of the background knowledge represented by the rules A B and 178 Research Track Paper B C, the rule A C is uninteresting. However, this cannot be discovered locally. See [25] for a detailed discussion of advantages of global versus local methods. Some more comparisons can be found in [18]. In this paper we present a method of finding interesting patterns using background knowledge represented by a Bayesian network. The main advantage of Bayesian networks is that they concisely represent full joint probability distributions, and allow for practically feasible probabilistic inference from those distributions [25, 15]. Other advantages include the ability to represent causal relationships, easy to understand graphical structure, as well as wide availability of modelling tools. Bayesian networks are also easy to modify by adding or deleting edges. We opt to compute interestingness of frequent itemsets instead of association rules, agreeing with [7] that directions of dependence should be decided by the user based on her experience and not suggested by interestingness measures. Our approach works by estimating supports of itemsets from Bayesian networks and comparing thus estimated supports with the data. Itemsets with strongly diverging supports are considered interesting. Further definitions of interestingness exploiting Bayesian network's structure are presented, as well as efficient methods for computing interestingness of large numbers of itemsets and for finding all attribute sets with given minimum interestingness. There are some analogies between mining emerging patterns [6] and our approach, the main differences being that in our case a Bayesian network is used instead of a second dataset, and that we use a different measure for comparing supports. Due to those differences our problem requires a different approach and a different set of algorithms. DEFINITIONS AND NOTATION Database attributes will be denoted with uppercase letters A, B, C, . . ., domain of an attribute A will be denoted by Dom(A). In this paper we are only concerned with categorical attributes, that is attributes with finite domains. Sets of attributes will be denoted with uppercase letters I, J, . . .. We often use database notation for representing sets of attributes, i.e. I = A 1 A 2 . . . A k instead of the set theoretical notation {A 1 , A 2 , . . . , A k }. Domain of an attribute set I = A 1 A 2 . . . A k is defined as Dom(I) = Dom(A 1 ) Dom(A 2 ) . . . Dom(A k ). Values from domains of attributes and attribute sets are denoted with corresponding lowercase boldface letters, e.g. i Dom(I). Let P I denote a joint probability distribution of the attribute set I. Similarly let P I |J be a distribution of I conditioned on J. When used in arithmetic operations such distributions will be treated as functions of attributes in I and I J respectively, with values in the interval [0, 1]. For example P I (i) denotes the probability that I = i. Let P I be a probability distribution, and let J I. Denote by P J I the marginalization of P I onto J, that is P J I = X I \J P I , (1) where the summation is over the domains of all variables from I \ J. Probability distributions estimated from data will be denoted by adding a hat symbol, e.g. ^ P I . An itemset is a pair (I, i), where I is an attribute set and i Dom(I). The support of an itemset (I, i) is defined as supp(I, i) = ^ P I (i), where the probability is estimated from some dataset. An itemset is called frequent if its support is greater than or equal to some user defined threshold minsupp. Finding all frequent itemsets in a given database table is a well known datamining problem [1]. A Bayesian network BN over a set of attributes H = A 1 . . . A n is a directed acyclic graph BN = (V, E) with the set of vertices V = {V A 1 , . . . , V A n } corresponding to attributes of H, and a set of edges E V V , where each vertex V A i has associated a conditional probability distribution P A i |par i , where par i = {A j : (V A j , V A i ) E} is the set of attributes corresponding to parents of V A i in G. See [25, 15] for a detailed discussion of Bayesian networks. A Bayesian network BN over H uniquely defines a joint probability distribution P BN H = n Y i =1 P A i |par i of H. For I H the distribution over I marginalized from P BN H will be denoted by P BN I P BN I = "P BN H " I . INTERESTINGNESS OF AN ATTRIBUTE SET WITH RESPECT TO A BAYESIAN NETWORK Let us first define the support of an itemset (I, i) in a Bayesian network BN as supp BN (I, i) = P BN I (i). Let BN be a Bayesian network over an attribute set H, and let (I, i) be an itemset such that I H. The interestingness of the itemset (I, i) with respect to BN is defined as I BN (I, i) = |supp(I, i) - supp BN (I, i)| that is, the absolute difference between the support of the itemset estimated from data, and the estimate of this support made from the Bayesian network BN . In the remaining part of the paper we assume that interestingness is always computed with respect to a Bayesian network BN and the subscript is omitted. An itemset is -interesting if its interestingness is greater than or equal to some user specified threshold . A frequent interesting itemset represents a frequently occurring (due to minimum support requirement) pattern in the database whose probability is significantly different from what it is believed to be based on the Bayesian network model. An alternative would be to use supp(I, i)/supp BN (I, i) as the measure of interestingness [6]. We decided to use absolute difference instead of a quotient since we found it to be more robust, especially when both supports are small. One could think of applying our approach to association rules with the difference in confidences as a measure of interestingness but, as mentioned in the Introduction, we think 179 Research Track Paper that patterns which do not suggest a direction of influence are more appropriate. Since in Bayesian networks dependencies are modelled using attributes not itemsets, it will often be easier to talk about interesting attribute sets, especially when the discovered interesting patterns are to be used to update the background knowledge. Definition 3.1. Let I be an attribute set. The interestingness of I is defined as I(I) = max iDom(I) I(I, i), (2) analogously, I is -interesting if I(I) . An alternative approach would be to use generalizations of Bayesian networks allowing dependencies to vary for different values of attributes, see [27], and deal with itemset interestingness directly. 3.1 Extensions to the Definition of Interestingness Even though applying the above definition and sorting attribute sets on their interestingness works well in practice, there might still be a large number of patterns retained, especially if the background knowledge is not well developed and large number of attribute sets have high interestingness values. This motivates the following two definitions. Definition 3.2. An attribute set I is hierarchically interesting if it is -interesting and none of its proper subsets is -interesting. The idea is to prevent large attribute sets from becoming interesting when the true cause of them being interesting lies in their subsets. There is also another problem with Definition 3.1. Consider a Bayesian network A B where nodes A and B have respective probability distributions P A and P B |A attached. Suppose also that A is interesting . In this case even if P B |A is the same as ^ P B |A , attribute sets B and AB may be considered -interesting. Below we present a definition of interestingness aiming at preventing such situations. A vertex V is an ancestor of a vertex W in a directed graph G if there is a directed path from V to W in G. The set of ancestors of a vertex V in a graph G is denoted by anc(V ). Moreover, let us denote by anc(I) the set of all ancestor attributes in BN of an attribute set I. More formally: anc (I) = {A i / I : V A i anc(V A j ) in BN, for some A j I}. Definition 3.3. An attribute set I is topologically -interesting if it is -interesting, and there is no attribute set J such that 1. J anc(I) I, and 2. I J, and 3. J is -interesting. The intention here is to prevent interesting attribute sets from causing all their successors in the Bayesian network (and the supersets of their successors) to become interesting in a cascading fashion. To see why condition 2 is necessary consider a Bayesian network A X B Suppose that there is a dependency between A and B in data which makes AB -interesting. Now however ABX may also become interesting, (even if P A |X and P B |X are correct in the network) and cause AB to be pruned. Condition 2 prevents AB from being pruned and ABX from becoming interesting. Notice that topological interestingness is stricter than hierarchical interestingness. Indeed if J I is -interesting, then it satisfies all the above conditions, and makes I not topologically -interesting. ALGORITHMS FOR FINDING INTERESTING ITEMSETS AND ATTRIBUTE SETS In this section we present algorithms using the definition of interestingness introduced in the previous section to select interesting itemsets or attribute sets. We begin by describing a procedure for computing marginal distributions for a large collection of attribute sets from a Bayesian network. 4.1 Computing a Large Number of Marginal Distributions from a Bayesian Network Computing the interestingness of a large number of frequent itemsets requires the computation of a large number of marginal distributions from a Bayesian network. The problem has been addressed in literature mainly in the context of finding marginals for every attribute [25, 15], while here we have to find marginals for multiple, overlapping sets of attributes. The approach taken in this paper is outlined below. The problem of computing marginal distributions from a Bayesian network is known to be NP-hard, nevertheless in most cases the network structure can be exploited to speed up the computations. Here we use exact methods for computing the marginals. Approximate methods like Gibbs sampling are an interesting topic for future work. Best known approaches to exact marginalizations are join trees [12] and bucket elimination [5]. We chose bucket elimination method which is easier to implement and according to [5] as efficient as join tree based methods. Also, join trees are mainly useful for computing marginals for single attributes, and not for sets of attributes. The bucket elimination method, which is based on the distributive law, proceeds by first choosing a variable ordering and then applying distributive law repeatedly to simplify the summation. For example suppose that a joint distribution of a Bayesian network over H = ABC is expressed as P BN ABC = P A P B |A P C |A , and we want to find P BN A . We need to compute the sum X B X C P A P B |A P C |A 180 Research Track Paper which can be rewritten as P A 0 @ X b Dom(B) P B |A 1 A 0@ X c Dom(C) P C |A 1 A. Assuming that domains of all attributes have size 3, computing the first sum directly requires 12 additions and 18 multiplications, while the second sum requires only 4 additions and 6 multiplications. The expression is interpreted as a tree of buckets, each bucket is either a single probability distribution, or a sum over a single attribute taken over a product of its child buckets in the tree. In the example above a special root bucket without summation could be introduced for completeness. In most cases the method significantly reduces the time complexity of the problem. An important problem is choosing the right variable ordering. Unfortunately that problem is itself NP-hard. We thus adopt a heuristic which orders variables according to the decreasing number of factors in the product depending on each variable. A detailed discussion of the method can be found in [5]. Although bucket elimination can be used to obtain supports of itemsets directly (i.e. P I (i)), we use it to obtain complete marginal distributions. This way we can directly apply marginalization to obtain distributions for subsets of I (see below). Since bucket elimination is performed repeatedly we use memoization to speed it up, as suggested in [21]. We remember each partial sum and reuse it if possible. In the example above P b Dom(B) P B |A , P c Dom(C) P C |A , and the computed P BN A would have been remembered. Another method of obtaining a marginal distribution P I is marginalizing it from P J where I J using Equation (1), provided that P J is already known. If |J \ I| is small, this procedure is almost always more efficient than bucket elimination , so whenever some P I is computed by bucket elimination , distributions of all subsets of I are computed using Equation (1). Definition 4.1. Let C be a collection of attribute sets. The positive border of C [19], denoted by Bd + (C), is the collection of those sets from C which have no proper superset in C: Bd + (C) = {I C : there is no J C such that I J}. It is clear from the discussion above that we only need to use bucket elimination to compute distributions of itemsets in the positive border. We are going to go further than this; we will use bucket elimination to obtain supersets of sets in the positive border, and then use Equation (1) to obtain marginals even for sets in the positive border. Experiments show that this approach can give substantial savings, especially when many overlapping attribute sets from the positive border can be covered by a single set only slightly larger then the covered ones. The algorithm for selecting the marginal distribution to compute is motivated by the algorithm from [9] for computing views that should be materialized for OLAP query processing. Bucket elimination corresponds to creating a materialized view, and marginalizing thus obtained distribution to answering OLAP queries. We first need to define costs of marginalization and bucket elimination. In our case the cost is defined as the total number of additions and multiplications used to compute the marginal distribution. The cost of marginalizing P J from P I , J I using Equation (1) is cost(P J I ) = | Dom(J)| (| Dom(I \ J)| - 1) . It follows from the fact that each value of P J I requires adding | Dom(I \ J)| values from P I . The cost of bucket elimination can be computed cheaply without actually executing the procedure. Each bucket is either an explicitly given probability distribution, or computes a sum over a single variable of a product of functions (computed in buckets contained in it) explicitly represented as multidimensional tables, see [5] for details. If the bucket is an explicitly given probability distribution, the cost is of course 0. Consider now a bucket b containing child buckets b 1 , . . . , b n yielding functions f 1 , . . . , f n respectively. Let Var(f i ) the set of attributes on which f i depends. Let f = f 1 f 2 f n denote the product of all factors in b. We have Var(f ) = n i =1 Var(f i ), and since each value of f requires n - 1 multiplications, computing f requires | Dom(Var(f ))| (n - 1) multiplications. Let A b be the attribute over which summation in b takes place. Computing the sum will require | Dom(Var(f ) \ {A b })| (| Dom(A b )| - 1) additions. So the total cost of computing the function in bucket b (including costs of computing its children) is thus cost(b) = n X i =1 cost(b i ) + | Dom(Var(f ))| (n - 1) + | Dom(Var(f ) \ {A b })| (| Dom(A b )| - 1). The cost of computing P BN I through bucket elimination, denoted cost BE (P BN I ), is the cost of the root bucket of the summation used to compute P BN I . Let C be a collection of attribute sets. The gain of using bucket elimination to find P BN I for some I while computing interestingness of attribute sets from C can be expressed as: gain(I) = -cost BE (P BN I ) + X J Bd + (C),J I hcost BE (P BN J ) - cost(P BN I J ) i. An attribute set to which bucket elimination will be applied is found using a greedy procedure by adding in each itera-tion the attribute giving the highest increase of gain. The complete algorithm is presented in Figure 1. 4.2 Computing The Interestingness of a Collection of Itemsets First we present an algorithm for computing interestingness of all itemsets in a given collection. Its a simple application of the algorithm in Figure 1. It is useful if we already 181 Research Track Paper Input: collection of attribute sets C, Bayesian network BN Output: distributions P BN I for all I C 1. S Bd + (C) 2. while S = : 3. I an attribute set from S. 4. for A in H \ I: 5. compute gain(I {A}) 6. pick A for which the gain in Step 5 was maximal 7. if gain(I {A }) &gt; gain(I): 8. I I {A } 9. goto 4 10. compute P BN I from BN using bucket elimination 11. compute P BN I J for all J S, J I using Equation (1) 12. remove from S all attribute sets included in I 13. compute P BN J for all J C \ Bd + (C) using Equation (1) Figure 1: Algorithm for computing a large number of marginal distributions from a Bayesian network. have a collection of itemsets (e.g. all frequent itemsets found in a database table) and want to select those which are the most interesting. The algorithm is given below Input: collection of itemsets K, supports of all itemsets in K, Bayesian network BN Output: interestingness of all itemsets in K. 1. C {I : (I, i) K for some i Dom(I)} 2. compute P BN I for all I C using algorithm in Figure 1 3. compute interestingness of all itemsets in K using distributions computed in step 2 4.3 Finding All Attribute Sets With Given Minimum Interestingness In this section we will present an algorithm for finding all attribute sets with interestingness greater than or equal to a specified threshold given a dataset and a Bayesian network BN . Let us first make an observation: Observation 4.2. If an itemset (I, i) has interestingness greater than or equal to with respect to a Bayesian network BN then its support must be greater than or equal to in either the data or in BN . Moreover if an attribute set is -interesting, by definition 3.1, at least one of its itemsets must be -interesting. It follows that if an attribute set is -interesting, then one of its itemsets must be frequent, with minimum support , either in the data or in the Bayesian network. Input: Bayesian network BN , minimum support minsupp. Output: itemsets whose support in BN is minsupp 1. k 1 2. Cand {(I, i) : |I| = 1} 3. compute supp BN (I, i) for all (I, i) Cand using the algorithm in Figure 1 4. F req k {(I, i) Cand : supp BN (I, i) minsupp} 5. Cand generate new candidates from F req k 6. remove itemsets with infrequent subsets from Cand 7. k k + 1; goto 3 Figure 2: The AprioriBN algorithm The algorithm works in two stages. First all frequent itemsets with minimum support are found in the dataset and their interestingness is computed. The first stage might have missed itemsets which are -interesting but don't have sufficient support in the data. In the second stage all itemsets frequent in the Bayesian network are found, and their supports in the data are computed using an extra database scan. To find all itemsets frequent in the Bayesian network we use the Apriori algorithm [1] with a modified support counting part, which we call AprioriBN. The sketch of the algorithm is shown in Figure 2, except for step 3 it is identical to the original algorithm. We now have all the elements needed to present the algorithm for finding all -interesting attribute sets, which is given in Figure 3. Step 4 of the algorithm can reuse marginal distributions found in step 3 to speed up the computations. Notice that it is always possible to compute interestingness of every itemset in step 6 since both supports of each itemset will be computed either in steps 1 and 3, or in steps 4 and 5. The authors implemented hierarchical and topological interestingness as a postprocessing step. They could however be used to prune the attribute sets which are not interesting without evaluating their distributions, thus providing a potentially large speedup in the computations. We plan to investigate that in the future. EXPERIMENTAL RESULTS In this section we present experimental evaluation of the method. One problem we were faced with was the lack of publicly available datasets with nontrivial background knowledge that could be represented as a Bayesian network . The UCI Machine Learning repository contains a few datasets with background knowledge (Japanese credit, molecular biology), but they are aimed primarily at Inductive Logic Programming: the relationships are logical rather than probabilistic, only relationships involving the class attribute are included. These examples are of little value for our approach. We have thus used networks constructed using our own 182 Research Track Paper Input: Bayesian network BN , dataset, interestingness threshold . Output: all attribute sets with interestingness at least , and some of the attribute sets with lower interestingness. 1. K {(I, i) : supp(I, i) } (using Apriori algorithm) 2. C {I : (I, i) K for some i Dom(I)} 3. compute P BN I for all I C using algorithm in Figure 1 4. K {(I, i) : supp BN (I, i) } (using AprioriBN algorithm) 5. compute support in data for all itemsets in K \ K by scanning the dataset 6. compute interestingness of all itemsets in K K 7. C {I : (I, i) K for some i Dom(I)} 8. compute interestingness of all attribute sets I in C C: I(I) = max{I(I, i) : (I, i) K K , i Dom(I)} Figure 3: Algorithm for finding all -interesting attribute sets. common-sense knowledge as well as networks learned from data. 5.1 An Illustrative Example We first present a simple example demonstrating the usefulness of the method. We use the KSL dataset of Danish 70 year olds, distributed with the DEAL Bayesian network package [4]. There are nine attributes, described in Table 1, related to the person's general health and lifestyle. All continuous attributes have been discretized into 3 levels using the equal weight method. FEV Forced ejection volume of person's lungs Kol Cholesterol Hyp Hypertension (no/yes) BMI Body Mass Index Smok Smoking (no/yes) Alc Alcohol consumption (seldom/frequently) Work Working (yes/no) Sex male/female Year Survey year (1967/1984) Table 1: Attributes of the KSL dataset. We began by designing a network structure based on au-thors' (non-expert) knowledge. The network structure is given in Figure 4a. Since we were not sure about the relation of cholesterol to other attributes, we left it unconnected. Conditional probabilities were estimated directly from the KSL dataset. Note that this is a valid approach since even when the conditional probabilities match the data perfectly interesting patterns can still be found because the network structure usually is not capable of representing the full joint distribution of the data. The interesting patterns can then be used to update the network's structure. Of course if both the structure and the conditional probabilities are given by a) b) Figure 4: Network structures for the KSL dataset constructed by the authors the expert, then the discovered patterns can be used to update both the network's structure and conditional probabilities . We applied the algorithm for finding all interesting attribute sets to the KSL dataset and the network, using the threshold of 0.01. The attribute sets returned were sorted by interestingness, and top 10 results were kept. The two most interesting attribute sets were {F EV, Sex} with interestingness 0.0812 and {Alc, Y ear} with interestingness 0.0810. Indeed, it is known (see [8]) that women's lungs are on average 20% - 25% smaller than men's lungs, so sex influences the forced ejection volume (FEV) much more than smoking does (which we thought was the primary influence). This fact, although not new in general, was overlooked by the authors, and we suspect that, due to large amount of literature on harmful effects of smoking, it might have been overlooked by many domain experts. This proves the high value of our approach for verification of Bayesian network models. The data itself implied a growth in alcohol consumption between 1967 and 1984, which we considered to be a plausible finding. We then decided to modify the network structure based on our findings by adding edges Sex F EV and Y ear Alc. One could of course consider other methods of modifying network structure, like deleting edges or reversing their direction . A brief overview of more advanced methods of Bayesian network modification can be found in [15, Chap. 3, Sect. 3.5]. Instead of adapting the network structure one could keep the structure unchanged, and tune conditional probabilities in the network instead, see [15, Chap. 3, Sect. 4] for details. 183 Research Track Paper As a method of scoring network structures we used the natural logarithm of the probability of the structure conditioned on the data, see [10, 26] for details on computing the score. The modified network structure had the score of -7162.71 which is better than that of the original network: -7356.68. With the modified structure, the most interesting attribute set was {Kol, Sex, Y ear} with interestingness 0.0665. We found in the data that cholesterol levels decreased between the two years in which the study was made, and that cholesterol level depends on sex. We found similar trends in the U.S. population based on data from American Heart Association [2]. Adding edges Y ear Kol and Sex Kol improved the network score to -7095.25. {F EV, Alc, Y ear} became the most interesting attribute set with the interestingness of 0.0286. Its interestingness is however much lower than that of previous most interesting attribute sets. Also, we were not able to get any improvement in network score after adding edges related to that attribute set. Since we were unable to obtain a better network in this case, we used topological pruning, expecting that some other attribute sets might be the true cause of the observed discrepancies . Only four attribute sets, given below, were topologically 0.01-interesting. {Kol, BM I} 0.0144 {Kol, Alc} 0.0126 {Smok, Sex, Y ear} 0.0121 {Alc, W ork} 0.0110 We found all those patters intuitively valid, but were unable to obtain an improvement in the network's score by adding related edges. Moreover, the interestingness values were quite small. We thus finished the interactive network structure improvement process with the final result given in Figure 4b. The algorithm was implemented in Python and used on a 1.7GHz Pentium 4 machine. The computation of interestingness for this example took only a few seconds so an interactive use of the program was possible. Further performance evaluation is given below. 5.2 Performance Evaluation We now present the performance evaluation of the algorithm for finding all attribute sets with given minimum interestingness . We used the UCI datasets and Bayesian networks learned from data using B-Course [26]. The results are given in Table 2. The max. size column gives the maximum size of frequent attribute sets considered. The #marginals column gives the total number of marginal distributions computed from the Bayesian network. The attribute sets whose marginal distributions have been cached between the two stages of the algorithm are not counted twice. The time does not include the initial run of the Apriori algorithm used to find frequent itemsets in the data (the time of the AprioriBN algorithm is included though). The times for larger networks can be substantial; however the proposed method has still a huge advantage over manually evaluating thousands of frequent patterns, and there are several possibilities to speed up the algorithm not yet implemented by the authors, discussed in the following section. 0 5000 10000 15000 0 50 100 150 200 250 300 no. of marginals time [s] Figure 5: Time of computation depending on the number of marginal distributions computed for the lymphography database 20 30 40 50 60 0 2000 4000 6000 8000 no. of attributes time [s] max. size = 3 max. size = 4 Figure 6: Time of computation depending on the number of attributes for datasets from Table 2 The maximum interestingness column gives the interestingness of the most interesting attribute set found for a given dataset. It can be seen that there are still highly interesting patterns to be found after using classical Bayesian network learning methods. This proves that frequent pattern and association rule mining has the capability to discover patterns which traditional methods might miss. To give a better understanding of how the algorithm scales as the problem size increases we present two additional figures . Figure 5 shows how the computation time increases with the number of marginal distributions that must be computed from the Bayesian network. It was obtained by varying the maximum size of attribute sets between 1 and 5. The value of = 0.067 was used (equivalent to one row in the database). It can be seen that the computation time grows slightly slower than the number of marginal distributions . The reason for that is that the more marginal distributions we need to compute, the more opportunities we have to avoid using bucket elimination by using direct marginalization from a superset instead. Determining how the computation time depends on the 184 Research Track Paper dataset #attrs max. size #marginals time [s] max. inter. KSL 9 0.01 5 382 1.12 0.032 soybean 36 0.075 3 7633 1292 0.064 soybean 36 0.075 4 61976 7779 0.072 breast-cancer 10 0.01 5 638 3.49 0.082 annealing 40 0.01 3 9920 1006 0.048 annealing 40 0.01 4 92171 6762 0.061 mushroom 23 0.01 3 2048 132.78 0.00036 mushroom 23 0.01 4 10903 580.65 0.00036 lymphography 19 0.067 3 1160 29.12 0.123 lymphography 19 0.067 4 5036 106.13 0.126 splice 61 0.01 3 37882 8456 0.036 Table 2: Performance evaluation of the algorithm for finding all -interesting attribute sets. size of the network is difficult, because the time depends also on the network structure and the number of marginal distributions computed (which in turn depends on the maximum size of attribute sets considered). We nevertheless show in Figure 6 the numbers of attributes and computation times plotted against each other for some of the datasets from Table 2. Data corresponding to maximum attribute set sizes equal to 3 and 4 are plotted sepa-rately . It can be seen that the algorithm remains practically usable for fairly large networks of up to 60 variables, even though the computation time grows exponentially. For larger networks approximate inference methods might be necessary , but this is beyond the scope of this paper. CONCLUSIONS AND DIRECTIONS OF FUTURE RESEARCH A method of computing interestingness of itemsets and attribute sets with respect to background knowledge encoded as a Bayesian network was presented. We built efficient algorithms for computing interestingness of frequent itemsets and finding all attribute sets with given minimum interestingness . Experimental evaluation proved the effectiveness and practical usefulness of the algorithms for finding interesting , unexpected patterns. An obvious direction for future research is increasing efficiency of the algorithms. Partial solution would be to rewrite the code in C, or to use some off-the-shelf highly op-timized Bayesian network library like Intel's PNL. Another approach would be to use approximate inference methods like Gibbs sampling. Adding or removing edges in a Bayesian network does not always influence all of its marginal distributions. Interactiv-ity of network building could be imporved by making use of this property. Usefulness of methods developed for mining emerging patterns [6], especially using borders to represent collections of itemsets, could also be investigated. Another interesting direction (suggested by a reviewer) could be to iteratively apply interesting patterns to modify the network structure until no further improvement in the network score can be achieved. A similar procedure has been used in [24] for background knowledge represented by rules. It should be noted however that it might be better to just inform the user about interesting patterns and let him/her use their experience to update the network. Manually up-dated network might better reflect causal relationships between attributes. Another research area could be evaluating other probabilistic models such as log-linear models and chain graphs instead of Bayesian networks. REFERENCES [1] R. Agrawal, T. Imielinski, and A. Swami. Mining association rules between sets of items in large databases. In Proc. ACM SIGMOD Conference on Management of Data, pages 207216, Washington, D.C., 1993. [2] American Heart Association. Risk factors: High blood cholesterol and other lipids. http://www.americanheart.org/downloadable/ heart/1045754065601FS13CHO3.pdf , 2003. [3] R. J. Bayardo and R. Agrawal. Mining the most interesting rules. In Proc. of the 5th ACM SIGKDD Int'l Conf. on Knowledge Discovery and Data Mining, pages 145154, August 1999. [4] Susanne G. Bttcher and Claus Dethlefsen. Deal: A package for learning bayesian networks. www.math.auc.dk/novo/Publications/ bottcher:dethlefsen:03.ps , 2003. [5] Rina Dechter. Bucket elimination: A unifying framework for reasoning. Artificial Intelligence, 113(1-2):4185, 1999. [6] Guozhu Dong and Jinyan Li. Efficient mining of emerging patterns: Discovering trends and differences. In Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data Mining (KDD'99), pages 4352, San Diego, CA, 1999. [7] William DuMouchel and Daryl Pregibon. Empirical bayes screening for multi-item associations. In Proceedings of the Seventh International Conference on Knowledge Discovery and Data Mining, pages 6776, 2001. [8] H. Gray. Gray's Anatomy. Grammercy Books, New York, 1977. [9] Venky Harinarayan, Anand Rajaraman, and Jeffrey D. Ullman. Implementing data cubes efficiently. In Proc. ACM SIGMOD, pages 205216, 1996. [10] David Heckerman. A tutorial on learning with Bayesian networks. Technical Report MSR-TR-95-06, Microsoft Research, Redmond, WA, 1995. 185 Research Track Paper [11] R. Hilderman and H. Hamilton. Knowledge discovery and interestingness measures: A survey. Technical Report CS 99-04, Department of Computer Science, University of Regina, 1999. [12] C. Huang and A. Darwiche. Inference in belief networks: A procedural guide. Intl. Journal of Approximate Reasoning, 15(3):225263, 1996. [13] S. Jaroszewicz and D. A. Simovici. A general measure of rule interestingness. In 5th European Conference on Principles of Data Mining and Knowledge Discovery (PKDD 2001), pages 253265, 2001. [14] S. Jaroszewicz and D. A. Simovici. Pruning redundant association rules using maximum entropy principle. In Advances in Knowledge Discovery and Data Mining, 6th Pacific-Asia Conference, PAKDD'02, pages 135147, Taipei, Taiwan, May 2002. [15] Finn V. Jensen. Bayesian Networks and Decision Graphs. Springer Verlag, New York, 2001. [16] Bing Liu, Wynne Hsu, and Shu Chen. Using general impressions to analyze discovered classification rules. In Proceedings of the Third International Conference on Knowledge Discovery and Data Mining (KDD-97), page 31. AAAI Press, 1997. [17] Bing Liu, Wynne Jsu, Yiming Ma, and Shu Chen. Mining interesting knowledge using DM-II. In Proceedings of the Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 430434, N.Y., August 1518 1999. [18] Heikki Mannila. Local and global methods in data mining: Basic techniques and open problems. In ICALP 2002, 29th International Colloquium on Automata, Languages, and Programming, Malaga, Spain, July 2002. Springer-Verlag. [19] Heikki Mannila and Hannu Toivonen. Levelwise search and borders of theories in knowledge discovery. Data Mining and Knowledge Discovery, 1(3):241258, 1997. [20] T.M. Mitchell. Machine Learning. McGraw-Hill, 1997. [21] Kevin Murphy. A brief introduction to graphical models and bayesian networks. http://www.ai.mit.edu/~murphyk/Bayes/ bnintro.html , 1998. [22] B. Padmanabhan and A. Tuzhilin. Belief-driven method for discovering unexpected patterns. In Proceedings. of the 4th International Conference on Knowledge Discovery and Data Mining (KDD'98), pages 94100, August 1998. [23] B. Padmanabhan and A. Tuzhilin. Small is beautiful: discovering the minimal set of unexpected patterns. In Proceedinmgs of the 6th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD'00), pages 5463, N. Y., August 2000. [24] B. Padmanabhan and A. Tuzhilin. Methods for knowledge refinement based on unexpected patterns. Decision Support Systems, 33(3):221347, July 2002. [25] Judea Pearl. Probabilistic Reasoning in Intelligent Systems. Morgan Kaufmann, Los Altos, CA, 1998. [26] P.Myllym aki, T.Silander, H.Tirri, and P.Uronen. B-course: A web-based tool for bayesian and causal data analysis. International Journal on Artificial Intelligence Tools, 11(3):369387, 2002. [27] D. Poole and N. L. Zhang. Exploiting contextual independence in probablisitic inference. Journal of Artificial Intelligence Research, 18:263313, 2003. [28] D. Shah, L. V. S. Lakshmanan, K. Ramamritham, and S. Sudarshan. Interestingness and pruning of mined patterns. In 1999 ACM SIGMOD Workshop on Research Issues in Data Mining and Knowledge Discovery, 1999. [29] Abraham Silberschatz and Alexander Tuzhilin. On subjective measures of interestingness in knowledge discovery. In Knowledge Discovery and Data Mining, pages 275281, 1995. [30] E. Suzuki. Autonomous discovery of reliable exception rules. In Proceedings of the Third International Conference on Knowledge Discovery and Data Mining (KDD-97), page 259. AAAI Press, 1997. [31] E. Suzuki and Y. Kodratoff. Discovery of surprising exception rules based on intensity of implication. In Proc of PKDD-98, Nantes, France, pages 1018, 1998. [32] P.-N. Tan, V. Kumar, and J. Srivastava. Selecting the right interestingness measure for association patterns. In Proc of the Eighth ACM SIGKDD Int'l Conf. on Knowledge Discovery and Data Mining (KDD-2002), pages 3241, 2002. [33] M. J. Zaki. Generating non-redundant association rules. In Proceedinmgs of the 6th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD-00), pages 3443, N. Y., August 2023 2000. 186 Research Track Paper

Bayesian network;frequent itemsets;association rules;interestingness;emerging pattern;association rule;background knowledge;frequent itemset

118

Interference Evaluation of Bluetooth and IEEE 802.11b Systems

The emergence of several radio technologies, such as Bluetooth and IEEE 802.11, operating in the 2.4 GHz unlicensed ISM frequency band, may lead to signal interference and result in significant performance degradation when devices are colocated in the same environment. The main goal of this paper is to evaluate the effect of mutual interference on the performance of Bluetooth and IEEE 802.11b systems. We develop a simulation framework for modeling interference based on detailed MAC and PHY models. First, we use a simple simulation scenario to highlight the effects of parameters, such as transmission power, offered load, and traffic type. We then turn to more complex scenarios involving multiple Bluetooth piconets and WLAN devices.

Introduction The proliferation of mobile computing devices including laptops , personal digital assistants (PDAs), and wearable computers has created a demand for wireless personal area networks (WPANs). WPANs allow closely located devices to share information and resources. A key challenge in the design of WPANs is adapting to a hostile radio environment that includes noise, time-varying channels, and abundant electromagnetic interference. Today, most radio technologies considered by WPANs (Bluetooth Special Interest Group [2], and IEEE 802.15) employ the 2.4 GHz ISM frequency band, which is also used by Local Area Network (WLAN) devices implementing the IEEE 802.11b standard specifications [9]. It is anticipated that some interference will result from all these technologies operating in the same environment . WLAN devices operating in proximity to WPAN devices may significantly impact the performance of WPAN and vice versa. The main goal of this paper is to present our findings on the performance of these systems when operating in close proximity to each other. Our results are based on detailed models for the MAC, PHY, and wireless channel. Recently, a number of research activities has led to the development of tools for wireless network simulation [1,16]. While some of these tools include a PHY layer implementation, it is often abstracted to a discrete channel model that does not implement interference per se. Therefore, in order to model interference and capture the time and frequency collisions, we chose to implement an integrated MAC-PHY module. Efforts to study interference in the 2.4 GHz band are relatively recent. For example, interference caused by microwave ovens operating in the vicinity of a WLAN network has been investigated [17] and requirements on the signal-to-noise ratio (SNR) are presented by Kamerman and Erkocevic [11]. Corresponding author. E-mail: nada.golmie@nist.gov In addition, there has been several attempts at quantifying the impact of interference on both the WLAN and Bluetooth performance. Published results can be classified into at least three categories depending on whether they rely on analysis, simulation, or experimental measurements. Analytical results based on probability of packet collision were obtained by Shellhammer [13], Ennis [4], and Zyren [18] for the WLAN packet error and by Golmie [6] for the Bluetooth packet error. In all these cases, the probability of packet error is computed based on the probability of packet collision in time and frequency. Although these analytical results can often give a first order approximation on the impact of interference and the resulting performance degradation , they often make assumptions concerning the traffic distributions and the operation of the media access protocol, which can make them less realistic. More importantly, in order for the analysis to be tractable, mutual interference that can change the traffic distribution for each system is often ig-nored . On the other hand, experimental results, such as the ones obtained by Kamerman [10], Howitt et al. [8], and Fumolari [5] for a two-node WLAN system and a two-node Bluetooth piconet, can be considered more accurate at the cost of being too specific to the implementation tested. Thus, a third alternative consists of using modeling and simulation to evaluate the impact of interference. This third approach can provide a more flexible framework. Zurbes et al. [19] present simulation results for a number of Bluetooth devices located in a single large room. They show that for 100 concurrent web sessions, performance is degraded by only 5%. Golmie et al. [7] use a detailed MAC and PHY simulation framework to evaluate the impact of interference for a pair of WLAN devices and a pair of Bluetooth devices. Similar results have been obtained by Lansford et al. [12] for the case of colocated WLAN and Bluetooth devices on the same laptop. Their simulation models are based on a link budget analysis and a theoretical calculation of the BER (Q function calculation). The work in this paper is an extension of [7]. 202 GOLMIE ET AL. Figure 1. Master TX/RX hopping sequence. This paper is organized as follows. In section 2, we give some general insights on the Bluetooth and IEEE 802.11 protocol operation. In section 3, we describe in great detail our modeling approach for the MAC, PHY and wireless channel. In section 4, we evaluate the impact of interference on both Bluetooth and WLAN performance and present simulation results . Concluding remarks are offered in section 5. Protocol overview In this section, we give a brief overview of the Bluetooth technology [2] and discuss the main functionality of its protocol specifications. Bluetooth is a short range (010 m) wireless link technology aimed at replacing non-interoperable proprietary cables that connect phones, laptops, PDAs and other portable devices together. Bluetooth operates in the ISM frequency band starting at 2.402 GHz and ending at 2.483 GHz in the USA and Europe. 79 RF channels of 1 MHz width are defined. The air interface is based on an antenna power of 1 mW with an antenna gain of 0 dB. The signal is modulated using binary Gaussian Frequency Shift Keying (GFSK). The raw data rate is defined at 1 Mbit/s. A Time Division Multiplexing (TDM) technique divides the channel into 625 s slots. Transmission occurs in packets that occupy an odd number of slots (up to 5). Each packet is transmitted on a different hop frequency with a maximum frequency hopping rate of 1600 hops/s. Two or more units communicating on the same channel form a piconet, where one unit operates as a master and the others (a maximum of seven active at the same time) act as slaves. A channel is defined as a unique pseudo-random frequency hopping sequence derived from the master device's 48-bit address and its Bluetooth clock value. Slaves in the piconet synchronize their timing and frequency hopping to the master upon connection establishment. In the connection mode, the master controls the access to the channel using a polling scheme where master and slave transmissions alternate . A slave packet always follows a master packet transmission as illustrated in figure 1, which depicts the master's view of the slotted TX/RX channel. There are two types of link connections that can be established between a master and a slave: the Synchronous Connection-Oriented (SCO), and the Asynchronous Connection-Less (ACL) link. The SCO link is a symmetric point-to-point connection between a master and a slave where the master sends an SCO packet in one TX slot at regular time intervals, defined by T SCO time slots. The slave responds with an SCO packet in the next TX opportunity. T SCO is set to either 2, 4 or 6 time slots for HV1, HV2, or HV3 packet formats, respectively. All three formats of SCO packets are defined to carry 64 Kbit/s of voice traffic and are never retransmitted in case of packet loss or error. The ACL link is an asymmetric point-to-point connection between a master and active slaves in the piconet. An Automatic Repeat Request (ARQ) procedure is applied to ACL packets where packets are retransmitted in case of loss until a positive acknowledgement (ACK) is received at the source. The ACK is piggy-backed in the header of the returned packet where an ARQN bit is set to either 1 or 0 depending on whether or not the previous packet was successfully received. In addition, a sequence number (SEQN) bit is used in the packet header in order to provide a sequential ordering of data packets in a stream and filter out retransmissions at the destination . Forward Error Correction (FEC) is used on some SCO and ACL packets in order to correct errors and reduce the number of ACL retransmissions. Both ACL and SCO packets have the same packet format. It consists of a 72-bit access code used for message identification and synchronization, a 54-bit header and a variable length payload that contains either a voice or a data packet depending on the type of link connection that is established between a master and a slave. A repetition code of rate 1/3 is applied to the header, and a block code with minimum distance, d min , equal to 14, is applied to the access code so that up to 13 errors are detected and (d min - 1)/2 = 6 can be corrected. Note that uncorrected errors in the header and the access code lead to a packet drop. Voice packets have a total packet length of 366 bits including the access code and header. A repetition code of 1/3 is used for HV1 packet payload. On the other hand, DM and HV2 packet payloads use a 2/3 block code where every 10 bits of information are encoded with 15 bits. DH and HV3 packets do not have any encoding on their payload. HV packets do INTERFERENCE EVALUATION OF BLUETOOTH AND IEEE 802.11b SYSTEMS 203 Figure 2. WLAN frame transmission scheme. Table 1 Summary of error occurrences in the packet and actions taken in case errors are not corrected. Error location Error correction Action taken Access code d min = 14 Packet dropped Packet header 1/3 repetition Packet dropped HV1 payload 1/3 repetition Packet accepted HV2 payload 2/3 block code Packet accepted HV3 payload No FEC Packet accepted DM1, DM3, DM5 payload 2/3 block code Packet dropped DH1, DH3, DH5 payload No FEC Packet accepted not have a CRC in the payload. In case of an error occurrence in the payload, the packet is never dropped. Uncorrected errors for DM and DH packets lead to dropped packets and the application of the ARQ and SEQN schemes. Table 1 summarizes the error occurrences in the packet and the actions taken by the protocol. 2.2. IEEE 802.11b The IEEE 802.11 standard [9] defines both the physical (PHY) and medium access control (MAC) layer protocols for WLANs. In this sequel, we shall be using WLAN and 802.11b interchangeably. The IEEE 802.11 standard calls for three different PHY specifications: frequency hopping (FH) spread spectrum, direct sequence (DS) spread spectrum, and infrared (IR). The transmit power for DS and FH devices is defined at a maximum of 1 W and the receiver sensitivity is set to -80 dBmW. Antenna gain is limited to 6 dB maximum. In this work, we focus on the 802.11b specification (DS spread spectrum) since it is in the same frequency band as Bluetooth and the most commonly deployed. The basic data rate for the DS system is 1 Mbit/s encoded with differential binary phase shift keying (DBPSK). Similarly , a 2 Mbit/s rate is provided using differential quadrature phase shift keying (DQPSK) at the same chip rate of 11 10 6 chips/s. Higher rates of 5.5 and 11 Mbit/s are also available using techniques combining quadrature phase shift keying and complementary code keying (CCK); all of these systems use 22 MHz channels. The IEEE 802.11 MAC layer specifications, common to all PHYs and data rates, coordinate the communication between stations and control the behavior of users who want to access the network. The Distributed Coordination Function (DCF), which describes the default MAC protocol operation, is based on a scheme known as carrier-sense, multiple access, collision avoidance (CSMA/CA). Both the MAC and PHY layers cooperate in order to implement collision avoidance procedures. The PHY layer samples the received energy over the medium transmitting data and uses a clear channel assessment (CCA) algorithm to determine if the channel is clear. This is accomplished by measuring the RF energy at the antenna and determining the strength of the received signal commonly known as RSSI, or received signal strength indicator. In addition, carrier sense can be used to determine if the channel is available . This technique is more selective since it verifies that the signal is the same carrier type as 802.11 transmitters. In all of our simulations, we use carrier sense and not RSSI to determine if the channel is busy. Thus, a Bluetooth signal will corrupt WLAN packets, but it will not cause the WLAN to defer transmission. A virtual carrier sense mechanism is also provided at the MAC layer. It uses the request-to-send (RTS) and clear-to-send (CTS) message exchange to make predictions of future traffic on the medium and updates the network allocation vector (NAV) available in stations. Communication is established when one of the wireless nodes sends a short RTS frame. The receiving station issues a CTS frame that echoes the sender's address. If the CTS frame is not received, it is assumed that a collision occurred and the RTS process starts over. Regardless of whether the virtual carrier sense routine is used or not, the MAC is required to implement a basic access procedure (depicted in figure 2) as follows. If a station has data to send, it waits for the channel to be idle through the use of the CSMA/CA algorithm. If the medium is sensed idle for a period greater than a DCF interframe space (DIFS), the station goes into a backoff procedure before it sends its frame. Upon the successful reception of a frame, the destination station returns an ACK frame after a Short interframe space (SIFS). The backoff window is based on a random value uniformly distributed in the interval [CW min , CW max ], where CW min and CW max represent the Contention Window parameters. If the medium is determined busy at any time during the backoff slot, the backoff procedure is suspended. It is resumed after the medium has been idle for the duration of the DIFS period. If an ACK is not received within an ACK timeout interval, the station assumes that either the data frame or the ACK was lost 204 GOLMIE ET AL. and needs to retransmit its data frame by repeating the basic access procedure. Errors are detected by checking the Frame Check Sequence (FCS) that is appended to the packet payload. In case an error is found, the packet is dropped and is then later retransmitted Integrated simulation model In this section, we describe the methodology and platform used to conduct the performance evaluation. The simulation environment consists of detailed models for the RF channel, the PHY, and MAC layers developed in C and OPNET (for the MAC layer). These detailed simulation models constitute an evaluation framework that is critical to studying the various intricate effects between the MAC and PHY layers. Although interference is typically associated with the RF channel modeling and measured at the PHY layer, it can significantly impact the performance of higher layer applications including the MAC layer. Similarly, changes in the behavior of the MAC layer protocol and the associated data traffic distribution can play an important factor in the interference scenario and affect the overall system performance. Figure 3 shows a packet being potentially corrupted by two interference packets. Consider that the desired packet is from the WLAN and the interference packets are Bluetooth (the figure is equally valid if the roles are reversed, except that the frequencies of the packets will be different). For interference to occur, the packets must overlap in both time and frequency. That is, the interference packets must be within the 22 MHz bandwidth of the WLAN. In a system with many Bluetooth piconets, there may be interference from more than one packet at any given time. We define a period of stationarity (POS) as the time during which the interference is constant . For example, t i t t i +1 is such a period, as is t i +1 t t i +2 . Even during a POS where there is one or more interferers, the number and location of bit errors in the desired packet depends on a number of factors: (1) the signal-to-interference ratio (SIR) and the signal-to-noise ratio at the receiver, (2) the type of modulation used by the transmitter and the interferer, and (3) the channel model. For this reason, it is essential to use accurate models of the PHY and channel, as described below. Just because two packets overlap in time and frequency does not necessary lead to bit errors and the consequent packet loss. While one can use (semi-)analytic models instead, such as approximating Bluetooth interference on WLAN as a narrowband tone jammer, the use of detailed signal processing-based models better allows one to handle multiple simultaneous interferers. In order to simulate the overall system, an interface module was created that allows the MAC models to use the physical layer and channel models. This interface module captures all changes in the channel state (mainly in the energy level). Consider the Bluetooth transmitterchannelreceiver chain of processes. For a given packet, the transmitter creates a set of Figure 3. Packet collision and placement of errors. The bit error rate (BER) is roughly constant during each of the three indicated periods. signal samples that are corrupted by the channel and input to the receiver; interference may be present for all or only specific periods of stationarity, as shown in figure 3. A similar chain of processing occurs for an 802.11b packet. The interface module is designed to process a packet at a time. At the end of each packet transmission, the MAC layer generates a data structure that contains all the information required to process the packet. This structure includes a list of all the interfering packets with their respective duration, timing offset, frequency, and transmitted power. The topology of the scenario is also included. The data structure is then passed to the physical layer along with a stream of bits representing the packet being transmitted. The physical layer returns the bit stream after placing the errors resulting from the interference . 3.1. MAC model We used OPNET to develop a simulation model for the Bluetooth and IEEE 802.11 protocols. For Bluetooth, we implemented the access protocol according to the specifications [2]. We assume that a connection is already established between the master and the slave and that the synchronization process is complete. The Bluetooth hopping pattern algorithm is implemented . Details of the algorithm are provided in section 2.1. A pseudo-random number generator is used instead of the implementation specific circuitry that uses the master's clock and 48-bit address to derive a random number. For the IEEE 802.11 protocol, we used the model available in the OPNET library and modified it to bypass the OPNET radio model and to use our MAC/PHY interface module. We focus in this study on the Direct Sequence mode which uses a fixed frequency that occupies 22 MHz of the frequency band. The center frequency is set to 2.437 GHz. At the MAC layer, a set of performance metrics are defined including probability of packet loss. Packet loss measures the number of packets discarded at the MAC layer due to errors in the bit stream. This measure is calculated after performing error correction. 3.2. PHY model The transmitters, channel, and receivers are implemented at complex baseband. For a given transmitter, inphase INTERFERENCE EVALUATION OF BLUETOOTH AND IEEE 802.11b SYSTEMS 205 and quadrature samples are generated at a sampling rate of 44 10 6 per second. This rate provides four samples/symbol for the 11 Mbit/s 802.11 mode, enough to implement a good receiver. It is also high enough to allow digital modulation of the Bluetooth signal to account for its frequency hopping . Specifically, since the Bluetooth signal is approximately 1 MHz wide, it can be modulated up to almost 22 MHz, which is more than enough to cover the 11 MHz bandwidth (one-sided) of the 802.11 signal. The received complex samples from both the desired transmitter and the interferer(s) are added together at the receiver. While there are a number of possible Bluetooth receiver designs, we chose to implement the noncoherent limiter-discriminator (LD) receiver [3,14]. Its simplicity and relatively low cost should make it the most common type for many consumer applications. Details of the actual design are given in [15]. In the 802.11b CCK receiver, each group of eight information bits chooses a sequence of eight consecutive chips that forms a symbol. As before, the inphase and quadrature components of these chips are transmitted. The receiver looks at the received symbol and decides which was the most likely transmitted one. While one can implement this decoding procedure by correlating against all 256 possible symbols, we chose a slightly sub-optimal, but considerably faster architecture similar to the WalshHadamard transform; again details can be found in [15]. 3.3. Channel model The channel model consists of a geometry-based propagation model for the signals, as well as a noise model. For the indoor channel, we apply a propagation model consisting of two parts: (1) line-of-sight propagation (free-space) for the first 8 m, and (2) a propagation exponent of 3.3 for distances over 8 m. Consequently, the path loss in dB is given by L p = 32.45 + 20 log(f d) if d &lt; 8 m, 58.3 + 33 log d8 otherwise, (1) where f is the frequency in GHz, and d is the distance in meters . This model is similar to the one used by Kamerman [10]. Assuming unit gain for the transmitter and receiver antennas and ignoring additional losses, the received power in dBmW is P R = P T - L p , (2) where P T is the transmitted power also in dBmW. Equation (2) is used for calculating the power received at a given point due to either a Bluetooth or an 802.11 transmitter, since this equation does not depend on the modulation method. The main parameter that drives the PHY layer performance is the signal-to-interference ratio between the desired signal and the interfering signal. This ratio is given in dB by SIR = P R - P I , (3) where P I is the interference power at the receiver. In the absence of interference, the bit error rate for either the Bluetooth or WLAN system is almost negligible for the transmitter powers and ranges under consideration. To complete the channel model, noise is added to the received samples, according to the specified SNR. In decibels, the signal-to-noise ratio is defined by SNR = P R -S R , where P R is the received signal power, and S R is the receiver's sensitivity in dBmW; this latter value is dependent on the receiver model and so is an input parameter. Additive white Gaussian noise (AWGN) is used to model the noise at the receivers. 3.4. Model validation The results obtained from the simulation models were validated against experimental and analytical results. Since the implementation of the PHY layer required choosing a number of design parameters, the first step in the validation process is comparing the PHY results against theoretical results. Complete BER curves of the Bluetooth and 802.11b systems are given in [15]; for the AWGN and flat Rician channels without interference, all the results match very closely to analytical bounds and other simulation results. Also, the simulation results for both the MAC and PHY models were compared and validated against analytical results for packet loss given different traffic scenarios [6]. For the experimental testing, we use the topology in figure 4 and compare the packet loss observed for Bluetooth voice and WLAN data with the simulation results in figure 5. The experimental and simulation results are in good agreement Simulation results We present simulation results to evaluate the performance of Bluetooth in the presence of WLAN interference and vice versa. First, we consider the effects of parameters such as transmitted power, offered load, hop rate, and traffic type on interference. Second, we look at two realistic interference scenarios to quantify the severity of the performance degradation for the Bluetooth and WLAN systems. 4.1. Factors effecting interference We first consider a four node topology consisting of two WLAN devices and two Bluetooth devices (one master and one slave) as shown in figure 4. The WLAN access point (AP) is located at (0, 15) m, and the WLAN mobile is fixed at ( 0, 1) m. The Bluetooth slave device is fixed at (0, 0) m and the master is fixed at (1, 0) m. In an effort to control the interference on Bluetooth and WLAN, we define two scenarios. In the first scenario, we let the mobile be the generator of 802.11 data, while the AP is the sink. In this case, the interference is from the mobile sending data packets to the AP and receiving acknowledgments (ACKs) from it. Since most of the WLAN traffic is 206 GOLMIE ET AL. Figure 4. Topology 1. Two WLAN devices and one Bluetooth piconet. Table 2 Summary of the scenarios. Scenario Desired Interferer WLAN WLAN signal signal AP mobile 1 Bluetooth WLAN Sink Source 2 WLAN Bluetooth Source Sink originating close to the Bluetooth piconet, both the master and the slave may suffer from serious interference. In the second scenario, the traffic is generated at the AP and received at the WLAN mobile. Because the data packets are generally longer then the ACKs, this is a more critical scenario for the WLAN then when the mobile is the source. Table 2 summarizes the two scenarios. For Bluetooth, we consider two types of applications, voice and data. For voice, we assume a symmetric stream of 64 Kbit/s each way using HV1 packet encapsulation. For data traffic, we consider a source that generates DM5 packets. The packet interarrival time is exponentially distributed, and its mean in seconds is computed according to t B = 2 n s T s , (4) where is the offered load; n s is the number of slots occupied by a packet. For DM5, n s = 5. T s is the slot size equal to 625 s. For WLAN, we use the 11 Mbit/s mode and consider a data application. Typical applications for WLAN could be ftp or http. However, since we are mainly interested in the MAC layer performance, we abstract the parameters for the application model to packet size and offered load and do not model the entire TCP/IP stack. We fix the packet payload to 12,000 bits which is the maximum size for the MAC payload data unit, and vary . The packet interarrival time in seconds, t W , is exponentially distributed, and its mean is computed acTable 3 Simulation parameters Simulation parameters Values Propagation delay 5 s/km Length of simulation run 30 s Bluetooth parameters ACL Baseband Packet Encapsulation DM5 SCO Baseband Packet Encapsulation HV1 Transmitted Power 1 mW WLAN parameters Transmitted power 25 mW Packet header 224 bits Packet payload 12,000 bits cording to t W = 192/1,000,000 + 12,224/11,000,000 , (5) where the 192-bit PLCP header is sent at 1 Mbit/s and the payload at 11 Mbit/s. Unless specified otherwise, we use the configuration and system parameters shown in table 3. For scenarios 1 and 2, we run 15 trials using a different random seed for each trial. In addition to plotting the mean value, confidence intervals, showing plus and minus two standard deviations, are also included. From figures 5 and 6, one sees that the statistical variation around the mean values are very small. In addition to the comparisons with analytical and experimental results described in section 3.4, this fact provides further validation for the results. 4.1.1. WLAN transmission power First, we look at the effect on Bluetooth of increasing the WLAN transmission power in scenario 1; that is, increasing the interferer transmission power on the victim signal. Since power control algorithms exist in many WLAN implementa-tions , it is important to consider how varying the transmitted power changes the interference. However, since Bluetooth was designed as a low power device, we fix its transmitter power at 1 mW for all simulations. We fix WLAN to 60% for different Bluetooth traffic types and values of . In figure 5(a), we note a saturation effect around 10 mW. A threshold, which is close to 22/79, corresponds to the probability that Bluetooth is hopping in the WLAN occupied band. Thus, increasing the WLAN transmission power beyond 10 mW does not affect the Bluetooth packet loss. Between 1 and 5 mW, a small change in the WLAN transmitted power triples the Bluetooth packet loss. Please note the relative positions of the packet loss curves for different values of between 1 and 5 mW; as increases, the packet loss is higher. Also, note that Bluetooth voice has the lowest packet loss, partly due to its short packet size. A second reason for the low loss probability is that voice packets are rejected only if there are errors in the access code or packet headers, cf. table 1. A packet may be accepted with a relatively large number of bit errors in the payload, which may lead to a substantial reduction in subjective voice quality. Figure 5(b) shows the probability of packet loss for the WLAN mobile device. This corresponds to ACKs being INTERFERENCE EVALUATION OF BLUETOOTH AND IEEE 802.11b SYSTEMS 207 (a) (b) (c) Figure 5. WLAN = 60%. (a) Scenario 1. Probability of packet loss for the Bluetooth slave. (b) Scenario 1. Probability of packet loss for the WLAN mobile. (c) Scenario 2. Probability of packet loss for the WLAN mobile. dropped at the WLAN source. The general trend is that the packet loss decreases as the WLAN transmitted power increases . However, we notice a slight "bump" between 1 and 5 mW. This is due to the effect of closed-loop interference. The WLAN source increases its transmitted power and causes more interference on the Bluetooth devices; as a result, there are more retransmissions in both the Bluetooth and WLAN piconets, which causes more lost ACKs at the WLAN source. Next, we consider the effect of increasing the WLAN transmission power on the WLAN performance in scenario 2. From figure 5(c), we observe that even if the WLAN transmission power is fifty times more than the Bluetooth transmission power (fixed at 1 mW), the packet loss for the WLAN does not change. This leads us to an interesting observation on power control. Basically, we note that increasing the transmission power does not necessarily improve the performance . However, decreasing the transmission power is usu-ally a "good neighbor" strategy that may help reduce the interference on other devices. 4.1.2. Offered load The offered load, also referred to in some cases as duty cycle , is an interesting parameter to track. Consider scenario 1 where Bluetooth is the interferer and fix the WLAN transmission power to 25 mW. We observe that for the WLAN, the packet loss is proportional to the Bluetooth offered load as shown in figure 6. For equal 20%, 50%, and 100%, the packet loss is 7%, 15%, and 25%, respectively. This observation has been confirmed analytically in [6], where the packet error is shown to depend not only on the offered load of the interferer system but also on the packet sizes of both systems. Also note that the probability of loss for the 30% WLAN of-208 GOLMIE ET AL. Figure 6. Scenario 2. Probability of packet loss for the WLAN mobile. fered load is slightly higher than for the 60% WLAN offered load. However, this difference is statistically insignificant. The significance of the packet size is apparent in figures 5(a) and (c), where short Bluetooth voice packets lead to less packet loss for Bluetooth but cause more interference for WLAN. However, for the WLAN 11 Mbit/s rate, the effect of changing the WLAN packet size over the range 1,000 to 12,000 bits has very little effect on the performance of both the WLAN and Bluetooth, and that is due to the relatively short transmission time of the WLAN packet. At the 1 Mbit/s rate, WLAN packets of the same bit lengths take considerably longer to transmit, and the effect of packet size is somewhat more pronounced. For a further discussion of the 1 Mbit/s case, please see [7]. 4.1.3. Bluetooth hop rate In order to highlight the effect of the Bluetooth hop rate on WLAN, we use different packet types, DM1, DM3, and DM5; these packets occupy 1, 3, and 5 time slots, respectively. The Bluetooth hop rate is determined by the number of time slots occupied by a packet. Thus, the hop rate is 1600, 533, and 320 hops/s for DM1, DM3, and DM5 packets, respectively. The offered load for Bluetooth is set to 100%. The results in table 4 clearly indicate that a faster hop rate leads to higher packet losses (44%, 28%, and 26% for DM1, DM3 and DM5, respectively). Note that the results are rather insensitive to the WLAN offered load. 4.1.4. Bluetooth traffic type The question here is, whether Bluetooth voice effects WLAN more than Bluetooth data, and vice versa. We use three types of packets for voice encapsulation, namely, HV1, HV2, and HV3. HV1 represents the worst case of interference for WLAN as shown in table 5 with 44% packet loss. HV2 and HV3, which contain less error correction and more user information , are sent less often and, therefore, interfer less with WLAN (25% and 16% for HV2 and HV3, respectively). The Table 4 Scenario 2. Probability of WLAN packet loss versus Bluetooth hop rate. BT WLAN = 30% WLAN = 60% DM1 0.449 0.449 DM3 0.286 0.277 DM5 0.269 0.248 Table 5 Scenario 2. Probability of WLAN packet loss versus Bluetooth traffic type. BT WLAN = 30% WLAN = 60% Voice HV1 0.446 0.470 HV2 0.253 0.257 HV3 0.166 0.169 Data, = 60% 0.191 0.177 Table 6 Scenario 1. Probability of Bluetooth packet loss versus Bluetooth traffic type. BT WLAN = 30% WLAN = 60% Voice HV1 0.077 0.141 HV2 0.075 0.149 HV3 0.069 0.136 Data, = 60% 0.2089 0.210 WLAN packet loss with Bluetooth data interference is 19%. Please note that the results do not depend on the WLAN offered load. On the other hand, the probability of packet loss for Bluetooth data (20%) is higher than for Bluetooth voice (7%) as shown in table 6. Note that doubling the WLAN offered load to 60% doubles the Bluetooth voice packet loss. Also, since all three types of voice packets suffer the same packet loss, it is preferable to use HV3, which causes less interference on the WLAN. The error correction coding in HV1 and HV2 packets may provide greater range in a noise-limited environment , but this coding is far too weak to protect the packets from interference. Instead, it is the frequency hopping ability of Bluetooth that limits the damage done by the WLAN. 4.1.5. Bluetooth transmission power While most Bluetooth devices will be operating at 1 mW, the specification also allows higher transmitter powers. Table 7 shows the probability of packet loss for both Bluetooth and the WLAN for three values of the BT transmitter power and two types of Bluetooth traffic. As expected, higher transmitter powers lead to more lost WLAN packets, regardless of the BT traffic type. Increasing the power from 1 to 10 mW leads to approximately a 50% increase in WLAN loss. Conversely, the Bluetooth packet error rate decreases. It still not clear how beneficial this decrease is for Bluetooth; even a loss probability of 0.0335 may lead to unacceptable voice quality. 4.1.6. Bluetooth packet error correction So far, the results shown for the Bluetooth data are with DM5 packets, which use a 2/3 block code on the packet payload. In order to show the effect of error correction on the probabil-INTERFERENCE EVALUATION OF BLUETOOTH AND IEEE 802.11b SYSTEMS 209 Figure 7. Scenario 1. Probability of packet loss for the Bluetooth slave. Table 7 Scenario 2. Probability of packet loss versus Bluetooth transmission power (mW). WLAN = 60%. BT traffic BT power BT loss WLAN loss (mW) probability probability = 60% 1 0.2125 0.0961 2.5 0.2085 0.1227 10 0.1733 0.1358 Voice 1 0.1417 0.1253 2.5 0.1179 0.1609 10 0.0335 0.1977 ity of packet loss, we repeat scenario 1 and compare the results given in figures 5(a) and 7, obtained with DM5 and DH5 packets, respectively. As expected, the probability of packet loss for DM5 packets (figure 5(a)) is slightly less than for DH5 packets (figure 7) for WLAN transmission powers less than 5 mW. Thus, for low levels of interference, a 2/3 block code can reduce the probability of loss by 4%. However, for WLAN transmission powers above 5 mW, the probability of packet loss is the same for both DM5 and DH5 packets. 4.2. Realistic interference topologies In this section, we consider two practical interference topologies . While they appear to be somewhat different, they actually complement each other. The first one has the WLAN device, in the midst of the Bluetooth piconets, acting at the source, while the second one has the WLAN access point acting as the source. 4.2.1. Topology 2 We first look at the topology illustrated in figure 8. It consists of one WLAN AP located at (0, 15) m, and one WLAN mobile at (0, 0) m. The WLAN traffic is generated at the mobile , while the AP returns acknowledgments. The distance between the WLAN AP and mobile is d W = 15 m. There Figure 8. Topology 2. Two WLAN devices and ten Bluetooth piconets. Table 8 Experiment 3 results. BT traffic WLAN BT loss WLAN loss d B = 1 m d B = 2 m = 30% 30% 0.056 0.157 0.121 60% 0.060 0.188 0.170 = 60% 30% 0.057 0.243 0.405 60% 0.061 0.247 0.381 Voice 30% 0.009 0.104 1 60% 0.008 0.106 1 are ten Bluetooth piconets randomly placed, covering a disk. The center of the disk is located at (0, 0) and its radius is r = 10 m. We define d B as the distance between a Bluetooth master and slave pair. d B = 1 m for half of the master and slave pairs, while d B = 2 m for the other half of the master and slave pairs. In this case, the main interference on Bluetooth is caused by the WLAN source located in the center of the disk; the aggregation of the ten piconets affects the WLAN source. We found that when the WLAN system is not operating, the Bluetooth packet loss is negligible (less than 1%). Table 8 gives the packet loss for the Bluetooth and WLAN devices. The packet loss for the Bluetooth devices is averaged over the master and slave devices and split into two groups: piconets with d B = 1 m and piconets with d B = 2 m. For WLAN, the packet loss is measured at the source. It is effectively zero at the sink. We observe that the WLAN packet loss depends on the Bluetooth traffic load value, . As is varied from 30% to 60%, the WLAN packet loss is significantly changed from 12% to 40%. However, the WLAN packet loss is insensitive to the WLAN offered load. Consistent with previous results, Bluetooth voice represents the worst case interference scenario for WLAN. 210 GOLMIE ET AL. In general, the Bluetooth packet loss for d B = 1 m is less than for d B = 2 m. The reason is that when the Bluetooth signal is stronger (over a shorter distance), the impact of interference is less significant. 4.2.2. Topology 3 We next consider the topology given in figure 9. It includes one WLAN AP and four WLAN mobile devices. The WLAN AP is located at (0, 15) m, and it is the source of the traffic generation. The four WLAN mobile devices are placed on a two-dimensional grid at ( -1, 1), (1, 1), (-1, -1), and ( 1, -1) m. In this topology, there are four Bluetooth piconets, each consisting of a masterslave device pair. The placement of the Bluetooth devices is as shown in the figure. In this case, we are looking at the effect of Bluetooth piconets on the four WLAN sink devices. The packet loss measure for WLAN is averaged over the four devices. As shown in table 9, the impact of WLAN interference on Bluetooth is minimal, given that the WLAN source is far from the Bluetooth piconets. As expected, the WLAN packet loss depends on the Bluetooth traffic conditions, and it is rather insensitive to the WLAN traffic activity. With Bluetooth voice, the WLAN packet loss is close to 84%. It is 57% for Bluetooth data with WLAN loads of = 30, 60%. Concluding remarks We presented results on the performance of Bluetooth and WLAN operating in the 2.4 GHz ISM band based on detailed channel, MAC, and PHY layer models for both systems. The evaluation framework used allows us to study the impact of interference in a closed loop environment where two systems are affecting each other, and explore the MAC and PHY layer interactions in each system. We are able to draw some useful conclusions based on our results. First, we note that power control may have limited benefits in this environment. Increasing the WLAN transmission power to even fifty times the power of Bluetooth is not sufficient to reduce the WLAN packet loss. On the other hand, limiting the WLAN power, may help avoid interference to Bluetooth. Second, using a slower hop rate for Bluetooth (i.e. longer packet sizes) may cause less interference to WLAN. Third, Bluetooth voice represents the worst type of interference for WLAN. In addition, the WLAN performance seems to degrade as the Bluetooth offered load is increased. Finally, the use of error correcting block codes in the Bluetooth payload does not improve performance. The errors caused by interference are often too many to correct. Overall, the results are dependent on the traffic distribution . Yet, there may be little room for parameter optimization especially for the practical scenarios. Not only does the complexity of the interactions and the number of parameters to adjust make the optimization problem intractable, but choosing an objective function is very dependent on the applications and the scenario. Thus, achieving acceptable performance for Figure 9. Topology 3. Five WLAN devices and four Bluetooth piconets. Table 9 Experiment 4 results. BT traffic WLAN BT loss WLAN loss = 30% 30% 0.007 0.574 60% 0.006 0.580 = 60% 30% 0.007 0.576 60% 0.006 0.580 Voice 30% 0.002 0.836 60% 0.001 0.828 a particular system comes at the expense of the other system's throughput. Therefore, we believe that the primary solutions to this problem lie in the development of coexistence mechanisms References [1] BlueHoc: Bluetooth Performance Evaluation Tool, Open-Source (2001) http://oss.software.ibm.com/developerworks/ opensource/~bluehoc [2] Bluetooth Special Interest Group, Specifications of the Bluetooth system , Vol. 1, v.1.0B Core, and Vol. 2, v1.0B Profiles (December 1999). [3] T. Ekvetchavit and Z. Zvonar, Performance of phase-locked loop receiver in digital FM systems, in: Ninth IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, Vol. 1 (1998) pp. 381385. [4] G. Ennis, Impact of Bluetooth on 802.11 direct sequence, IEEE P802.11 Working Group Contribution, IEEE P802.11-98/319 (September 1998). [5] D. Fumolari, Link performance of an embedded Bluetooth personal area network, in: Proceedings of IEEE ICC'01, Helsinki, Finland (June 2001). [6] N. Golmie and F. Mouveaux, Interference in the 2.4 GHz ISM band: Impact on the Bluetooth access control performance, in: Proceedings of IEEE ICC'01, Helsinki, Finland (June 2001). INTERFERENCE EVALUATION OF BLUETOOTH AND IEEE 802.11b SYSTEMS 211 [7] N. Golmie, R.E. Van Dyck, and A. Soltanian, Interference of Bluetooth and IEEE 802.11: Simulation modeling and performance evaluation , in: Proceedings of the Fourth ACM International Workshop on Modeling, Analysis, and Simulation of Wireless and Mobile Systems, MSWIM'01, Rome, Italy (July 2001). [8] I. Howitt, V. Mitter and J. Gutierrez, Empirical study for IEEE 802.11 and Bluetooth interoperability, in: Proceedings of IEEE Vehicular Technology Conference (VTC) (Spring 2001). [9] IEEE Standard 802-11, IEEE standard for wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specification (June 1997). [10] A. Kamerman, Coexistence between Bluetooth and IEEE 802.11 CCK: Solutions to avoid mutual interference, IEEE P802.11 Working Group Contribution, IEEE P802.11-00/162r0 (July 2000). [11] A. Kamerman and N. Erkocevic, Microwave oven interference on wireless LANs operating in the 2.4 GHz ISM band, in: Proceedings of the 8th IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, Vol. 3 (1997) pp. 12211227. [12] J. Lansford, A. Stephens and R. Nevo, Wi-Fi (802.11b) and Bluetooth: Enabling coexistence, IEEE Network Magazine (September/October 2001). [13] S. Shellhammer, Packet error rate of an IEEE 802.11 WLAN in the presence of Bluetooth, IEEE P802.15 Working Group Contribution, IEEE P802.15-00/133r0 (May 2000). [14] M.K. Simon and C.C. Wang, Differential versus limiter-discriminator detection of narrow-band FM, IEEE Transactions on Communications COM-31(11) (November 1983) 12271234. [15] A. Soltanian and R.E. Van Dyck, Physical layer performance for coexistence of Bluetooth and IEEE 802.11b, in: Virginia Tech. Symposium on Wireless Personal Communications (June 2001). [16] M. Takai, R. Bagrodia, A. Lee and M. Gerla, Impact of channel models on simulation of large scale wireless networks, in: Proceedings of ACM/IEEE MSWIM'99, Seattle, WA (August 1999). [17] S. Unawong, S. Miyamoto and N. Morinaga, Techniques to improve the performance of wireless LAN under ISM interference environments, in: Fifth Asia-Pacific Conference on Communications, 1999 and Fourth Optoelectronics and Communications Conference, Vol. 1 (1999) pp. 802805. [18] J. Zyren, Reliability of IEEE 802.11 WLANs in presence of Bluetooth radios, IEEE P802.11 Working Group Contribution, IEEE P802.15-99/073r0 (September 1999). [19] S. Zurbes, W. Stahl, K. Matheus and J. Haartsen, Radio network performance of Bluetooth, in: Proceedings of IEEE International Conference on Communications, ICC 2000, New Orleans, LA, Vol. 3 (June 2000) pp. 15631567. Nada Golmie received the M.S.E degree in computer engineering from Syracuse University, Syracuse , NY, in 1993, and the Ph.D. degree in computer science from University of Maryland, College Park, MD, in 2002. Since 1993, she has been a research engineer at the advanced networking technologies division at the National Institute of Standards and Technology (NIST). Her research in traffic management and flow control led to several papers presented at professional conferences, journals and numerous contributions to international standard organizations and industry led consor-tia . Her current work is focused on the performance evaluation of protocols for Wireless Personal Area Networks. Her research interests include modeling and performance analysis of network protocols, media access control, and Quality of Service for IP and wireless network technologies. She is the vice-chair of the IEEE 802.15 Coexistence Task Group. E-mail: nada.golmie@nist.gov Robert E. Van Dyck received the B.E and M.E.E degrees from Stevens Institute of Technology, Hoboken , NJ, in 1985 and 1986, respectively, and the Ph.D. degree in electrical engineering from the North Carolina State University at Raleigh in 1992. Since June 2000, he has been a member of the Advanced Network Technologies Division of the National Institute of Standards and Technology, Gaithersburg, MD. Prior to that, he was an Assistant Professor in the Department of Electrical Engineering, the Pennsylvania State University, University Park, PA. During 1999, he was a Summer Faculty Research Fellow at Rome Laboratory. His other previous affiliations include GEC-Marconi Electronic Systems, Wayne, NJ (19951996), the Center for Computer Aids for Industrial Productivity, Rutgers University, Piscataway, NJ (19921995), the Computer Science Corporation, Research Triangle Park NC, (1989), and the Communications Laboratory, Raytheon Co., Marlborough, MA (19851988). His present research interests are in self-organization of sensor networks, multimedia communications and networking , and source and channel coding for wireless communications. Amir Soltanian received his M.S. degree from Sharif University of Technology, Tehran, Iran, in 1994. He has been working in the industry for 6 years doing research on GSM receivers. Currently, he is a guest researcher at National Institute of Standards and Technology. His current research is the study of the interference cancellation methods for the physical layer of the Bluetooth and IEEE802.11 WLAN. Arnaud Tonnerre is a graduate student at the cole Nationale Suprieure des Telecommunications (ENST) in Bretagne, France. He is currently doing an internship at the National Institute of Standards and Technology (NIST) in Gaithersburg, MD. He will receive the Diplome d'Ingenieur in June 2003. His research interests are in wireless personal area networks. Olivier Rbala received a computer science degree from the Institut suprieur d'informatique, de modlisation et de leurs applications (ISIMA) in Clermont-Ferrand, France, in September 2001. He is currently a Guest Researcher at the National Institute of Standards and Technology (NIST) in the advanced networking technologies division. His research interests includes the performance evaluation of wireless networking protocols.

evaluation;packet loss;performance degradation;IEEE 802.11b;simulation framework;Bluetooth;interference;hop rate;tranmission power;topology;WPANs;WLAN;offered load

119

Is a Picture Worth a Thousand Words?

What makes a peripheral or ambient display more effective at presenting awareness information than another ? Presently, little is known in this regard and techniques for evaluating these types of displays are just beginning to be developed. In this article, we focus on one aspect of a peripheral display's effectiveness-its ability to communicate information at a glance. We conducted an evaluation of the InfoCanvas, a peripheral display that conveys awareness information graphically as a form of information art, by assessing how well people recall information when it is presented for a brief period of time. We compare performance of the InfoCanvas to two other electronic information displays , a Web portal style and a text-based display, when each display was viewed for a short period of time. We found that participants noted and recalled significantly more information when presented by the InfoCanvas than by either of the other displays despite having to learn the additional graphical representations employed by the InfoCanvas.

Introduction The Peripheral awareness displays are systems that reside in a user's environment within the periphery of the user's attention. As such, the purpose of these displays is not for monitoring vital tasks. Rather, peripheral displays best serve as communication media that people can opportunistically examine to maintain information awareness [11, 17]. The term ambient display [22] has been used to describe systems like this as well, but to avoid confusion, throughout this document we use this term to describe peripheral awareness systems that generally convey only one piece of information. We use the term peripheral display to describe peripheral awareness systems that may present multiple information items. Both peripheral and ambient displays are designed not to distract people from their tasks at hand, but to be subtle, calm reminders that can be occasionally noticed. In addition to presenting information, the displays also frequently contribute to the aesthetics of the locale in which they are deployed [1]. Dozens of peripheral/ambient displays have been created in many shapes and form factors. Some displays , such as the dangling string [21], tangible displays including water lamps and pinwheels [4], and the Information Percolator [7] have utilized physical (and often everyday) objects. Other displays, such as Informative Artwork [8] and the Digital Family Portrait [16] use electronic displays to represent information in a graphical manner. All these systems primarily communicate one item of information. Other peripheral/ambient displays exist that are capable of conveying more than one information item simultaneously. The Digital Family Portrait, although primarily intended to allow geographically separated family members maintain awareness of each other, allows for the optional displaying of additional information such as weather [16]. Audio cues, instead of visual displays, have also been utilized in peripheral displays to convey multiple nuggets of information in the Audio Aura system [15]. The Kandinsky system [5] attempts to create artistic collages of various pieces of information , and the Scope system is an abstract visualization displaying notification information from multiple sources [19]. SideShow [3] provides a display sidebar containing multiple awareness icons such as traffic and weather indicators. The InfoCanvas [14], the focus of this article, differs from the initial set of systems above by explicitly promoting the conveyance of multiple pieces of information concurrently. It differs from the latter set of Copyright is held by the author/owner originally published by the Canadian Human-Computer Communications Society in the Proceedings of Graphics Interface 2004, May 17-19, London, Ontario. 117 systems in promoting greater flexibility of information monitored and its subsequent visual representation, as well as allowing for greater user control in specifying those mappings. Although many types of displays exist and new ones are being developed, little is known about what makes a particular peripheral/ambient display more successful at presenting information than another [10]. Furthermore, such displays are inherently difficult to evaluate formally since they are designed not to distract the user. As a result, evaluation techniques have been limited, as Mankoff et al. note [10], to formative ethnographies [16] and within-lab studies where displays are developed and subsequently refined over time by their designers [6]. However, there has been recent work on developing new evaluation techniques for ambient displays , most notably Mankoff et al.'s set of discount formative techniques [10] and McCrickard et al.'s notification system categorization framework [13]. The goal of this study is not to evaluate peripheral displays in general. Rather, we focus on one particular component of a peripheral display's effectiveness, its ability to communicate information. More specifically, we examine how the abstract data mappings of electronic information artwork affect people's interpretation and memory of the data. Both the InfoCanvas [14] and the Informative Artwork [8] projects make use of dynamic pieces of electronic artwork to represent information in an eye-appealing manner. Such displays are placed within a person's work environment or are publicly displayed, enabling at-a-glance information awareness. How well the systems convey information is not known, however. Note that the success of a peripheral/ambient display involves more than simple information acquisition. Because these displays are positioned in people's environments , aesthetics and attractiveness influence adop-tion as well. The research reported here, though, focuses solely on such displays' ability to convey information . In a companion study, the issues of aesthetics and longer-term use of the InfoCanvas system are currently being explored. Experimental Design This study examines if an electronic picture "is worth a thousand words." That is, how well are users able to learn mappings and subsequently comprehend and recall information when it is presented in the form of electronic artwork in comparison to more traditional methods. We accomplish this by designing an InfoCanvas display as well as two more conventional information displays and evaluating participants' memories of them when they only see the displays for short periods of time. Study participants viewed three examples of each display with each example encoding different data values (described in detail in the next section). After viewing a display for eight seconds, participants recalled the information presented using a multiple-choice questionnaire. 2.1 Materials Ten items of information were selected to be monitored : time of day, a weather forecast, a temperature forecast, traffic conditions, a news headline, the Dow Jones stock index value, an airfare price, updates to a Web site, a count of new emails, and a baseball score. These items are examples of information people typically seek to maintain awareness of [14]. Three information screens were designed including an InfoCanvas beach scene, a minimalist text-based display, and a Web portal-like display. These three displays were chosen to represent interesting points in a spectrum of possibilities, as depicted in Figure 1, for representing awareness information on electronic ambient displays. Styles range from pure textual presentations to highly abstract, graphical imagery. The InfoCanvas and the Text-based display inhabit positions near the endpoints of that spectrum. The Web Portal display was designed to incorporate a hybrid of textural and graphical representations, and resemble the types of Web "start pages" that people frequently use to maintain information awareness today [14]. Other interesting points in the spectrum include more direct graphical (typically iconic) representations of information as embodied by systems such as Sideshow [3], and could be the subject of future experiments . For this study, we compare the InfoCanvas to two widely deployed types, Web portals (e.g. MyYahoo !) and text-heavy news summaries or Web pages. Highly Textual InfoCanvas [14] Sideshow [3] Web Portal My Yahoo! Text-Based Informative Artwork [6] Highly Graphical Figure 1: A spectrum of awareness displays ranging from textual to graphical presentations of information. 118 All three displays in the study were designed seeking a balance of experimental control and representation of ecologically valid real-world use. Extensive pilot testing and redesign was used to refine their appearance . We designed the three displays to encode the ten pieces of information in an appropriate manner for that display style. In all three, we added a small amount of extra information beyond the ten queried information values, much as similar real world displays would undoubtedly do. All displays were presented full-screen on a Viewsonic 15" LCD display running at a resolution of 1024 x 768. The InfoCanvas used the entire screen area, and the other two displays used slightly less of the entire display as will be explained below. In the following subsections, we describe each of the displays in more detail. InfoCanvas Display The InfoCanvas system supports a variety of artistic scenes or themes. We chose to use a beach scene as shown in Figure 2 for the experiment due to its popularity with trial users. Individual objects in the scene represented the ten data values as follows: Airfare price: Represented by the vertical height of the kite in the sky from $0 (near the water level) to $400 (top of the screen). News headline: Shown on the banner behind the plane. Time of day: Denoted by the sailboat moving from the left side (12:01 AM) to the right side (11:59 PM). Web site update: Represented by the color of the leaves on the palm tree, green indicates a recent update and brown indicates no recent changes. Weather forecast: Illustrated through the actual weather shown in the sky (e.g., clouds represents a forecast of cloudy weather). Temperature forecast: Represented by the height of the large seagull in the sky, ranging from 50 degrees at water level to 90 degrees at the top of the screen. Dow Jones stock market change: Displayed by the arrangement of seashells on the shoreline. Shells form an arrow to indicate whether stocks are up or down and the quantity of shells indicates the value (three shells indicate a change of 0 50 points, five shells indicate a change of more than 50 points). New email messages: Depicted by the height of liquid in the glass ranging from 0 new emails (empty glass) to 20 new emails (a full glass). Current traffic speed on a local roadway: Sym-bolized by the color of the woman's bathing suit with red indicating speed less than 25 MPH, yellow indicating a speed between 25 and 50 MPH, and green indicating a speed greater than 50 MPH. Baseball score: Shown by the size of two beach balls: A larger ball indicates a winning team and identical ball sizes indicate a tied score. Color is used to distinguish the two teams. These mappings were chosen to reflect a variety of objects moving or changing size or color. In addition, some mappings were chosen for being more intuitive and direct, such as using weather icons to represent weather or the metaphor of a kite flying in the sky to reflect airfare price. Other mappings, such as representing updates to a Website by tree leaf color, were intended to be more abstract and indirect. A pilot study of four InfoCanvas users revealed a wide variety of mapping styles, both natural and abstract. As a result, we wanted the scene used in this study to reflect this. Furthermore, as also done in actual use, we placed additional items in the scene such as the chair, umbrella, and crab simply for aesthetic purposes. Several items within this display present information as a precise point along a continuous scale, including the time-of-day, airfare, and forecasted temperature, by displaying objects that move along a line. Other items, including the traffic speed, stock update, and baseball score, are represented using categorical encod-ings . For example, the different shell arrangements representing the Dow Jones stock update indicate four different ranges of values. The implications of this difference will be explored more fully later when describing the questionnaire formats. Text-Based Display The Text-based display (shown in Figure 2) predomi-nantly uses text to display information. Web pages such as MyYahoo were the inspiration for the Text-based display, but the use of images, different colors, and graphics were removed. Thus, the display represents a position near the endpoint of the graphics-text spectrum presented earlier. As a result, we restricted formatting on this display to changes in point size and the use of bold text with the exception of using a fixed-width font to indicate stock change values. (The fixed-width font helps to align numerical stock values, providing a clean and orderly appearance similar to the style used by existing services.) Extra information beyond the ten data values on this display included a few lines from a news article related to the current headline, the current date, and additional stock information for the Standard & Poor's 500 and 119 Figure 2: Examples of the InfoCanvas beach scene (top), text-based (middle), and Web Portal displays (bottom) used in the study. 120 NASDAQ indices--items likely to appear on such a display. The Text-based display consisted of a region 970 pixels wide to 330 pixels high on the screen. Pilot testing found this size optimal in allowing the use of columns , section headers, and white space to make an effective and visually pleasing display. Furthermore, pilot testing indicated that information recall suffered as the display's size increased, perhaps due to increased eye movement, even though the data elements remained located in the same position. Web Portal Display The Web Portal display (shown in Figure 2) also mim-icked the look and feel of popular no-cost "start" Web pages such as My Yahoo. However, we added additional formatting and iconic graphics/images as found in awareness displays such as Sideshow [3] to differentiate this display from the Text-based display. Web portals, in actuality, tend to make relatively limited use of images and graphics. Our introduction of graphics and images served two main purposes--making the display more of a hybrid between the highly artistic InfoCanvas and a display utilizing only text, and also to increase the effectiveness of the design by using graphics to position items or convey information. Graphics that encode values--those that change to reflect information--in the Web Portal display include the weather icon indicating the weather forecast, the speedometer icon with a meter indicating the current speed of vehicles, and an icon indicating the presence of new email messages. In addition, an image related to the news headline was displayed. Iconic images that did not change and were used solely as positional anchors included a picture frame icon for the Web site update item, baseball team logos, and an airline logo. In addition, colors and arrows were used to indicate stock trends and the baseball team currently winning was displayed in bold text. The Web Portal display's extra information (e.g. not encoding the ten queried values) included a few lines of a news story related to the headline and the current date, and the two other stock indices as done in the Text-based display. The Web Portal display used an area of 968 pixels wide by 386 pixels high on the display. Again, iterative development and pilot testing helped determine this size was best to create a balanced and ordered layout and be an effective presenter of information. As in the Text-based display, each element on the Web Portal display remained in the same relative position. Design Considerations As noted above, wherever we faced a design choice in creating the Web Portal and Text-based displays, we attempted to optimize the display to promote comprehension . For example, both the Web Portal and Text-based displays represent substantial improvements over real-life examples. The Web Portal design contained more graphics and images than what typically appears on these Web pages. Pilot subjects found these graphics and images to be beneficial in remembering information . Furthermore, individual items were modified during pilot testing to assist recall. For example, we made the size of the weather forecast image substan-tially larger than what is typically found on Web portals . Likewise, we designed the Text-based display to be a substantial improvement over existing text-based information displays, such as tickers or small desktop window applications, by introducing columns, section headers, and white space. Initial full-screen presentations used for the Web Portal and Text-based display tended to look unwieldy and resulted in lower recall of information during pilot testing. We attributed this to the larger screen area that participants had to visually parse. Hence, we reduced the screen area occupied by those displays to promote comprehension. Following that logic, InfoCanvas' larger size should have served to negatively impact its performance, if anything. 2.2 Participants Forty-nine (11 female) individuals with normal or corrected -to-normal eyesight participated in this study. Participants ranged from 18 to 61 years of age (mean 24.2). 27 were graduate students, 17 were undergraduates , and 5 were non-students. Participants were com-pensated $10 for their time. 2.3 Procedure Testing occurred in individual sessions lasting approximately 45 minutes. Participants sat two feet in front of the LCD monitor. The keyboard and mouse were removed from the area, leaving empty desk space between the participant and the display. The experi-menter informed participants that they were participating in a study to determine how much they could remember from different information screens when they could only see the screen for a brief amount of time. A within-subjects experimental design was used and the ordering of the display conditions was counterbal-anced . Participants were randomly assigned to an ordering sequence. For each of the three displays, an in-121 troductory tour, preparation task, and practice task were given prior to performing three actual trials. The introductory session included an explanation of the display and the information found on it. For the InfoCanvas and Web Portal displays, the behaviors of the elements on the displays were also explained. Due to the display's more complex and dynamic nature, the introductory tour took longer to perform with the InfoCanvas , approximately 3.5 minutes in duration, than with the Web Portal and Text-based display, both approximately 1.5 minutes in duration. Initially, especially with InfoCanvas, we had concerns that the introductory tour might not be sufficient to allow participants to learn each display. Pilot testing, however, revealed that participants were able to quickly learn the information mappings. To further ensure that we would be testing information comprehension and recall but not mapping recall with respect to the InfoCanvas , participants were asked to point out the different objects on a sample display and say aloud what information each object represented. We also provided participants with a reference sheet labeling the mappings between information and objects on the InfoCanvas . In practice, we found that participants seldom looked at the sheet and some actually turned it over. During the preparation task, participants were shown an example display and instructed to complete a sample recall questionnaire (explained in more detail later in this section), much as they would in the actual trials. In this phase, however, no time limit was en-forced for viewing the display. This task then allowed the participant to better familiarize him or herself with the display, the questionnaire style, and to ask additional questions regarding the display, all while it was visible. Next, in the practice task, participants were exposed to what the actual trials would be like. A recall questionnaire was placed text side down in front of the participant and then an information display was shown for eight seconds. Pilot testing determined that this was a suitable amount of exposure time to avoid ceiling or floor effects, with recall averaging about five or six items. Furthermore, participants during pilot testing felt that this amount of time was indicative of the duration of a glance of a person seeking multiple information updates. Upon completion of the exposure, the computer prompted the individual to turn over and complete the recall sheet. Participants were instructed to not guess on the recall questionnaire; if the participant did not remember an item at all, he or she left that item blank on the questionnaire. The actual trials followed the practice task and consisted of three exposure and recall activities involving different data sets and hence data displays. Again, specific emphasis was made to discourage the participant from guessing on the recall. The same data values were used for each position of the nine total experiment trials independent of the display ordering, ensuring a balance across the experiment. Upon completion of the three different display conditions , participants were given several concluding surveys that captured subjective feedback from the participants regarding perceived performance and display preferences. Recall Task Ten questions, one per each information item, were presented to participants after exposure to an information display. We varied the question topic order across trials to discourage participants from becoming accustomed to a particular topic being the subject of the first few questions and then seeking out information from the displays on those topics. While participants were not explicitly informed of this, the varied order came as no surprise when they performed actual trials since they had already encountered the recall sheet in the preparation and practice tasks. To minimize cognitive load, the questions were designed to elicit the comprehension and recall of information in the same manner that it had been encoded. For all questions about the Text and Web Portal displays , and for the majority of questions about the InfoCanvas display, the question style was multiple-choice, typically including four exact-value answers spread relatively evenly across the range of possible answers. For instance, the potential answers for the time of day might have been 3:42am, 8:36am, 5:09pm, and 10:11pm. The newspaper headline question used four possible answers containing some similarity (usually using the same key words such as "Iraq" or "President Bush") to ensure the recall of the headline by context, not by recognition of a key word. The Web site update question simply asked whether the site had been up-dated , with yes and no as the possible answers. Finally, the baseball score question asked which team was currently winning and offered the choices of the Braves, Pirates, or tied game. The data values used to generate displays for the nine trials also were chosen to range across the possible set of values. "Exact Value" "Categorical" What is the status of the Dow Jones? What is the status of the Dow Jones? + 89 points + 42 points - 2 points - 75 points Up over 50 points Up 0 50 points Down 0 50 points Down over 50 points Figure 3: Example of exact value and categorical recall questions. 122 For topics that the InfoCanvas presented categories or ranges of values (e.g., traffic conditions, baseball score, and stock updates), answer choices to the recall questions were also presented in the form of ranges. Figure 3 shows an example of how these differed using stocks as an example. Note how the exact-value answers lie within the intervals used; the questions and answers were designed to be as similar as possible. Furthermore, we felt that the more general issue of participants needing to translate pictures into exact, usually numeric, values would counter any benefit received by the InfoCanvas in using ranges for a few questions. Adjacent to each multiple-choice question on the recall questionnaire was a confidence level scale with choices for high, medium, or low confidence. Participants were instructed to indicate their relative confidence for each item. We did this to further lessen the "guessing factor" and identify whether confidence would play a measure. Following the nine cumulative trials for all three displays, participants completed a Likert scale survey rating all the displays for facilitating the recall of information , being an effective presenter of information, and visual appeal. In addition, participants rank-ordered each display for facilitating recall and visual appeal. Lastly, participants responded to open-ended questions regarding which display they would employ at their workstation or on a wall if a dedicated display would be available. Results Table 1 presents the means and standard deviations across all conditions of the raw number of correct responses for each of the three trials under each display. A repeated measures ANOVA identified an overall effect of the display for accurately recalled items, F(2,96) = 22.21, MSE = 2.31, p &lt; .0001, and there was no effect for order. Additionally, pair-wise comparisons between display types found an advantage of the InfoCanvas display over the Web Portal, F(1,48) = 14.65, MSE = 2.66, p &lt; .0005), the Web Portal over the Text-based display, F(1,48) = 8.17, MSE = 1.76, p &lt; .007), and the InfoCanvas over the Text-based display, F(1,48) = 40.01, MSE =2.51, p &lt; .0001). To take into account participants' confidence of their answers, a second method to evaluate performance was developed. Weights of value 3, 2, and 1 were assigned for the high, medium, and low confidence levels, respectively (e.g. a correct answer with medium confidence yielded +2 points, while an incorrect answer also with a medium confidence yielded 2 points). Questions not answered on the recall task were assigned a weighted score value of 0. Participants forgot to assign a confidence on 13 of the 4410 responses collected in the study. Since this number of accidental omissions was quite low, items with omitted confidence ratings were assigned a medium level, the median of the obtainable point values. Of the 13 questions with omitted confidence, 3 were answered incorrectly. In examining the weighted scores shown in Table 2, an overall effect was found on the display, F(2,96) = 10.40, MSE = 25.35, p &lt; .001, and again there was no effect of order. Furthermore, pair-wise comparisons between the displays again found an advantage of the InfoCanvas display over the Web Portal, F(1,48) = 7.29, MSE = 30.56, p = .0095, and of the InfoCanvas display over the Text-based display, F(1,48) = 22.21, MSE = 22.93, p &lt; .0001. However, the weighted scores gave no advantage of the Web Portal over the Text-based display, F(1,48) = 2.59, MSE = 2.51, p = 0.11. Figure 3 presents an item-by-item breakdown of the percentage of correctly answered questions for each display. The InfoCanvas had the highest average on Ease of Info. Recall 1 2 3 4 5 Mean Text-Based 7 18 14 10 0 2.6 (1.0) Web Portal 1 8 18 17 5 3.3 (0.9) InfoCanvas 2 4 13 20 10 3.7 (1.0) Effective Data Pres. 1 2 3 4 5 Mean Text-Based 6 18 16 7 2 2.6 (1.0) Web Portal 2 3 14 24 6 3.6 (0.9) InfoCanvas 5 9 13 18 4 3.1 (1.1) Visual Appeal 1 2 3 4 5 Mean Text-Based 20 19 8 2 0 1.8 (0.9) Web Portal 1 2 12 22 12 3.9 (0.9) InfoCanvas 1 1 10 17 20 4.1 (0.9) Table 3: Likert scale responses for display characteristics , with 1 = low rating and 5 = high rating. 1st Trial 2nd Trial 3rd Trial Text-Based 5.14 (1.59) 5.12 (1.33) 5.02 (1.57) Web Portal 5.67 (1.61) 5.65 (1.54) 5.29 (1.89) InfoCanvas 6.27 (1.80) 6.22 (1.79) 6.31 (1.76) Table 1: Means and standard deviations of correct responses for three trials of each display. 1st Trial 2nd Trial 3rd Trial Text-Based 11.47 (4.92) 11.78 (4.81) 10.57 (5.02) Web Portal 12.88 (5.09) 12.35 (5.84) 11.27 (6.40) InfoCanvas 13.88 (5.96) 14.02 (5.89) 13.82 (6.63) Table 2. Means and standard deviations of correct responses for weighted scores for three trials of each display 123 seven of the ten items. The Web Portal score was higher on the time and baseball items, and the Text display was best for the airfare price. Table 3 contains a breakdown of participants' Likert ratings captured during the post-experiment surveys. These results mirror the performance data with the InfoCanvas generally being rated higher with the exception that participants generally ranked the Web Portal higher as being a more effective presenter of data. Participants' order rankings of the three displays for facilitation of recall and personal preference are shown in Table 4. Here, the Text-based display fared poorly along both dimensions. More participants preferred the Web Portal but rated the InfoCanvas as best for recall. Discussion Participants in the study recalled information best using the InfoCanvas display despite having the greater cognitive load of remembering mappings and representations used in the art paradigm. This cognitive load also includes translating pictorial InfoCanvas objects to the values used in the recall questions, while the two other displays presented data values more closely to the format of the questions. Even with these disadvantages, the InfoCanvas conveyed information better and was more vividly recalled. Another possible interpretation is that the InfoCanvas system actually reduces the cognitive load of the individuals. In this scenario, it follows that it is easier and cognitively more efficient to remember and recall the InfoCanvas images, and then translate later to the values desired. Regardless of their cognitive interpretation, the study's results should not be too surprising. People are able to process images rapidly by leveraging the sophisticated , parallel, and high-bandwidth nature of the perceptual system [20]. Umanath and Scamell showed that graphics are conducive towards recall tasks involving simple fact retrieval in a series of studies investigating the role of recall in real-time decision-making [18]. Furthermore, "ecological" layouts with objects in natural positions have been shown to facilitate faster browsing [2]. This study, however, confirms our intuition that the InfoCanvas, and displays like it, has potential to be an effective peripheral display where people seek to obtain information at a glance. Several interesting observations emerged from the results of this study. We noted that participants generally expressed preference for the Web Portal display over the InfoCanvas display even though they felt that the InfoCanvas display had best facilitated the recall of information. When asked about this preference, one participant remarked that the Web Portal design was "more professional looking" and "more common than the other two." Other participants praised the Web Portal for its ability to display information in a more "logi-cal and precise" manner and providing "accurate information that is not influenced by my interpretation." These comments seem to imply a conservative attitude about adopting a new and unconventional technology such as an ambient display Other participants appeared to capture the essence of peripheral/ambient displays and their abilities to be subtle communication channels, not distracting a user. One participant remarked that, "I think I could choose to ignore it [InfoCanvas] while I was working. I think once I got used to what all the icons meant and what the scales were, I could easily look at it to see the information I was interested in." Others also echoed this sentiment : "[InfoCanvas] is the quickest and easiest to see at a glance the information you want" and "[InfoCanvas ] is informative but also relaxing." Finally, one participant summarized the benefits of the InfoCanvas as being "able to keep working and not get distracted by details; [InfoCanvas is] faster to see and interpret from a distance." In the context of this study, the InfoCanvas was evaluated on its abilities as an information purveyor. The mappings between information and graphical elements used in this study were designed by the authors, and as such, did not always feel instinctive to participants . Some participants indicated they had difficulty in learning the mappings; one participant remarked that "I struggled with the visual mappings" and another felt that InfoCanvas was "counterintuitive." As was mentioned earlier, this was a concern in the design of the 0% 20% 40% 60% 80% 100% Web Site Update Weather Forecast Traffic Conditions Time of Day Stock Updates News Headline New Email Forecasted Temp Baseball Score Airfare Text Web InfoCanvas Figure 3. Mean percentage for correctly recalled items for each display type 124 study--would individuals even be able to learn these mappings in such a short period of time? Pilot studies and the final study data both indicated that despite not being able to define their own mappings for the information , participants were able to recall more information when presented on the InfoCanvas. A crucial implication lies in this; the InfoCanvas is designed to be a highly personalized peripheral display where users specify their own mappings and layouts. Since participants were able to recall information quite well when they did not specify the mappings, it seems logical to conclude that comprehension and recall would benefit even more when people design their own display and it is constantly present in their environment. Several interesting discussion points arise from the breakdown of correctly recalled items shown in Figure 3. First, note that on the whole, the InfoCanvas yielded the largest percentage of correctly recalled items per category, with the exception of the airfare, time of day, and baseball score items. However, performance of the three displays on the baseball score item was comparable , averaging a recall rate of 64-70%. In regards to the airfare and time of day items, the InfoCanvas produced the second best percentage of correctly recalled items and was outperformed by the Text and Web Portal displays , respectively. Slightly lower performance was somewhat expected with these two items, since their representations moved along a straight line to indicate a point on a scale. Pilot participants often remarked that these representations were more difficult to keep track of since they could be found in different areas. Interestingly , even with these representations, the InfoCanvas performed better than the Web Portal (for the airfare item) and the Text display (for the time item), indicating that despite their moving nature, graphical representations still worked relatively well. The temperature element, also represented by a moving object, illustrates this point as well, generating a higher recall than the other displays. Interestingly, the InfoCanvas appeared to have the largest advantage over the other two displays with the traffic conditions item. While some may argue that this is due to the use of intervals to represent conditions, as opposed to the exact-value representations on the Web Portal and Text displays, note that the use of intervals for the baseball score did not yield such an effect. This difference implies that the representation used to indicate traffic conditions--the color of the woman's bathing suit--provided an excellent mapping. Therefore, we speculate that if individuals create their own mappings , leveraging their personal experiences, recall with InfoCanvas will benefit even more. This study examined the information conveyance abilities of three specific examples of displays involving a sample population consisting mainly of academic-related , relatively young individuals. Generalizing its findings too much would be unwise. Nevertheless, we speculate that the results would extend to other similar types of displays and people of different demographics. The lessons learned from this study could be applied to the design of new information systems. For example , in designing a system using a docked PDA as an information display, a graphical representation of information , such as using a miniature InfoCanvas, might convey information more effectively than a traditional text-based manner. Conclusion and Future Work In this paper, we present a formal evaluation of information recall from three different electronic information displays, the InfoCanvas, a Web Portal-like, and a Text-based display. We present results indicating that participants comprehended and recalled more awareness information when it was represented in graphical manners; participants recalled more information from the InfoCanvas display than the Web Portal and Text-based displays. Likewise, participants recalled more information from the Web Portal display than the Text-based display. Our results suggest that there are benefits for comprehension, when a person may only glance at a display for a short period of time, by displaying information in a highly graphical or stylized nature. A number of potential directions for follow-on work exist. It would be interesting to compare a more abstract graphical presentation of information as embodied by the InfoCanvas with a purely graphical, but more direct iconic encoding, such as in Sideshow [3]. In this study, we positioned the information displays directly in front of participants. Another possible experiment could position the display further away, perhaps on a neighboring wall, from the person's main computer display. Yet another possibility is to introduce an explicit primary task thus making information comprehension more truly peripheral. For instance, participants could perform a primary task such as document editing while information is presented for comprehension and recall on a display in another location as done in several other studies [9,12]. Text-based Web Portal InfoCanvas Best Recall Facilitator 2 (4%) 16 (33%) 31 (63%) Worst Recall Facilitator 41 (84%) 5 (10%) 3 (6%) Most Preferred 2 (4%) 35 (71%) 12 (25%) Least Preferred 35 (71%) 2 (4%) 12 (25%) Table 4: Rankings of displays for facilitating recall and personal preference. 125 Acknowledgements This research has been supported in part by a grant from the National Science Foundation, IIS-0118685 and the first author's NDSEG Graduate Fellowship The authors would like to express gratitude to Richard Catrambone and Mary Czerwinski for providing valuable insights into the development and analysis of this study

evaluation;peripheral display;graphical representation;awareness information;ambient display;text-based display;information conveyance;InfoCanvas display;Peripheral display;information recall;empirical evaluation;information visualization;Web portal-like display

12

A Geometric Constraint Library for 3D Graphical Applications

Recent computer technologies have enabled fast high-quality 3D graphics on personal computers, and also have made the development of 3D graphical applications easier. However , most of such technologies do not sufficiently support layout and behavior aspects of 3D graphics. Geometric constraints are, in general, a powerful tool for specifying layouts and behaviors of graphical objects, and have been applied to 2D graphical user interfaces and specialized 3D graphics packages. In this paper, we present Chorus3D, a geometric constraint library for 3D graphical applications. It enables programmers to use geometric constraints for various purposes such as geometric layout, constrained dragging, and inverse kinematics. Its novel feature is to handle scene graphs by processing coordinate transformations in geometric constraint satisfaction. We demonstrate the usefulness of Chorus3D by presenting sample constraint-based 3D graphical applications.

INTRODUCTION Recent advances in commodity hardware have enabled fast high-quality 3D graphics on personal computers. Also, software technologies such as VRML and Java 3D have made the development of 3D graphical applications easier. However, most of such technologies mainly focus on rendering aspects of 3D graphics, and do not sufficiently support layout and behavior aspects. Constraints are, in general, a powerful tool for specifying layouts and behaviors of graphical objects. It is widely recognized that constraints facilitate describing geometric layouts and behaviors of diagrams in 2D graphical user interfaces such as drawing editors, and therefore constraint solvers for this purpose have been extensively studied [3, 7, Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to distribute to lists, requires prior specific permission and/or fee. Int. Symp. on Smart Graphics, June 11-13, 2002, Hawthorne, NY, USA. Copyright 2002 ACM 1-58113-555-6/02/0600... $ 5.00 8, 9, 11, 12, 13, 17, 18]. Also, many specialized 3D graphics packages enable the specification of object layouts and behaviors by using constraints or similar functions. It is natural to consider that various 3D graphical applications can also be enhanced by incorporating constraints. It might seem sufficient for this purpose to modify existing 2D geometric constraint solvers to support 3D geometry. It is, however, insufficient in reality because of the essential difference between the ways of specifying 2D and 3D graphics; typical 2D graphics handles only simple coordinate systems, whereas most 3D graphics requires multiple coordinate systems with complex relations such as rotations to treat scene graphs. It means that we need to additionally support coordinate transformations in 3D geometric constraint solvers. In this paper, we present Chorus3D, a geometric constraint library for 3D graphical applications. The novel feature of Chorus3D is to handle scene graphs by processing coordinate transformations in geometric constraint satisfaction. We have realized Chorus3D by adding this feature to our previous 2D geometric constraint library Chorus [13]. Another important point of Chorus3D is that it inherits from Chorus the capability to handle "soft" constraints with hierarchical strengths or preferences (i.e., constraint hierarchies [7]), which are useful for specifying default layouts and behaviors of graphical objects. It determines solutions so that they satisfy as many strong constraints as possible, leaving weaker inconsistent constraints unsatisfied. Chorus3D also inherits from Chorus a module mechanism which allows user-defined kinds of geometric constraints. This feature enables programmers to use geometric constraints for various purposes including the following: Geometric layout: A typical use of Chorus3D is to lay out graphical objects. For example, it allows putting objects parallel or perpendicular to others without requiring predetermined positioning parameters. Also, it provides constraint-based general graph layout based on the spring model [14]. Constrained dragging: Chorus3D enables dragging objects with positioning constraints. For example, it can constrain a dragged object to be on the surface of a sphere. Constrained dragging is important for 3D graphics because it provides a sophisticated way to ac-94 commodate ordinary mouse dragging to 3D spaces. Inverse kinematics: Chorus3D is applicable to inverse kinematics, which is a problem of finding desired configurations of "articulated" objects [1, 20]. It allows the specification of articulated objects by using coordinate transformations, and can automatically calculate the parameters of the transformations that satisfy constraints. This method is also applicable to camera control by aiming at a possibly moving target object. In this paper, we demonstrate the usefulness of Chorus3D by presenting sample constraint-based 3D graphical applications . This paper is organized as follows: We first present our approach to the use of constraints for 3D graphics. Second, we describe our basic framework of constraints. Next, we present a method for processing coordinate transformations in our framework. We then provide the implementation of Chorus3D, and demonstrate examples of using constraints in 3D graphics. After giving related work and discussion, we mention the conclusions and future work of this research. OUR APPROACH In this research, we integrate geometric constraints with 3D graphics. Basically, we realize this by extending our previous 2D geometric constraint solver Chorus [13] to support 3D geometry. However, as already mentioned, it is not a straightforward task because 3D graphics typically requires handling scene graphs with hierarchical structures of coordinate systems, which is not covered by the 2D version of the Chorus constraint solver. To support hierarchies of coordinate systems, we introduce the following new model of constraints: Point variables: Each point variable (which consists of three real-valued constrainable variables) is associated with one coordinate system, and its value is expressed as local coordinates. Geometric constraints: Geometric constraints on point variables are evaluated by using the world coordinates of the point variables (they can also refer to 1D variables for, e.g., distances and angles by using their values directly). A single constraint can refer to point variables belonging to different coordinate systems. Coordinate transformations: Parameters of coordinate transformations are provided as constrainable variables , and the solver is allowed to change the parameters of transformations to appropriately satisfy given constraints. With this model, we can gain the benefit of the easy maintenance of geometric relations by using constraints, as well as the convenience of modeling geometric objects by employing scene graphs. In our actual implementation, we provide the following three elemental kinds of coordinate transformations: Translation: A translation transformation is characterized with three variables t x , t y , and t z , and specifies the translation of vector (t x , t y , t z ). Rotation: A rotation transformation is parameterized with four variables r x , r y , r z , and r w , and specifies the rotation of angle r w about the axis (r x , r y , r z ). Scale: A scale transformation is represented with three variables s x , s y , and s z , and specifies the axis-wise scale (s x , s y , s z ) about the origin. We can express many practically useful transformations by using such elemental ones. In fact, any transformations represented with Transform nodes in VRML can be realized by combining these kinds of transformations [4]. CONSTRAINT FRAMEWORK In this section, we briefly describe our framework for handling constraints. We base it on the framework for the 2D version of the Chorus constraint solver. See [13] for further detail. 3.1 Problem Formulation We first present the mathematical formulation for modeling constraints and constraint systems. In the following, we write x to represent a variable vector (x 1 , x 2 , . . . , x n ) of n variables, and also v to indicate a variable value vector (v 1 , v 2 , . . . , v n ) of n real numbers (v i expresses the value of x i ). To support various geometric constraints in a uniform manner , we adopt error functions as a means of expressing constraints . An error function e(x) is typically associated with a single arithmetic constraint, and is defined as a function from variable value vectors to errors expressed as non-negative real numbers; that is, e(v) gives the error of the associated constraint for v. An error function returns a zero if and only if the constraint is exactly satisfied. For example, e(x) = (x i - x j ) 2 can be used for the constraint x i = x j . We assume that, for each e(x), its gradient is known: e(x) = e(x) x 1 , e(x) x 2 , . . . , e(x) x n . In the same way as constraint hierarchies [7], constraint systems in our framework can be divided into levels consisting of constraints with equal strengths. Constraints with the strongest preference are said to be required (or hard), and are guaranteed to be always satisfied (if it is impossible, there will be no solution). By contrast, constraints with weaker preferences are said to be preferential (or soft), and may be relaxed if they conflict with stronger constraints. Solutions to constraint systems are defined as follows: let e i,j (x) be the error function of the j-th constraint (1 j m i ) at strength level i (0 i l); then solutions v are determined with the optimization problem minimize v E(v) subject to e 0,j (v) = 0 (1 j m 0 ) 95 where E is an objective function defined as E(x) = l i=1 m i j=1 w i e i,j (x) in which w i indicates the weight associated with strength i, and the relation w 1 w 2 w l holds. In this formulation , level 0 corresponds to required constraints, and the others to preferential ones. Intuitively, more weighted (or stronger) preferential constraints should be more satisfied. Our framework simulates constraint hierarchies. Particularly , if the squares of constraint violations are used to compute error functions, a system in our framework will obtain approximate solutions to the similar hierarchy solved with the criterion least-squares-better [3, 17]. The largest difference is that a system in our framework slightly considers a weak constraint inconsistent with a stronger satisfiable one in computing its solutions, while the similar hierarchy would discard such a weak one. Our actual implementation of the Chorus3D constraint solver provides four external strengths required, strong, medium, and weak as well as two internal strengths very strong (used to approximately handle required nonlinear or inequality constraints) and very weak (exploited to make new solutions as close to previous ones as possible). It typically assigns weights 32 4 , 32 3 , 32 2 , 32 1 , and 1 to strengths very strong, strong, medium, weak, and very weak respectively . These weights were determined according to the precision of the actual numerical algorithm (described in the next subsection). To know how much these weights affect solutions, suppose a system of strong constraint x = 0 and medium one x = 100. Then the unique solution will be obtained as x = 3.0303 (= 100/33). Thus the difference of strengths is obvious. According to our actual experience, this precision allows us to discriminate constraint strengths in most graphical applications. 3.2 Algorithm To actually find solutions to constraint systems presented above, we need to solve their corresponding optimization problems. For this purpose, we designed a constraint satisfaction algorithm by combining a numerical optimization technique with a genetic algorithm. It uses numerical optimization to find local solutions, while it adopts a genetic algorithm to search for global solutions. For numerical optimization, we mainly use the quasi-Newton method based on Broyden-Fletcher-Goldfarb-Sahnno updating formula [2, 6], which is a fast iterative technique that exhibits superlinear convergence. Since it excludes fruitless searches by utilizing its history, it is usually faster than straightforward Newton's method. We introduced a genetic algorithm to alleviate the problem that some kinds of geometric constraints suffer from local optimal but global non-optimal solutions [11, 16]. Generally, a genetic algorithm is a stochastic search method that repeatedly transforms a population of potential solutions into another next-generation population [10, 15]. We typically necessitate it only for computing initial solutions; in other words, we can usually re-solve modified constraint systems without the genetic algorithm, only by applying numerical optimization to previous solutions. PROCESSING COORDINATE TRANSFORMATIONS In this section, we propose a method for integrating coordinate transformations with our constraint framework. As already mentioned, we use world coordinates of points to evaluate 3D geometric constraints. A naive method for this is to duplicate point variables in all ancestor coordinate systems, and then to impose required constraints that represent coordinate transformations between the point variables . However, this method requires an optimization routine supporting required nonlinear constraints, which limits the availability of actual techniques (in fact, we cannot use the quasi-Newton method for this purpose). Also, this method tends to yield many variables and constraints, and therefore requires an extra amount of memory. Below we propose a more widely applicable method for handling coordinate transformations. Its characteristic is to hide transformations from optimization routines, which is realized by embedding transformations in error functions. 4.1 Model To begin with, we introduce another variable vector x = (x 1 , x 2 , . . . , x n ), which is created by replacing variables for local coordinates of 3D points in x with the corresponding ones for world coordinates (1D variables remain the same). We can mathematically model this process as follows: Consider the sequence of the s transformations y 0 (= x) t 0 - y 1 t 1 - t s-2 - y s-1 t s-1 - y s (= x ) where y 0 and y s are equal to x and x respectively, each y k (1 k s - 1) is an "intermediate" vector, and each t k (0 k s - 1) is a function that transforms y k into y k+1 . Intuitively, t k corresponds to a coordinate transformation, and transforms related point variables from its source coordinate system into its destination system. It should be noted that, although transformations are, in general, hierarchical (or tree-structured), we can always find such a linear sequence by "serializing" them in an appropriate order. By using such transformations, we can compute x as follows : x = t s-1 (t s-2 ( (t 1 (t 0 (x))) )) t(x) where t is defined as the composition of all the elemental transformations. In the following description, we write y k,i to denote the i-th element of y k , and also t k,i to represent the i-th element of t k ; that is, y k+1 = (y k+1,1 , y k+1,2 , . . . , y k+1,n ) = (t k,1 (y k ), t k,2 (y k ), . . . , t k,n (y k )) = t k (y k ). 4.2 Method Geometric constraints are evaluated by using world coordinates of points, which means that their error functions are 96 defined as e(x ). Using the composed transformations, we can evaluate them as e(x ) = e(t(x)). Importantly, we can efficiently realize this computation by applying only necessary transformations to actually used variables. We also need to compute the gradient of e(t(x)), i.e., e(t(x)) = e(t(x)) x 1 , e(t(x)) x 2 , . . . , e(t(x)) x n . Basically, we can decompose each partial derivative e(t(x))/x i into primitive expressions by repeatedly using the chain rule. However, we should avoid the simple application of the chain rule since it would result in a large number of expressions. Instead, we perform a controlled way of decomposing such partial derivatives; it appropriately arranges the chain rule to restrict the computation to only necessary components. First, we decompose e(t(x))/x i as follows: e(t(x)) x i = j e(x ) x j t s-1,j (y s-1 ) x i = j e(x ) x j j s-1 t s-1,j (y s-1 ) y s-1,j s-1 t s-2,j s-1 (y s-2 ) x i = j s-1 j e(x ) x j t s-1,j (y s-1 ) y s-1,j s-1 t s-2,j s-1 (y s-2 ) x i = j s-1 e(x ) y s-1,j s-1 t s-2,j s-1 (y s-2 ) x i . Note that each e(x )/x j is given by the definition of the geometric constraint, and also that each t s-1,j (y s-1 )/y s-1,j s-1 is a partial derivative in the gradient of a single coordinate transformation t s-1 . Thus we can obtain each e(x )/y s-1,j s-1 . Also, by repeating this process, we can compute, for each k, e(t(x)) x i = j k e(x ) y k,j k t k-1,j k (y k-1 ) x i and finally achieve e(t(x)) x i = j 1 e(x ) y 1,j 1 t 0,j 1 (x) x i where each t 0,j 1 (x)/x i is a component of the gradient of t 0 . Therefore, e(t(x))/x i is now determined. Furthermore, we can considerably reduce the number of the computations of e(x )/y k,j k in practice. We can make the following observations about the above computation: For each variable x j , e(x )/x j can be non-zero only if x j is actually needed to evaluate the designated constraint . If x i is originated in the coordinate system associated with t k (that is, x i is either a local coordinate or a parameter of the coordinate transformation), we have y k,i = x i , which means that we have t k,j (y k )/x i . Therefore, we can compute e(x )/x i immediately. These observations reveal that we need to transfer a partial derivative e(x )/y k,j to the next step only when x j represents a really necessary coordinate that has not reached its local coordinate system. Also, since we can handle each necessary point independently, we can implement this process with a linear recursive function that hands over only three derivatives e(x )/y k,j at each recursive call. IMPLEMENTATION We implemented the proposed method by developing a constraint solver called Chorus3D, which is a 3D extension to our previous 2D geometric constraint solver Chorus [13]. We constructed Chorus3D as a C++ class library, and also developed a native method interface to make it available to Java programs. Chorus3D allows programmers to add a new kind of arithmetic constraints (e.g., Euclidean geometric constraints) by constructing a new constraint class with a method that evaluates their error functions. Also, programmers can introduce a new kind of non-arithmetic (or pseudo) constraints (for, e.g., general graph layout) by developing a new evalua-tion module which computes an "aggregate" error function for a given set of constraints. Chorus3D currently provides linear equality, linear inequality , edit (update a variable value), stay (fix a variable value), Euclidean geometric constraints (for, e.g., parallelism, per-pendicularity , and distance equality), and graph layout constraints based on the spring model [14]. Linear equality/ inequality constraints can refer to only 1D variables (including elements of 3D point variables), while edit and stay constraints can be associated with 1D and 3D point variables. Euclidean geometric constraints typically refer to point variables although they sometimes require 1D variables for angles and distances. Each graph layout constraint represents a graph edge, and refers to two point variables as its associated graph nodes. As stated earlier, constraints on such point variables are evaluated by using world coordinates of the points. Also, a single constraint can refer to point variables belonging to different coordinate systems. The application programming interface of Chorus3D is a natural extension to that of Chorus, which provides a certain compatibility with a recent linear solver called Cassowary [3]; in a similar way to Cassowary and Chorus, Chorus3D allows programmers to process constraint systems by creating variables and constraints as objects, and by adding/ removing constraint objects to/from the solver object. In addition, Chorus3D handles coordinate transformations as objects, and presents an interface for arranging them hier-archically EXAMPLES In this section, we present three examples to demonstrate how to incorporate geometric constraints into 3D graphics by using the Chorus3D constraint solver. All the examples are implemented in Java by using Java 3D as a graphics 97 Figure 1: A 3D geometric layout of a general graph structure. programming interface as well as the native method interface with Chorus3D. We also provide computation times taken for constraint satisfaction in these examples. 6.1 Graph Layout The first example is an application which lays out a set of points with a general graph structure in a 3D space as shown in Figure 1. This application also allows a user to drag graph nodes with a mouse. 1 The used graph layout technique is based on a 3D extension to the spring model [14]. This kind of 3D graph layout is practically useful to information visualization, and has actually been adopted in a certain system [19]. The constraint system of this graph layout consists of 26 point variables (i.e., 78 real-valued variables), 31 graph layout constraints, and three linear equality constraints for fixing one of the point variables at the origin. When executed on an 866 MHz Pentium III processor running Linux 2.2.16, Chorus3D obtained an initial solution in 456 milliseconds. It performed constraint satisfaction typically within 250 milliseconds to reflect the user's dragging a graph node. 6.2 Constrained Dragging The second example is an application which allows a user to drag an object constrained to be on another spherical object. Figure 2 depicts this application, where the smaller solid spherical object is constrained to be on the surface of the larger wireframe one. The application declares a strong Euclidean geometric constraint which specifies a constant distance between the centers of these objects. When the user tries to drag the smaller object with a mouse, the application imposes another medium Euclidean constraint which collinearly locates the viewpoint, the 3D position of the mouse cursor (which is considered to be on the screen), and 1 Unlike constrained dragging in the next example, this mouse operation is simply implemented with Java 3D's PickMouseBehavior classes. Figure 2: Dragging an object constrained to be on a sphere. Viewpoint Mouse cursor which is on the screen Screen Object which is on the sphere surface Collinearity constraint Distance constraint Sphere Figure 3: Implementation of constrained dragging. the center of the dragged object as shown in Figure 3. This collinearity constraint reflects the motion of the mouse in the position of the dragged object. Since the collinearity constraint is weaker than the first Euclidean constraint, the user cannot drag the smaller object to the outside of the larger sphere. The application initially declares one Euclidean geometric constraint on two point variables, and solved it in 1 millisecond on the same computer as the first example. When the user tries to drag the smaller object, it adds another Euclidean constraint as well as two edit constraints for the viewpoint and mouse position. The solver maintained this constraint system usually within 2 milliseconds. 6.3 Inverse Kinematics The final example applies inverse kinematics to a virtual robot arm by using constraints. Unlike the previous examples , it takes advantage of coordinate transformations to express its constraint system. 98 (a) (b) (c) (d) (e) (f ) Figure 4: A robot arm application which performs inverse kinematics. As illustrated in Figure 4(a), the robot arm consists of four parts called a base, a shoulder, an upper arm, and a forearm. Constraint satisfaction for inverse kinematics is performed to position its hand (the end of the forearm) at the target object if possible, or otherwise to make it maximally close to the target. Figures 4(b)(f) show the movement of the robot arm. In Figures 4(b)(e), its hand is positioned at the exact location of the target by using appropriate angles of its joints. By contrast, in Figure 4(f), the hand cannot reach the target, and therefore the arm is extended toward the target instead. Figure 5 describes the constraint program used in the robot arm application. After constructing a constraint solver s, it creates six coordinate transformations shldrTTfm, shldrRTfm, uarmTTfm, uarmRTfm, farmTTfm, and farmRTfm. Here the rotation angle parameters of the rotation transformations shldrRTfm, uarmRTfm, and farmRTfm will actually work as variables that can be altered by the solver. Next, it generates a point variable handPos to represent the position of the hand, and then suggests the target position to the hand by using a preferential edit constraint editHandPos. Finally, executing the solver, it obtains the desired angles shldrAngle, uarmAngle, and farmAngle of the rotation transformations. These angles will be passed to the Java 3D library to render the properly configured robot arm. This program generates a constraint system which contains three translation and three rotation transformations, one explicit point variable as well as six point variables and three 1D variables for coordinate transformations, and one edit constraint. The solver found an initial solution to this system in 18 milliseconds, and obtained each new solution for a frame update typically within 10 milliseconds. RELATED WORK AND DISCUSSION There has been work on integrating constraints or similar functions with 3D graphics languages to facilitate the specification of graphical objects. For example, we can view the event routing mechanism in VRML [4] as a limited form of one-way propagation constraints. Also, there is an attempt to extend VRML by introducing one-way propagation and finite-domain combinatorial constraints [5]. However, they cannot handle more powerful simultaneous nonlinear constraints such as Euclidean geometric constraints. Although many constraint solvers have been developed in 99 // constraint solver s = new C3Solver(); // translation transformation for the shoulder: fixed to (0, .1, 0) shldrTTfm = new C3TranslateTransform(new C3Domain3D(0, .1, 0)); s.add(shldrTTfm); // shldrTTfm is parented by the world coordinate system // rotation transformation for the shoulder: axis fixed to (0, 1, 0); angle ranging over [-10000, 10000] shldrRTfm = new C3RotateTransform(new C3Domain3D(0, 1, 0), new C3Domain(-10000, 10000)); s.add(shldrRTfm, shldrTTfm); // shldrRTfm is parented by shldrTTfm // translation transformation for the upper arm: fixed to (0, .1, 0) uarmTTfm = new C3TranslateTransform(new C3Domain3D(0, .1, 0)); s.add(uarmTTfm, shldrRTfm); // uarmTTfm is parented by shldrRTfm // rotation transformation for the upper arm: axis fixed to (0, 0, 1); angle ranging over [-1.57, 1.57] uarmRTfm = new C3RotateTransform(new C3Domain3D(0, 0, 1), new C3Domain(-1.57, 1.57)); s.add(uarmRTfm, uarmTTfm); // uarmRTfm is parented by uarmTTfm // translation transformation for the forearm: fixed to (0, .5, 0) farmTTfm = new C3TranslateTransform(new C3Domain3D(0, .5, 0)); s.add(farmTTfm, uarmRTfm); // farmTTfm is parented by uarmRTfm // rotation transformation for the forearm: axis fixed to (0, 0, 1); angle ranging over [-3.14, 0] farmRTfm = new C3RotateTransform(new C3Domain3D(0, 0, 1), new C3Domain(-3.14, 0)); s.add(farmRTfm, farmTTfm); // farmRTfm is parented by farmTTfm // variable for the hand's position, associated with farmRTfm and fixed to (0, .5, 0) handPos = new C3Variable3D(farmRTfm, new C3Domain3D(0, .5, 0)); // medium-strength edit constraint for the hand's position editHandPos = new C3EditConstraint(handPos, C3.MEDIUM); s.add(editHandPos); // suggest the hand being located at the target's position editHandPos.set(getTargetWorldCoordinates()); // solve the constraint system s.solve(); // get solutions double shldrAngle = shldrRTfm.rotationAngle().value(); double uarmAngle = uarmRTfm.rotationAngle().value(); double farmAngle = farmRTfm.rotationAngle().value(); Figure 5: Constraint program for the robot arm application. the field of graphical user interfaces [3, 7, 11, 12, 13, 17, 18], most of them do not provide special treatment for 3D graphics . In general, the role of nonlinear geometric constraints is more important in 3D applications than in 2D interfaces. Most importantly, 3D graphics usually requires rotations of objects which are rarely used in 2D interfaces. The main reason is that we often equally treat all "horizontal" directions in a 3D space even if we may clearly distinguish them from "vertical" directions. Therefore, nonlinear constraint solvers are appropriate for 3D applications. In addition, coordinate transformations should be supported since they are typically used to handle rotations of objects. Gleicher proposed the differential approach [8, 9], which supports 3D geometric constraints and coordinate transformations . In a sense, it shares a motivation with Chorus3D; in addition to support for 3D graphics, it allows user-defined kinds of geometric constraints. However, it is based on a different solution method from Chorus3D; it realizes constraint satisfaction by running virtual dynamic simulations. This difference results in a quite different behavior of solutions as well as an interface for controlling solutions. By contrast, Chorus3D provides a much more compatible interface with recent successful solvers such as Cassowary [3]. Much research on inverse kinematics has been conducted in the fields of computer graphics and robotics [1, 20]. However , inverse kinematics is typically implemented as specialized software which only provides limited kinds of geometric constraints. Chorus3D has two limitations in its algorithm: one is on the precision of solutions determined by preferential constraints; the other is on the speed of the satisfaction of large constraint systems. These limitations are mainly caused by the treatment of multi-level preferences of constraints in addition to required constraints (i.e., constraint hierarchies). Although many numerical optimization techniques have been proposed and implemented in the field of mathematical programming [2, 6], most of them do not handle preferential constraints. To alleviate the limitations of Chorus3D, we are pursuing a more sophisticated method for processing multi-level preferential constraints. We implemented Chorus3D as a class library which can be exploited in C++ and Java programs. However, more high-level authoring tools will also be useful for declarative approaches to 3D design. One possible direction is to extend VRML [4] to support geometric constraints. Standard VRML requires scripts in Java or JavaScript to realize complex layouts and behaviors. By contrast, constraint-enabled VRML will cover a wider range of applications without such additional scripts. CONCLUSIONS AND FUTURE WORK In this paper, we presented Chorus3D, a geometric constraint library for 3D graphical applications. It enables programmers to use geometric constraints for various purposes 100 such as geometric layout, constrained dragging, and inverse kinematics. Its novel feature is to handle scene graphs by processing coordinate transformations in geometric constraint satisfaction. Our future work includes the development of other kinds of geometric constraints to further prove the usefulness of our approach. In particular, we are planning to implement non-overlapping constraints [13] in Chorus3D so that we can use it for the collision resolution of graphical objects. Another future direction is to improve Chorus3D in the scalability and accuracy of constraint satisfaction. REFERENCES [1] Badler, N. I., Phillips, C. B., and Webber, B. L. Simulating Humans: Computer Graphics, Animation, and Control. Oxford University Press, Oxford, 1993. [2] Bertsekas, D. P. Nonlinear Programming, 2nd ed. Athena Scientific, 1999. [3] Borning, A., Marriott, K., Stuckey, P., and Xiao, Y. Solving linear arithmetic constraints for user interface applications. In Proc. ACM UIST , 1997, 8796. [4] Carey, R., Bell, G., and Marrin, C. The Virtual Reality Modeling Language (VRML97). ISO/IEC 14772-1:1997, The VRML Consortium Inc., 1997. [5] Diehl, S., and Keller, J. VRML with constraints. In Proc. Web3D-VRML, ACM, 2000, 8186. [6] Fletcher, R. Practical Methods of Optimization, 2nd ed. John Wiley & Sons, 1987. [7] Freeman-Benson, B. N., Maloney, J., and Borning, A. An incremental constraint solver. Commun. ACM 33, 1 (1990), 5463. [8] Gleicher, M. A graphical toolkit based on differential constraints. In Proc. ACM UIST , 1993, 109120. [9] Gleicher, M. A differential approach to graphical manipulation (Ph.D. thesis). Tech. Rep. CMU-CS-94-217, Sch. Comput. Sci. Carnegie Mellon Univ., 1994. [10] Herrera, F., Lozano, M., and Verdegay, J. L. Tackling real-coded genetic algorithms: Operators and tools for behavioural analysis. Artif. Intell. Rev. 12, 4 (1998), 265319. [11] Heydon, A., and Nelson, G. The Juno-2 constraint-based drawing editor. Research Report 131a, Digital Systems Research Center, 1994. [12] Hosobe, H. A scalable linear constraint solver for user interface construction. In Principles and Practice of Constraint Programming--CP2000 , vol. 1894 of LNCS, Springer, 2000, 218232. [13] Hosobe, H. A modular geometric constraint solver for user interface applications. In Proc. ACM UIST , 2001, 91100. [14] Kamada, T., and Kawai, S. An algorithm for drawing general undirected graphs. Inf. Process. Lett. 31, 1 (1989), 715. [15] Kitano, H., Ed. Genetic Algorithms. Sangyo-Tosho, 1993. In Japanese. [16] Kramer, G. A. A geometric constraint engine. Artif. Intell. 58, 13 (1992), 327360. [17] Marriott, K., Chok, S. S., and Finlay, A. A tableau based constraint solving toolkit for interactive graphical applications. In Principles and Practice of Constraint Programming--CP98 , vol. 1520 of LNCS, Springer, 1998, 340354. [18] Sannella, M. Skyblue: A multi-way local propagation constraint solver for user interface construction. In Proc. ACM UIST , 1994, 137146. [19] Takahashi, S. Visualizing constraints in visualization rules. In Proc. CP2000 Workshop on Analysis and Visualization of Constraint Programs and Solvers, 2000. [20] Zhao, J., and Badler, N. I. Inverse kinematics positioning using nonlinear programming for highly articulated figures. ACM Trans. Gr. 13, 4 (1994), 313336. 101

layout;scene graphs;3D graphics;geometric layout;constraint satisfaction;3D graphical applications;geometric constraints;graphical objects;behaviors;coordinate transformation

120

KDDCS: A Load-Balanced In-Network Data-Centric Storage Scheme for Sensor Networks

We propose an In-Network Data-Centric Storage (INDCS) scheme for answering ad-hoc queries in sensor networks. Previously proposed In-Network Storage (INS) schemes suffered from Storage Hot-Spots that are formed if either the sensors' locations are not uniformly distributed over the coverage area, or the distribution of sensor readings is not uniform over the range of possible reading values. Our K-D tree based Data-Centric Storage (KDDCS) scheme maintains the invariant that the storage of events is distributed reasonably uniformly among the sensors. KDDCS is composed of a set of distributed algorithms whose running time is within a poly-log factor of the diameter of the network. The number of messages any sensor has to send, as well as the bits in those messages, is poly-logarithmic in the number of sensors. Load balancing in KDDCS is based on defining and distributively solving a theoretical problem that we call the Weighted Split Median problem . In addition to analytical bounds on KDDCS individual algorithms , we provide experimental evidence of our scheme's general efficiency, as well as its ability to avoid the formation of storage hot-spots of various sizes, unlike all previous INDCS schemes.

INTRODUCTION Sensor networks provide us with the means of effectively monitoring and interacting with the physical world. As an illustrative example of the type of sensor network application that concerns us here, consider an emergency/disaster scenario where sensors are deployed in the area of the disaster [17]. It is the responsibility of the sensor network to sense and store events of potential interest. An event is composed of one or more attributes (e.g. temperature, carbon monoxide level, etc.), the identity of the sensor that sensed the event, and the time when the event was sensed. As first responders move through the disaster area with hand-held devices, they issue queries about recent events in the network. For example, the first responder might ask for the location of all sensor nodes that recorded high carbon monoxide levels in the last 15 minutes, or he might ask whether any sensor node detected movement in the last minute. Queries are picked up by sensors in the region of the first responder. The sensor network is then responsible for answering these queries. The first responders use these query answers to make decisions on how to best manage the emergency. The ad-hoc queries of the first responders will generally be multi-dimensional range queries [9], that is, the queries concern sensor readings that were sensed over a small time window in the near past and that fall in a given range of the attribute values. In-Network Storage (INS) is a storage technique that has been specifically presented to efficiently process this type of queries. INS involves storing events locally in the sensor nodes. Storage may be in-network because it is more efficient than shipping all the data (i.e., raw sensor readings) out of the network (for example to base stations), or simply because no out-of-network storage is available. All INS schemes already presented in literature were Data-Centric Storage (DCS) schemes [15]. In any In-Network Data-Centric Storage (INDCS ) scheme, there exists a function from events to sensors that maps each event to an owner sensor based on the value of the attributes of that event. The owner sensor will be responsible for storing this event. The owner may be different than the sensor that originally generated the event. To date, the Distributed Index for Multi-dimensional data (DIM) scheme [9] has been shown to exhibit the best performance among all proposed INDCS schemes in dealing with sensor networks whose query loads are basically composed of ad-hoc queries . In DIM [9], the events-to-sensors mapping is based on a K-D tree [3], where the leaves R form a partition of the coverage area, and each element of R contains either zero or one sensor. The formation of the K-D tree consists of rounds. Initially, R is a one element set containing the whole coverage area. In each odd/even round r, each region R R that contains more than one sensor is bisected horizontally/vertically. Each time that a region is split, each sensor in that region has a bit appended to its address specifying which side of the split the sensor was on. Thus, the length of a sensor's address (bit-code) is its depth in the underlying K-D tree. When a sensor generates an event, it maps such event to a binary code based on a repetitive fixed uniform splitting of the attributes' ranges in a round robin fashion. For our purposes, it is sufficient for now to 317 consider the cases that the event consists of only one attribute, say temperature. Then, the high order bits of the temperature are used to determine a root-to-leaf path in the K-D tree, and if there is a sensor in the region of the leaf, then this sensor is the owner of this event. Due to the regularity of regions in this K-D tree, the routing of an event from the generating sensor to the owner sensor is particularly easy using Greedy Perimeter Stateless Routing (GPSR) [6]. Full description of DIM is presented in Section 2. Though it is the best DCS scheme so far, DIM suffers from several problems. One problem is that events may well be mapped to orphan regions that contain no sensors. Thus, DIM requires some kludge to assign orphan regions to neighboring sensors. Another major problem in DIM is that of storage hot-spots. Storage hot-spots may occur if the sensors are not uniformly distributed. A storage hot-spot occurs when relatively many events are assigned to a relatively small number of the sensors. For example, if there was only one sensor on one side of the first bisection, then half of the events would be mapped to this sensor if the events were uniformly distributed over the range of possible events. Due to their storage constraints, the presence of a storage hot-spot leads to increasing the dropping rate of events by overloaded sensors. Clearly, this has a significant impact on the quality of data (QoD) generated by the sensor network. Queries for events in a storage hot-spot may be delayed due to contention at the storage sensors and the surrounding sensors. More critically, the sensors in and near the hot-spot may quickly run out of energy, due to the high insertion/query load imposed to them. This results in a loss of the events generated at these sensors, the events stored at these sensors, and possibly a decrease in network connectivity. Increased death of sensors results in decreasing the coverage area and causes the formation of coverage gaps within such area. Both of which consequently decrease QoD. Certainly, it is not desirable to have a storage scheme whose performance and QoD guarantees rest on the assumption that the sensors are uniformly distributed geographically. Storage hot-spots may also occur in DIM if the distribution of events is not uniform over the range of possible events. It is difficult to imagine any reasonable scenario where the events are uniformly distributed over the range of all possible events. Consider the situation where the only attribute sensed is temperature. One would expect that most temperature readings would be clustered within a relatively small range rather than uniform over all possible temperatures . Without any load balancing, those sensors responsible for temperatures outside this range would store no events. In this paper, we provide a load-balanced INDCS scheme based on K-D trees, that we, not surprisingly, call K-D tree based DCS (KDDCS). In our KDDCS scheme, the refinement of regions in the formation of the K-D tree has the property that the numbers of sensors on both sides of the partition are approximately equal. As a result of this, our K-D tree will be balanced, there will be no orphan regions, and, regardless of the geographic distribution of the sensors, the ownership of events will uniformly distributed over the sensors if the events are uniformly distributed over the range of possible events. We present a modification of GPSR routing, namely Logical Stateless Routing (LSR), for the routing of events from their generating sensors to their owner sensors, that is competitive with the GPSR routing used in DIM. In order to maintain load balance in the likely situation that the events are not uniformly distributed, we present a re-balancing algorithm that we call K-D Tree Re-balancing (KDTR). Our re-balancing algorithm guarantees load balance even if the event distribution is not uniform. KDTR has essentially minimal overhead. We identify a problem, that we call the weighted split median problem, that is at the heart of both the construction of the initial K-D tree, and the re-balancing of the K-D tree. In the weighted split median problem, each sensor has an associated weight/multiplicity, and the sensors' goal is to distributively determine a vertical line with the property that the aggregate weight on each side of the line is approximately equal. We give a distributed algorithm for the weighted split median problem, and show how to use this algorithm to construct our initial K-D tree, and to re-balance the tree throughout the network lifetime. We are mindful of the time, message complexity, and node storage requirements, in the design and implementation of all of our algorithms. The time for all of our algorithms is within a poly-log factor of the diameter of the network. Obviously, no algorithm can have time complexity less than the diameter of the network. The number of messages, and number of bits in those messages, that any particular node is required to send by our algorithms is poly-logarithmic in number of sensors. The amount of information that each node must store to implement one of our algorithms is logarithmic in the number of sensors. Experimental evaluation shows that the main advantages of KDDCS , when compared to the pure DIM, are: Achieving a better data persistence by balancing the storage responsibility among sensor nodes. Increasing the QoD by distributing the storage hot-spot events among a larger number of sensors. Increasing the energy savings by achieving a well balanced energy consumption overhead among sensor nodes. The rest of the paper is organized as follows. Section 2 presents an overview of the differences between DIM and KDDCS. Section 3 describes the weighted split median problem, and our distributed solution. Section 4 describes the components of KDDCS. Section 5 presents our K-D tree re-balancing algorithm. Experimental results are discussed in Section 6. Section 7 presents the related work. OVERVIEW OF DIM VS KDDCS In this section, we will briefly describe the components of both schemes, DIM and KDDCS, and highlight the differences between the two schemes using a simple example. We assume that the sensors are arbitrarily deployed in the convex bounded region R. We assume also that each sensor is able to determine its geographic location (i.e., its x and y coordinates), as well as, the boundaries of the service area R. Each node is assumed to have a unique NodeID, like a MAC address. Sensor nodes are assumed to have the capacity for wireless communication, basic processing and storage, and they are associated with the standard energy limitations. The main components of any DCS scheme are: the sensor to address mapping that gives a logical address to each sensor, and the event to owner-sensor mapping that determines which sensor will store the event. The components of DIM and KDDCS are: Repetitive splitting of the geographic region to form the underlying K-D tree, and the logical sensor addresses. Repetitive splitting of the attribute ranges to form the bit-code for an event. The routing scheme to route an event from the generating sensor to the owner sensor. We now explain how DIM implements these components. Let us start with the formation of the K-D tree in DIM. DIM starts the network operation with a static node to bit-code mapping phase. In such phase, each sensor locally determines its binary address by uniformly splitting the overall service area in a round 318 Figure 1: Initial network configuration Figure 2: DIM K-D tree robin fashion, horizontally then vertically, and left shifting its bit-code with every split by 0 (or 1) bit when falling above (or below) the horizontal split line (similarly, by a 0 bit if falling on the left of the vertical split line, or a 1 bit otherwise). Considering the region as partitioned into zones, the process ends when every sensor lies by itself in a zone, such that the sensor address is the zone bit code. Thus, the length of the binary address of each sensor (in bits) represents its depth in the underlying K-D tree. Note that from a sensor address, one can determine the physical location of the sensor. In case any orphan zones exist (zones physically containing no sensors in their geographic area), the ownership of each of these zones is delegated to one of its neighbor sensors. As an example, consider the simple input shown in Figure 1. The K-D tree formed by DIM is shown in Figure 2. In this figure, the orphan zone (01) is assumed to be delegated to node 001, which is the least loaded among its neighbors. We now turn to the construction of an event bit-code in DIM. The generation of the event bit-code proceeds in rounds. As we proceed, there is a range R j associated with each attribute j of the event. Initially, the range R j is the full range of possible values for attribute j. We now describe how a round i 0 works. Round i, determines the (i+1) th high order bit in the code. Round i depends on attribute j = i mod k of the event, where k is the number of attributes in the event. Assume the current value of R j is [a, c], and let b = (a + c)/2 be the midpoint of the range R j . If the value of attribute j is in the lower half of the range R j , that is in [a, b], then the i th bit is 0, and R j is set to be the lower half of R j . If the value of attribute j is in the upper half of the range R j , that is in [b, c], then the i th bit is 1, and R j is set to be the upper half of R j . To show the events to bit-code mapping in DIM, consider that the events in our example (shown in Figure 2) are composed of two attributes, temperature and pressure, with ranges (30, 70) and (0, 2), respectively. Let an event with values (55, 0.6) be generated by Node N3(11). The 4 high-order bits of the bit-code for this event are 1001. This is because temperature is in the top half of the range [30, 70], pressure is in the bottom half of the range [0, 2], then temperature is in the bottom half of the range [50, 70], and pressure is in the top half of the range [0, 1]. Thus, the event should be routed toward the geometric location specified by code 1001. In DIM, an event is routed using Greedy Perimeter Stateless Routing (GPSR) [6] to the geographic zone with an address matching the high order bits of the event bit-code. In our example, the sensor 10 will store this event since this is the sensor that matches Figure 3: KDDCS K-D tree the high order bits of the bit-code 1001. If there is no sensor in this region, then, the event is stored in a neighboring region. We now highlight the differences between our proposed KDDCS scheme, and DIM. The first difference is how the splitting is accomplished during the formation of the K-D tree. In KDDCS, the split line is chosen so that there are equal numbers of sensors on each side of the split line. Recall that, in DIM, the split line was the geometric bisector of the region. Thus, in KDDCS, the address of a sensor is a logical address and does not directly specify the location of the sensor. Also, note that the K-D tree in KDDCS will be balanced , while this will not be the case in DIM if the sensors are not uniformly distributed. This difference is illustrated by the K-D tree formed by KDDCS shown in Figure 3 for the same simple input shown in Figure 1. The second difference is that in determining the owner sensor for an event, the range split point b need not be the midpoint of the range R j . The value of b is selected to balance the number of events in the ranges [a, b] and [b, c]. Thus, in KDDCS, the storage of events will be roughly uniform over the sensors. The third difference is that, since addresses are not geographic, KDDCS needs a routing scheme that is more sophisticated than GPSR. THE WEIGHTED SPLIT MEDIAN PROBLEM Before presenting our KDDCS scheme, we first define the weighted split median problem in the context of sensor networks and present an efficient distributed algorithm to solve the problem. Each sensor s i initially knows w i associated values v 1 , . . . v w i . Let W = P n i=1 w i be the number of values. The goal for the sensors is to come to agreement on a split value V with the property that approximately half of the values are larger than V and half of the values are smaller than V . We present a distributed algorithm to solve this problem. The time complexity of our algorithm is O(log n) times the diameter of the communication network in general, and O(1) times the diameter if n is known a priori within a constant factor. Each node is required to send only O(log n) sensor ID's. The top level steps of this algorithm are: 1. Elect a leader sensor s , and form a breadth first search (BFS) tree T of the communication network that is rooted at s . 2. The number of sensors n, and the aggregate number of values W is reported to s . 3. The leader s collects a logarithmically-sized uniform random sample L of the values. The expected number of times that a value from sensor s i is included in this sample is " w i log n W " . 4. The value of V is then the median of the reported values in L, which s reports to all of the sensors. We need to explain how these steps are accomplished, and why the algorithm is correct. 319 We start with the first step. We assume that each sensor has a lower bound k on the number of sensors in R. If a sensor has no idea of the number of other sensors, it may take k = 2. Then, each sensor decides independently, with probability ` ln k k , to become a candidate for the leader. Each candidate sensor s c initiates the construction of a BFS tree of the communication graph rooted at s c by sending a message Construct(s c ) to its neighbors. Assume a sensor s i gets a message Construct(s c ) from sensor s j . If this is the first Construct(s c ) message that it has received, and s c 's ID is larger than the ID of any previous candidates in prior Construct messages, then: s i makes s j its tentative parent in the BFS tree T , and forwards the Construct(s c ) message to its neighbors. If the number of candidates was positive, then, after time proportional to the diameter of the communication network, there will be a BFS tree T rooted at the candidate with the largest ID. Each sensor may estimate an upper bound for the diameter of the communication graph to be the diameter of R divided by the broadcast radius of a sensor. After this time, the sensors know that they have reached an agreement on T , or that there were no candidates. If there were no candidates, each sensor can double its estimate of k, and repeat this process. After O(log n) rounds, it will be the case that k = (n). Once k = (n), then, with high probability (that is, with probability 1 1 poly(n) ), the number of candidates is (log n). Thus, the expected time complexity to accomplish the first step is O(n log n). Assuming that each ID has O(log n) bits, the expected number of bits that each sensors has to send is O(log 2 n) since there are are likely only O(log n) candidates on the first and only round in which there is a candidate. A log n factor can be removed if each sensor initially knows an estimate of n that is accurate to within a multiplicative constant factor. The rest of the steps will be accomplished by waves of root-to-leaves and leaves-to-root messages in T . The second step is easily accomplished by a leave-to-root wave of messages reporting on the number of sensors and number of values in each subtree. Let T i be the subtree of T rooted at sensor s i , and W i the aggregate number of values in T i . The value W i that s i reports to its parents is w i plus the aggregate values reported to s i by its children in T . The sensor count that s i reports to its parents is one plus the sensor counts reported to s i by its children in T . The third step is also accomplished by a root-to-leaves wave and then a leaves-to-root wave of messages. Assume a sensor s i wants to generate a uniform random sample of L i of the values stored in the sensors in T i . The value of L for the leader is (log n). Let s i 1 , . . . , s i d be the children of s i in T . Node s i generates the results to L i Bernoulli trials, where each trial has d + 1 outcomes corresponding to s i and its d children. The probability that the outcome of a trial is s i is w i W i , and the probability that the outcome is the child s i j is w ij W i . Then, s i informs each child s i j how often it was selected, which becomes the value of L i j s i , then waits until it receives samples back from all of its children. s i then unions these samples, plus a sample of values of the desired size from itself, and then passes that sample back to its parent. Thus, each sensor has to send O(log n) ID's. The leader s then sets V to be the median of the values of the sample L, then, in a root-to-leaves message wave, informs the other sensors of the value of V . We now argue that, with high probability, the computed median of the values is close to the true median. Consider a value ^ V such that only a fraction &lt; 1 2 of the values are less than ^ V . One can think of each sampled value as being a Bernoulli trial with outcomes less and more depending on whether the sampled value is Figure 4: Logical address assignment algorithm less than ^ V . The number of less outcomes is binomially distributed with mean L. In order for the computed median to be less than ^ V , one needs the number of less outcomes to be at least L/2, or equivalently ( 1 2 -)L more than the mean L. But the probability that a binomially distributed variable exceeds its mean by a factor of 1 + is at most e -2 3 . Thus, by picking the multiplicative constant in the sample size to be sufficiently large (as a function of ), one can guarantee that, with high probability, the number of values less than the computed median V cannot be much more than L/2. A similar argument shows that the number more than the computed median V can not be much more than L/2. If the leader finds that n is small in step 2, it may simply ask all sensors to report on their identities and locations, and then compute V directly. Now that we solved the weighted split median problem, we present the components of the KDDCS scheme in the next section. KDDCS We now present our KDDCS scheme in details. We explain how the initial K-D tree is constructed, how events are mapped to sensors , and how events are routed to their owner sensors. 4.1 Distributed Logical Address Assignment Algorithm The main idea of the algorithm is that the split lines used to construct the K-D tree are selected so that each of the two resulting regions contain an equal number of sensors. The split line can be determined using our weighted split median algorithm with each sensor having unit weight, and the value for each sensor is either its x coordinate or its y coordinate. The recursive steps of the algorithm are shown in Figure 4. We now describe in some greater detail how a recursive step works. The algorithm starts by partitioning the complete region R horizontally . Thus, the distributed weighted split median algorithm (presented in section 3) is applied for R using the y-coordinates of the sensors to be sent to the BFS root. Upon determining weighted split median of R, sensors having lower y-coordinate than the median value (we refer to these sensors as those falling in the lower region of R) assign their logical address to 0. On the other hand, those sensor falling on the upper region of R assign themselves a 1 logical address. At the end of the first recursive step, the terrain can be looked at as split into two equally logically loaded partitions (in terms of the number of sensors falling in each partition). At the next step, the weighted split median algorithm is applied locally in each of the sub-regions (lower/upper), while using the sensors' x-coordinates, thus, partitioning the sub-regions vertically rather than horizontally. Similarly, sensors' logical addresses are updated by left-shifting them with a 0 bit for those sensors falling 320 in the lower regions (in other words, sensor nodes falling on the left of the weighted median line), or with a 1 bit for sensor nodes falling in the upper regions (i.e., sensor nodes falling on the right of the weighted median line). The algorithm continues to be applied distributively by the different subtrees until each sensor obtains a unique logical address, using x and y coordinates of sensors, in a round robin fashion, as the criterion of the split. The algorithm is applied in parallel on the different subtrees whose root nodes fall at the same tree level. At the i th recursive step, the algorithm is applied at all intermediate nodes falling at level i- 1 of the tree. Based on the definition of the weighted split median problem, the algorithm results in forming a balanced binary tree, such that sensors represent leaf nodes of this tree (intermediate nodes of the tree are logical nodes, not physical sensors). The algorithm terminates in log n recursive steps. At the end of the algorithm, the size of the logical address given to each sensor will be log n bits. Recall that the time complexity of our weight split median algorithm is O(d log n), where d is the diameter of the region. Thus, as the depth of our K-D tree is O(log n), we get that the time complexity for building the tree is O(d log 2 n). If the sensors are uniformly distributed, then, as the construction algorithm recurses, the diameters of the regions will be geometrically decreasing. Thus, in the case of uniformly distributed sensors, one would expect the tree construction to take time O(d log n). As our weighted split median algorithm requires each sensor to send O(log n) ID's, and our K-D tree has depth O(log n), we can conclude that during the construction of our K-D tree, the number of ID's sent by any node is O(log 2 n). 4.2 Event to Bit-code Mapping In this section, we explain how the event to bit-code mapping function is determined. Recall that the main idea is to set the split points of the ranges so that the storage of events is roughly uniform among sensor nodes. To construct this mapping requires a probability distribution on the events. In some situations, this distribution might be known. For example, if the network has been operational for some period of time, a sampling of prior events might be used to estimate a distribution. In cases where it is not known, say when a network is first deployed, we can temporarily assume a uniform distribution. In both cases, we use the balanced binary tree as the base tree to overlay the attribute-specific K-D tree on (Recall that a K-D tree is formed by k attributes). This is basically done by assigning a range for each of the k attributes to every intermediate node in the tree. Note that the non-leaf nodes in the K-D tree are logical nodes that do not correspond to any particular sensor. One may think of non-leaf nodes as regions. Any split point p of a node x of tree level l, where l%k = i, represents a value of attribute i. Such split point partitions the range of attribute i falling under responsibility of node x into two subranges such that the the subrange lower than p is assigned to the left child of x, while the other range is assigned to x's right child. Note that the other k - 1 ranges of node x, corresponding to the remaining k-1 attributes, are simply inherited by both children of x. Knowing the data distribution, the split points of the tree should be predefined in a way to cope with any expected irregularity in the load distribution among the K-D tree leaf nodes. For example, given an initial temperature range (30, 70) and knowing that 50% of the events will fall in the temperature range (65, 70), the root split point should be defined as 65 (assuming that the temperature is the first attribute in the event). Therefore, based on the selected root split point, the left child subtree of the root will be responsible of storing events falling in the temperature range (30, 65), while the right child subtree will store events falling in the range (65, 70). Figure 3 gives an example of non-uniform initialization of split points. We finish by describing what information is stored in each sensor node. Each sensor node corresponds to a leaf in the K-D tree. Each sensor knows its logical address in the tree. Further, each leaf in the K-D tree knows all the pertinent information about each of its ancestors in the tree. The pertinent information about each node is: The geographic region covered. The split line separating its two children. The attribute range, and attribute split point, associated with this region. From this information, each leaf/sensor can determine the range of events that will be stored at this sensor. Note that each sensor only stores O(log n) information about the K-D tree. 4.3 Incremental Event Hashing and Routing Strictly speaking, the events-to-sensors mapping in DIM actually produces a geographic location. GPSR routing can then be used to route that event towards that geographic location. If the destination is contained in a leaf region with one sensor, then that sensor stores the event. If the leaf region is an orphan, then one of the sensors in the neighboring regions will store this event. In our scheme, the events-to-sensors mapping provides a logical address. Essentially, all that the sensor generating the event can determine from this logical address is a general direction of the owner sensor. Thus, our routing protocol, which we call Logical Stateless Routing (LSR), is in some sense less direct. LSR operates in O(log n) rounds. We explain how a round works. Assume that a source sensor with a logical address s wants to route an event e to a sensor with logical address t. However, s does not know the identity of the sensor t. Recall that s knows the pertinent information about its ancestors in the K-D tree. In particular, s knows the range split values of its ancestors. Thus, s can compute the least common ancestor (LCA) of s and t in the K-D tree. Assume that the first bit of disagreement between s and t is the th bit. So, the least common ancestor (LCA) of s and t in the K-D tree has depth . Let R be the region corresponding to the LCA of s and t, L the split line corresponding to this region, and R 0 and R 1 the two subregions of R formed by L. Without loss of generality, assume that s R 0 and t R 1 . From its own address, and the address of t, the sensor s can conclude that t is in the region R 1 . Recall that s knows the location of the split line L. The sensor s computes a location x in the region R 1 . For concrete-ness here, let us assume that x is some point in R 1 that lies on the line intersecting s and perpendicular to L (Although there might be some advantages to selecting x to be the geometric center of the region R 1 ). LSR then directs a message toward the location x using GPSR. The message contains an additional field noting that this is a th round message. The th round terminates when this message first reaches a sensor s whose address agrees with the address of t in the first + 1 bits. The sensor s will be the first sensor reached in R 1 . Round + 1 then starts with s being the new source sensor. We explain how range queries are routed by means of an example . This example also essentially illustrates how events are stored. Figure 5 gives an example of a multi-dimensional range query and shows how to route it to its final destination. In this example, a multi-dimensional range query arises at node N 7(111) asking for the number of events falling in the temperature range (30, 32) and pressure range (0.4, 1) that were generated throughout the last 2 minutes. Node N 7 knows that the range split point for the root 321 Figure 5: Example of routing a query on KDDCS was temperature 40, and thus, this query needs to be routed toward the left subtree of the root, or geometrically toward the top of the region, using GPSR. The first node in this region that this event reaches is say N 3. Node N 3 knows that the first relevant split point is pressure = 0.5. Thus, the query is partitioned into two sub-queries, ((30, 32), (0.4, 0.5)) and ((30, 32), (0.5, 1)). When processing the first subquery, node N 3 forwards it to the left using GPSR. N 3 can then tell that the second query should be routed to the other side of its parent in the K-D tree since the range split for its parent is temperature 34. The logical routing of this query is shown on the right in Figure 5, and a possible physical routing of this query is shown on the left in Figure 5. As LSR does not initially know the geometric location of the owner sensor, the route to the owner sensor cannot possibly be as direct as it is in DIM. But, we argue that the length of the route in LSR should be at most twice the length of the route in DIM. Assume for the moment that all messages are routed by GPSR along the direct geometric line between the source sensor and the destination location. Let us assume, without loss of generality, that LSR is routing horizontally in the odd rounds. Then, the routes used in the odd rounds do not cross any vertical line more than once. Hence, the sum of the route distances used by LSR in the odd rounds is at most the diameter of the region. Similarly, the sum of the route distances used by LSR in the even rounds is at most the diameter of the region. Thus, the sum of the route distances for LSR, over all rounds, is at most twice the diameter. The geometric distance between the source-destination pair in DIM is obviously at most the diameter. So we can conclude that the length of the route found by LSR is at most twice the length of the route found by DIM, assuming that GPSR is perfect. In fact, the only assumption that we need about GPSR to reach this conclusion is that the length of the path found by GPSR is roughly a constant multiple times the geometric distance between the source and destination. Even this factor of two can probably be improved significantly in expectation if the locations of the sensors are roughly uniform. A simple heuristic would be to make the location of the target x equal to the location of the destination sensor t if the sensors in R 1 where uniformly distributed . The location of x can easily be calculated by the source sensor s given information local to s. KDTR K-D TREE RE-BALANCING ALGORITHM Based on the KDDCS components presented so far, KDDCS avoids the formation of storage hot-spots resulting from skewed sensor deployments, and from skewed events distribution if the distribution of events was known a priori. However, storage hot-spots may be formed if the initial presumed events distribution was not correct, or if events distribution evolves over times. We present a K-D tree re-balancing algorithm, KDTR, to re-balance the load. In the next subsections, we first explain how to determine the roots of the subtrees that will re-balance, and then show how a re-balancing operation on a subtree works. We assume that this re-balancing is performed periodically with a fixed period. 5.1 Selection of Subtrees to be Re-Balanced The main idea is to find the highest unbalanced node in the KD tree. A node is unbalanced if the ratio of the number of events in one of the child subtrees over the number of events stored in the other child subtree exceeds some threshold h. This process of identifying nodes to re-balance proceeds in O(log n) rounds from the leaves to the root of the K-D tree. We now describe how round i 1 works. Intuitively, round i occurs in parallel on all subtrees rooted at nodes of height i + 1 in the K-D tree. Let x be a node of height i + 1. Let the region associated with x be R, the split line be L, and the two subregions of R be R 0 and R 1 . At the start of this round, each sensor in R 0 and R 1 knows the number of stored events C 0 and C 1 in R 0 and R 1 , respectively. The count C 0 is then flooded to the sensors in R 1 , and the count C 1 is flooded to the sensors in R 0 . After this flooding, each sensor in R knows the number of events stored in R, and also knows whether the ratio max( C 0 C 1 , C 1 C 0 ) exceeds h. The time complexity per round is linear in the diameter of a region considered in that round. Thus, the total time complexity is O(D log n), where D is the diameter of the network, as there are O(log n) rounds. The number of messages sent per node i in a round is O(d i ) , where d i is the degree of node i in the communication network. Thus, the total number of messages sent by a node i is O(d i log n). Re-Balancing is then performed in parallel on all unbalanced nodes, that have no unbalanced ancestors. Note that every leaf knows if an ancestor will re-balance, and is so, the identity of the unique ancestor that will balance. All the leaves of a node that will re-balance, will be aware of this at the same time. 5.2 Tree Re-balancing Algorithm Let x be an internal node to the K-D tree that needs to be re-balanced . Let the region associated with x be R. Let the attribute associated with node x be the j'th attribute. So, we need to find a new attribute split L for the j'th attribute for node x. To accomplish this, we apply the weighted split median procedure, where the weight w i associated with sensor i is the number of events stored at sensor i, and the values are the j'th attributes of the w i events stored at that sensor. Thus, the computed attribute split L has the property that, in expectation, half of the events stored in R have their j'th attribute larger than L, and half of the events stored in R have their j'th attribute smaller than L. Let R 0 and R 1 be the two subregions of R. Eventually, we want to recursively apply this process in parallel to the regions R 0 and R 1 . But before recursing, we need to route some events from one of R 0 or R 1 to the other. The direction of the routing depends on whether the attribute split value became larger or smaller. Let us assume, without loss of generality, that events need to be routed from R 0 to R 1 . Consider an event e stored at a sensor s in R 0 that needs to be routed to R 1 . The sensor s picks a destination logical address t, uniformly at random, from the possible addresses in the region R 1 . The event e is then routed to t using the routing scheme described in section 4.3. The final owner for e in R 1 cannot be determined until our process is recursively applied to R 1 , but this process cannot be recursively applied until the events that should be stored in R 1 are contained in R 1 . The fact the the destination addresses in R 1 were picked uniformly at random ensures load balance . This process can now be recursively applied to R 0 and R 1 . 322 Figure 6: KDDCS original K-D tree We now discuss the complexity of this procedure. We break the complexity into two parts: the cost of performing the weighted split median operation, and the cost of migrating the events. One application of the weighted split median has time complexity O(D log n), where D is the diameter of the region, and messages sent per node of O(log 2 n) messages. Thus, we get time complexity O(D log 2 n) and messages sent per node of O(log 3 n) for all of the applications of weighted split median. Every period re-balance requires each event to travel at most twice the diameter of the network (assuming that GPSR routes on a direct line). The total number of events that can be forced to migrate as a result of k new events being stored is O(k log k). Thus, the amortized number of migrations per event is logarithmic, O(log k) in the number of insertions. This amount of re-balancing per insertion is required for any standard dynamic data structure (e.g. 2-3 trees, AVL trees, etc.). Figures 6 and 7 show a detailed example illustrating how KDTR works. Continuing on the same example we presented in Section 4.2, we monitor how KDTR maintains the K-D tree balancing in the course of successive insertions. Starting with an equal number of 3 events stored at each sensor, a storage hot-spot arises in node N 7 after 6 event insertions. By checking the ratio of N 7 storage to that of N 7, KDTR identifies the subtree rooted at node 11 as an unbalanced subtree. As none of node 11's ancestors is unbalanced at this point, KDTR selects 11 to be re-balanced. However, the storage load remains skewed toward subtree 1, thus, after another 6 insertions, KDTR re-balances the subtree rooted at 1. After 12 more insertions aiming the right subtree of the root, KDTR re-balances the root of the tree, basically changing the attribute-based split points of almost all internal nodes, in order to maintain the balance of the tree. Note that, as long as the average loads of sensors which are falling outside the hot-spot area increases, the frequency of re-balancing decreases. We digress slightly to explain a method that one might use to trigger re-balancing, as opposed to fixed time period re-balancing. Each sensor s i knows the number of events that are stored in each region corresponding to an ancestor of s i in the K-D tree when this region was re-balanced. Let C j be the number of events at the last re-balancing of the region R j corresponding to node of depth j on the path from the root to s i in the K-D tree. Assume that the region R j has elected a leader s j . Then, the number of events that have to be stored in R j , since the last re-balancing, to cause another re-balancing in R j is something like hC j , where h is the unbalancing ratio that we are willing to tolerate. Then, each insertion to s i is reported by s i to s j with probability something like " log n hC j " . Thus, after seeing (log n) such notifications, the leader s j can be confident that there have been very close to hC j insertions into the region R j , and a re-balancing might be warranted. Note that the role of leader requires only receiving O(log n) messages. Figure 7: KDTR example EXPERIMENTAL RESULTS In order to evaluate our KDDCS scheme, we compared its performance with that of the DIM scheme, that has been shown to be the best among current INDCS schemes [9]. In our simulation, we assumed having sensors of limited buffer and constrained energy. We simulated networks of sizes ranging from 50 to 500 sensors, each having an initial energy of 50 units, a radio range of 40m, and a storage capacity of 10 units. For simplicity , we assumed that the size of a message is equal to the size of a storage unit. We also assumed that the size of a storage unit is equal to the size of an event. When sent from a sensor to its neighbor, a message consumes 1 energy unit from the sender energy and 0.5 energy unit from the receiver energy. The service area was computed such that each node has on average 20 nodes within its nominal radio range. As each sensor has a limited storage capacity, it is assumed to follow a FIFO storage approach to handle its cache. Thus, a sensor replaces the oldest event in its memory by the newly incoming event to be stored in case it is already full when receiving this new event. We modeled a network of temperature sensors. The range of possible reading values was [30, 70]. We modeled storage hot-spots by using a random uniform distribution to represent sensors' locations, while using a skewed distribution of events among the attributes ranges. Note that the regular sensor deployment assumption does not affect our ability to assess the effectiveness of KDDCS as the storage hot-spot can result from either skewed sensor deployments, or skewed data distributions, or both. The storage hot-spot size is characterized by the skewness dimensions, which are the percentage of the storage hot-spot events to the total number of events generated by the sensor network and the percentage of the read-323 0 200 400 600 800 1000 1200 1400 1600 1800 2000 50 100 150 200 250 300 350 400 450 500 Dropped Events Network Size DIM KDDCS/KDTR Figure 8: Number of dropped events for networks with a 80%-10% hot-spot 0 200 400 600 800 1000 1200 1400 1600 1800 2000 50 100 150 200 250 300 350 400 450 500 Dropped Events Network Size DIM KDDCS/KDTR Figure 9: Number of dropped events for networks with a 80%-5% hot-spot ings' range in which the hot-spot events fall to the total possible range of temperature readings. We assumed that a single storage hot-spot is imposed on the sensor network. To follow the behavior of KDDCS toward storage hot-spots of various sizes, we simulated, for each network size, a series of hot-spots where a percentage of 10 % to 80% of the events fell into a percentage of 5% to 10% of the reading' range. Note that we always use the term x%-y% hot-spot to describe a storage hot-spot where x% of the total generated events fall into y% of the readings' range. We used a uniform split points initialization to setup the attribute range responsibilities of all internal nodes of the K-D tree. For the re-balancing threshold, we used a value of 3 to determine that a specific subtree is unbalanced. Node failures were handled in the same way as DIM. When a node fails, its stored events are considered lost. Futher events directed to the range responsibility of such node are directed to one of its close neighbors. We ran the simulation for each network size and storage hot-spot size pair. Each simulation run consisted of two phases: the insertion phase and the query phase. During the insertion phase, each sensor generates (i.e. reads) 5 events, according to the predefined hot-spot size and distribution, and forward each of these event to its owner sensor. In the query phase, each sensor generates queries of sizes ranging from 10% to 90% of the [30, 70] range. The query phase is meant to measure the damages, in terms of QoD and energy losses, caused by the storage hot-spot. The results of the simulations are shown in the Figures 8 to 17. In these figures, we compare the performance of our KDDCS scheme versus that the DIM scheme with respect to various performance measures. Note that we only show some of our findings due to space constraints. R1. Data Persistence: Figures 8 and 9 present the total number events dropped by all network nodes in networks with 80%-10% and 80%-5% hot-spots, respectively. By analyzing the difference 0 200 400 600 800 1000 1200 1400 50 100 150 200 250 300 350 400 450 500 Events Returned for (0.5 * attribute range) Query Network Size DIM KDDCS/KDTR Figure 10: Query size of a 50% query for networks with a 80%-10% hot-spot 0 200 400 600 800 1000 1200 1400 50 100 150 200 250 300 350 400 450 500 Events Returned for (0.8 * attribute range) Query Network Size DIM KDDCS/KDTR Figure 11: Query size of a 80% query for networks with a 80%-5% hot-spot between KDDCS and DIM, we can find out that the number of dropped events in the first is around 40% to 60% of that in the second. This can be interpreted by the fact that KDDCS achieves a better load balancing of storage among the sensors. This leads to decreasing the number of sensors reaching their maximum storage, and decreasing the total number of such nodes compared to that in the pure DIM. This directly results in decreasing the total number of dropped events and achieving a better data persistence. Another important remark to be noted based on the two figures is that decreasing the size of the hot-spot by making the same number of events to fall into a smaller attributes' range does not highly affect the overall performance of KDDCS compared to that of DIM. R2. Quality of Data: Figures 10 and 11 show the average query sizes of 50% and 80% of the attribute ranges for networks with a 80%-10% and 80%-5% hot-spots, respectively. It is clear that KDDCS remarkebly improves the QoD provided by the sensor network . This is mainly due to dropping less information (as pointed at in R1), thus, increasing the number of events resulting in each query. The gap between DIM and KDDCS, in terms of resulting query sizes, is really huge for in both graphs, which indicates that KDDCS outperforms DIM for different storage hot-spot sizes. This result has a very important implication on the data accuracy of the sensor readings output from a network experiencing a hot-spot. The success of the KDDCS in avoiding hot-spots results in improving the network ability to keep a higher portion of the hot-spot data. This ameliorates the degree of correctness of any aggregate functions on the network readings, for example, an average of the temperature or pressure values where a high percentage of the data is falling within a small range of the total attributes' range. We consider this to be a good achievement compared to the pure DIM scheme. R3. Load Balancing: Figures 12 and 13 show the average node storage level for networks with 70%-10% and 60%-5% hot-spots, 324 0 1 2 3 4 5 50 100 150 200 250 300 350 400 450 500 Average Storage Level Network Size DIM KDDCS/KDTR Figure 12: Average node storage level for networks with a 70%-10% hot-spot (numbers rounded to ceiling integer) 0 1 2 3 4 5 50 100 150 200 250 300 350 400 450 500 Average Storage Level Network Size DIM KDDCS/KDTR Figure 13: Average node storage level for networks with a 60%-5% hot-spot (numbers rounded to ceiling integer) respectively. By a node storage level, we mean the number of events stored in the node's cache. The figures show that KDDCS has a higher average storage level than DIM, especially for less skewed hot-spots. This can be interpreted as follows. When a storage hot-spot arises in DIM, the hot-spot load is directed to a small number of sensors. These nodes rapidly reach their storage maximum , while almost all other sensor nodes are nearly empty. Therefore , the load distribution is highly skewed among nodes leadind to a low average storage level value. However, in KDDCS, the number of nodes effectively storing events increases. Subsequently, the average storage load value increases. This gives us a truthful figure about the better storage balancing the network. It is worth mentioning that each of the values in both figures is rounded to the ceiling integer. Thus, in both cases, the average in DIM does not exceed one event per sensor for all network sizes. R4. Energy Consumption Balancing: Figures 14 and 15 show the average node energy level at the end of the simulation for networks with 70%-10% and 50%-5% hot-spots, respectively. The figures show that this average generally decreases with the increase 30 35 40 45 50 50 100 150 200 250 300 350 400 450 500 Average Energy Level Network Size DIM KDDCS/KDTR Figure 14: Average sensors' energy levels for networks with a 70%-10% hot-spot 30 35 40 45 50 50 100 150 200 250 300 350 400 450 500 Average Energy Level Network Size DIM KDDCS/KDTR Figure 15: Average sensors' energy levels for networks with a 50%-5% hot-spot 0 500 1000 1500 2000 2500 3000 3500 50 100 150 200 250 300 350 400 450 500 Moved Events for x%-10% hot-spot Network Size x=40 x=60 x=80 Figure 16: Number of event movements for networks with a x%-10% hot-spot of the network size for both schemes. The interesting result that these figures show is that both KDDCS and DIM result in fairly close average energy consumption among the sensors. However, as we mentioned in R3 and based on the way DIM works, most of the energy consumed in DIM is effectively consumed by a small number of nodes, namely those falling in the hot-spot region. On the other hand, the number of nodes consuming energy increases in KDDCS due to the better load balancing KDDCS achieves, while the average energy consumed by each sensor node decreases. Thus, although the overall energy consumption is the same in both KDDCS and DIM, this result is considered as a positive result in terms of increasing the overall network lifetime, as well as avoiding the early death of sensor nodes, which leads to avoid network partitioning . R5. Events Movements: Figures 16 and 17 show the number of migrated events for networks with x% - 10% and x% - 5% hot-spots , respectively, where x varies from 40 to 80. For both sets of hot-spot sizes, the number of event movements lineraly increases with the network size. The important result to be noted in both 0 500 1000 1500 2000 2500 3000 3500 50 100 150 200 250 300 350 400 450 500 Moved Events for x%-5% hot-spot Network Size x=40 x=60 x=80 Figure 17: Number of event movements for networks with a x%-5% hot-spot 325 figure is that the total number of movements is not highly dependent on the hot-spot size. This is mainly because KDDCS avoids storage hot-spots in their early stages instead of waiting for a large storage hot-spot to be formed, and then try to decompose it. Therefore , most of the event movements are really done at the start of the formation of the storage hot-spot. This leads to the fact that, for highly skewed data distributions, (i.e. large hot-spot sizes), the number of event movements does not highly change with changing the exact storage hot-spot size. RELATED WORK Many approaches have been presented in literature defining how to store the data generated by a sensor network. In the early age of sensor networks research, the main storage trend used consisted of sending all the data to be stored in base stations, which lie within, or outside, the network. However, this approach may be more appropriate to answer continuous queries, which are queries running on servers and mostly processing events generated by all the sensor nodes over a large period of time [4, 10, 18, 14, 12, 11]. In order to improve the lifetime of the sensor network, as well as the QoD of ad-hoc queries, INS techniques have been proposed. All INS schemes presented so far were based on the idea of DCS [15]. These INDCS schemes differ from each other based on the events-to-sensors mapping used. The mapping was done using hash tables in DHT [15] and GHT [13], or using K-D trees in DIM [9]. The formation of storage hot-spots due to irregularity, in terms of sensor deployment or events distribution, represent a vital issue in current INDCS techniques [5]. Some possible solutions for irregular sensors deployments were highlighted by [5], such as routing based on virtual coordinates, or using heuristics to locally adapt to irregular sensor densities. Recently, some load balancing heuristics for the irregular events distribution problem were presented by [2, 8]. Such techniques were limited in their capability to deal with storage hot-spots of large sizes as they were basically acting like storage hot-spots detection and decomposition schemes, rather than storage hot-spots avoidance schemes like KDDCS. To the best of our knowledge, no techniques have been provided to cope with both types of irregularities at the same time. A complentary work to our paper is that on exploting similarities in processing queries issued by neighboring sensors in a DCS scheme [16]. Query Hot-Spots is another important problem that is orthogonal to the storage hot-spots problem. The problem arizes when a large percentage of queries ask for data stored in few sensors. We identified the problem in an earlier paper [1] and presented two algorithms , Zone Partitioning (ZP) and Zone Partial Replication (ZPR), to locally detect and decompose query hot-spots in DIM. We believe that KDDCS is able to cope with query hot-spots provided minor changes are added to the scheme. We aim at addressing this problem in the KDDCS testbed that we plan to develop. Recently, Krishnamurthy et al. [7] presented a novel DCS scheme, called RESTORE, that is characterized by real time event correlation . It would be interesting to compare the performance of both KDDCS and RESTORE in terms of load balacing. CONCLUSIONS Sensor databases are becoming embedded in every aspect of our life from merchandise tracking, healthcare, to disaster responds. In the particular application of disaster management, it has been ar-gued that it is more energy efficient to store the sensed data locally in the sensor nodes rather than shipping it out of the network, even if out-of-network storage is available. The formation of Storage Hot-Spots is a major problem with the current INDCS techniques in sensor networks. In this paper, we presented a new load-balanced INDCS scheme, namely KDDCS, that avoids the formation of storage hot-spots arising in the sensor network due to irregular sensor deployment and/or irregular events distribution. Further, we proposed a new routing algorithm called Logical Stateless Routing, for routing events from the generating sensors to the storage sensors, that is competitive with the popular GPSR routing. Our experimental evaluation has confirmed that our proposed KDDCS both increases the quality of data and the energy savings by distributing events of the storage hot-spots among a larger number of sensors. Acknowledgments We would like to thank Mohamed Sharaf for his useful feedback. We would also like to thank the anonymous referees for their helpful comments. REFERENCES [1] M. Aly, P. K. Chrysanthis, and K. Pruhs. Decomposing data-centric storage query hot-spots in sensor networks. In Proc. of MOBIQUITOUS, 2006. [2] M. Aly, N. Morsillo, P. K. Chrysanthis, and K. Pruhs. Zone Sharing: A hot-spots decomposition scheme for data-centric storage in sensor networks. In Proc. of DMSN, 2005. [3] J. L. Bentley. Multidimensional binary search trees used for associative searching. In CACM, 18(9), 1975. [4] P. Bonnet, J. Gehrke, and P. Seshadri. Towards sensor database systems. In Proc. of MDM, 2001. [5] D. Ganesan, S. Ratnasamy, H. Wang, and D. Estrin. Coping with irregular spatio-temporal sampling in sensor networks. In Proc. of HotNets-II, 2003. [6] B. Karp and H. T. Kung. GPSR: Greedy perimeter stateless routing for wireless sensor networks. In Proc. of ACM Mobicom, 2000. [7] S. Krishnamurthy, T. He, G. Zhou, J. A. Stankovic, and S. H. Son. Restore: A real-time event correlation and storage service for sensor networks. In Proc. of INSS, 2006. [8] X. Li, F. Bian, R. Govidan, and W. Hong. Rebalancing distributed data storage in sensor networks. Technical Report No. 05-852, CSD, USC, 2005. [9] X. Li, Y. J. Kim, R. Govidan, and W. Hong. Multi-dimensional range queries in sensor networks. In Proc. of ACM SenSys, 2003. [10] S. Madden, M. Franklin, J. Hellerstein, and W. Hong. TAG: a tiny aggregation service for ad-hoc sensor networks. In Proc. of OSDI, 2002. [11] S.-J. Park, R. Vedantham, R. Sivakumar, and I. F. Akyildiz. A scalable approach for reliable downstream data delivery in wireless sensor networks. In Proc. of MobiHoc, 2004. [12] T. Pham, E. J. Kim, and W. M. Moh. On data aggregation quality and energy efficiency of wireless sensor network protocols. In Proc. of BROADNETS, 2004. [13] S. Ratnasamy, B. Karp, L. Yin, F. Yu, D. Estrin, R. Govidan, and S. Shenker. GHT: A grographic hash table for data-centric storage. In Proc. of WSNA, 2002. [14] M. A. Sharaf, J. Beaver, A. Labrinidis, and P. K. Chrysanthis. TiNA: A scheme for temporal coherency-aware in-network aggregation. In Proc. of MobiDE, 2003. [15] S. Shenker, S. Ratnasamy, B. Karp, R. Govidan, and D. Estrin. Data-centric storage in sensornets. In Proc. of HotNets-I, 2002. [16] P. Xia, P. K. Chrysanthis, and A. Labrinidis. Similarity-aware query processing in sensor networks. In Proc. of WPDRTS, 2006. [17] T. Yan, T. He, and J. A. Stankovic. Differentiated surveillance for sensor networks. In Proc. of SenSys, 2003. [18] Y. Yao and J. Gehrke. Query processing for sensor networks. In Proc. of CIDR, 2003. 326

quality of data (QoD);KDDCS;routing algorithm;Power-Aware;energy saving;Sensor Network;sensor network;Distributed Algorithms;weighted split median problem;DIM;data persistence;storage hot-spots;ad-hoc queries

121

Language-specific Models in Multilingual Topic Tracking

Topic tracking is complicated when the stories in the stream occur in multiple languages. Typically, researchers have trained only English topic models because the training stories have been provided in English. In tracking, non-English test stories are then machine translated into English to compare them with the topic models. We propose a native language hypothesis stating that comparisons would be more effective in the original language of the story. We first test and support the hypothesis for story link detection. For topic tracking the hypothesis implies that it should be preferable to build separate language-specific topic models for each language in the stream. We compare different methods of incrementally building such native language topic models.

INTRODUCTION Topic detection and tracking (TDT) is a research area concerned with organizing a multilingual stream of news broadcasts as it arrives over time. TDT investigations sponsored by the U.S. government include five different tasks: story link detection, clustering (topic detection), topic tracking, new event (first story) detection , and story segmentation. The present research focuses on topic tracking, which is similar to filtering in information retrieval . Topics are defined by a small number of (training) stories, typically one to four, and the task is to find all the stories on those topics in the incoming stream. TDT evaluations have included stories in multiple languages since 1999. TDT2 contained stories in English and Mandarin. TDT3 and TDT4 included English, Mandarin, and Arabic. Machine-translations into English for all non-English stories were provided , allowing participants to ignore issues of story translation. All TDT tasks have at their core a comparison of two text models. In story link detection, the simplest case, the comparison is between pairs of stories, to decide whether given pairs of stories are on the same topic or not. In topic tracking, the comparison is between a story and a topic, which is often represented as a centroid of story vectors, or as a language model covering several stories. Our focus in this research was to explore the best ways to compare stories and topics when stories are in multiple languages. We began with the hypothesis that if two stories originated in the same language, it would be best to compare them in that language, rather than translating them both into another language for comparison . This simple assertion, which we call the native language hypothesis, is easily tested in the TDT story link detection task. The picture gets more complex in a task like topic tracking, which begins with a small number of training stories (in English) to define each topic. New stories from a stream must be placed into these topics. The streamed stories originate in different languages, but are also available in English translation. The translations have been performed automatically by machine translation algorithms, and are inferior to manual translations. At the beginning of the stream, native language comparisons cannot be performed because there are no native language topic models (other than English ). However, later in the stream, once non-English documents have been seen, one can base subsequent tracking on native-language comparisons, by adaptively training models for additional languages. There are many ways this adaptation could be performed , and we suspect that it is crucial for the first few non-English stories to be placed into topics correctly, to avoid building non-English models from off-topic stories. Previous research in multilingual TDT has not attempted to compare the building of multiple language-specific models with single -language topic models, or to obtain native-language models through adaptation. The focus of most multilingual work in TDT for example [2] [12] [13], has been to compare the efficacy of machine translation of test stories into a base language, with other means of translation. Although these researchers normalize scores for the source language, all story comparisons are done within the base language. This is also true in multilingual filtering, which is a similar task [14]. The present research is an exploration of the native language hypothesis for multilingual topic tracking. We first present results on story link detection, to support the native language hypothesis in a simple, understandable task. Then we present experiments that test the hypothesis in the topic tracking task. Finally we consider several different ways to adapt topic models to allow native language comparisons downstream. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SIGIR '04, July 25-29, 2003, Sheffield, South Yorkshire, UK. Copyright 2004 ACM 1-58113-881-4/04/0007...$5.00. 402 Although these experiments were carried out in service of TDT, the results should equally apply to other domains which require the comparison of documents in different languages, particularly filtering, text classification and clustering. EXPERIMENTAL SETUP Experiments are replicated with two different data sets, TDT3 and TDT4, and two very different similarity functions - cosine similarity , and another based on relevance modeling, described in the following two sections. Cosine similarity can be seen as a basic default approach, which performs adequately, and relevance modeling is a state of the art approach which yields top-rated performance . Confirming the native-language hypothesis in both systems would show its generality. In the rest of this section, we describe the TDT data sets, then we describe how story link detection and topic tracking are carried out in cosine similarity and relevance modeling systems. Next, we describe the multilingual aspects of the systems. 2.1 TDT3 Data TDT data consist of a stream of news in multiple languages and from different media - audio from television, radio, and web news broadcasts, and text from newswires. Two forms of transcription are available for the audio stream. The first form comes from automatic speech recognition and includes transcription errors made by such systems. The second form is a manual transcription, which has few if any errors. The audio stream can also be divided into stories automatically or manually (so-called reference boundaries). For all the research reported here, we used manual transcriptions and reference boundaries. The characteristics of the TDT3 data sets for story link detection and topic tracking are summarized in Tables 1-3. Table 1: Number of stories in TDT3 Corpus English Arabic Mandarin Total TDT3 37,526 15,928 13,657 67,111 Table 2: Characteristics of TDT3 story link detection data sets Number of topics 8 Number of link pairs Same topic Different topic English-English 605 3999 Arabic-Arabic 669 3998 Mandarin-Mandarin 440 4000 English-Arabic 676 4000 English-Mandarin 569 4000 Arabic-Mandarin 583 3998 Total 3542 23,995 Table 3: Characteristics of TDT3 topic tracking data sets N t =2 N t =4 Number of topics 36 30 Num. test stories On-topic All On-topic All English 2042 883,887 2042 796,373 Arabic 572 372,889 572 336,563 Mandarin 405 329,481 369 301,568 Total 3019 1,593,782 2983 1,434,504 2.2 Story Representation and Similarity 2.2.1 Cosine similarity To compare two stories for link detection, or a story with a topic model for tracking, each story is represented as a vector of terms with tfidf term weights: ( ) ( ) ( ) 1 log 5 . 0 log + + = N df N tf a i (1) where tf is the number of occurrences of the term in the story, N is the total number of documents in the collection, and df is the number of documents containing the term. Collection statistics N and df are computed incrementally, based on the documents already in the stream within a deferral period after the test story arrives. The deferral period was 10 for link detection and 1 for topic tracking. For link detection, story vectors were pruned to the 1000 terms with the highest term weights. The similarity of two (weighted, pruned) vectors n a a a ,..., 1 = r and m b b b ,..., 1 = r is the inner product between the two vectors: ( ) ( )( ) = i i i i i i i b a b a Sim 2 2 cos (2) If the similarity of two stories exceeds a yes/no threshold, the stories are considered to be about the same topic. For topic tracking, a topic model is a centroid, an average of the vectors for the N t training stories. Topic models are pruned to 100 terms based on the term weights. Story vectors pruned to 100 terms are compared to centroids using equation (2). If the similarity exceeds a yes/no threshold, the story is considered on-topic. 2.2.2 Relevance modeling Relevance modeling is a statistical technique for estimating language models from extremely small samples, such as queries, [9]. If Q is small sample of text, and C is a large collection of documents , the language model for Q is estimated as: ) | ( ) | ( ) | ( Q M P M w P Q w P d C d d = (3) A relevance model, then, is a mixture of language models M d of every document d in the collection, where the document models are weighted by the posterior probability of producing the query P(M d |Q). The posterior probability is computed as: = C d Q q d Q q d d M q P d P M q P d P Q M P ) | ( ) ( ) | ( ) ( ) | ( (4) Equation (4) assigns the highest weights to documents that are most likely to have generated Q, and can be interpreted as nearest-neighbor smoothing, or a massive query expansion technique. To apply relevance modeling to story link detection, we estimate the similarity between two stories A and B by pruning the stories to short queries, estimating relevance models for the queries, and measuring the similarity between the two relevance models. Each story is replaced by a query consisting of the ten words in the query with the lowest probability of occurring by chance in ran-domly drawing |A| words from the collection C: 403 = A C A A C C A C A P w w w w w chance ) ( (5) where |A| is the length of the story A, A w is the number of times word w occurs in A, |C| is the size of the collection, and C w is the number of times word w occurs in C. Story relevance models are estimated using equation (4). Similarity between relevance models is measured using the symmetrized clarity-adjusted divergence [11]: + = w A B w B A RM GE w P Q w P Q w P GE w P Q w P Q w P Sim ) | ( ) | ( log ) | ( ) | ( ) | ( log ) | ( (6) where P(w|Q A ) is the relevance model estimated for story A, and P(w|GE) is the background (General English, Arabic, or Mandarin ) probability of w, computed from the entire collection of stories in the language within the same deferral period used for cosine similarity. To apply relevance modeling to topic tracking, the asymmetric clarity adjusted divergence is used: = w track GE w P S w P T w P S T Sim ) | ( ) | ( log ) | ( ) , ( (7) where P(w|T) is a relevance model of the topic T. Because of computational constraints, smoothed maximum likelihood estimates rather than relevance models are used for the story model P(w|S). The topic model, based on Equation (3), is: = t S d d t M w P S T w P ) | ( 1 ) | ( (8) where S t is the set of training stories. The topic model is pruned to 100 terms. More detail about applying relevance models to TDT can be found in [2]. 2.3 Evaluation TDT tasks are evaluated as detection tasks. For each test trial, the system attempts to make a yes/no decision. In story link detection, the decision is whether the two members of a story pair belong to the same topic. In topic tracking, the decision is whether a story in the stream belongs to a particular topic. In all tasks, performance is summarized in two ways: a detection cost function (C Det ) and a decision error tradeoff (DET) curve. Both are based on the rates of two kinds of errors a detection system can make: misses, in which the system gives a no answer where the correct answer is yes, and false alarms, in which the system gives a yes answer where the correct answer is no. The DET curve plots the miss rate (P Miss ) as a function of false alarm rate (P Fa ), as the yes/no decision threshold is swept through its range. P Miss and P Fa are computed for each topic, and then averaged across topics to yield topic-weighted curves. An example can be seen in Figure 1 below. Better performance is indicated by curves more to the lower left of the graph. The detection cost function is computed for a particular threshold as follows: C Det = (C Miss * P Miss * P Target + C Fa * P Fa * (1-P Target )) (9) where: P Miss = #Misses / #Targets P Fa = #False Alarms / #NonTargets C Miss and C Fa are the costs of a missed detection and false alarm, respectively, and are specified for the application, usually at 10 and 1, penalizing misses more than false alarms. P Target is the a priori probability of finding a target, an item where the answer should be yes, set by convention to 0.02. The cost function is normalized: (C Det ) Norm = C Det / MIN(C Miss * C Target , C Fa * (1-P Target )) (10) and averaged over topics. Each point along the detection error tradeoff curve has a value of (C Det ) Norm . The minimum value found on the curve is known as the min(C Det ) Norm . It can be interpreted as the value of C Det ) Norm at the best possible threshold. This measure allows us to separate performance on the task from the choice of yes/no threshold. Lower cost scores indicate better performance. More information about these measures can be found in [5]. 2.4 Language-specific Comparisons English stories were lower-cased and stemmed using the kstem stemmer [6]. Stop words were removed. For native Arabic comparisons, stories were converted from Unicode UTF-8 to windows (CP1256) encoding, then normalized and stemmed with a light stemmer [7]. Stop words were removed. For native Mandarin comparisons, overlapping character bigrams were compared. STORY LINK DETECTION In this section we present experimental results for story link detection , comparing a native condition with an English baseline. In the English baseline, all comparisons are in English, using machine translation (MT) for Arabic and Mandarin stories. Corpus statistics are computed incrementally for all the English and translated-into-English stories. In the Native condition, two stories originating in the same language are compared in that language . Corpus statistics are computed incrementally for the stories in the language of the comparison. Cross language pairs in the native condition are compared in English using MT, as in the baseline. Figure 1: DET curve for TDT3 link detection based on English versions of stories, or native language versions, for cosine and relevance model similarity 404 Table 4: Min(C det ) Norm for TDT3 story link detection Similarity English Native Cosine .3440 .2586 Relevance Model .2625 .1900 Figure 1 shows the DET curves for the TDT3 story link detection task, and Table 4 shows the minimum cost. The figure and table show that native language comparisons (dotted) consistently outperform comparisons based on machine-translated English (solid). This difference holds both for the basic cosine similarity system (first row) (black curves), and for the relevance modeling system (second row) (gray curves). These results support the general conclusion that when two stories originate in the same language, it is better to carry out similarity comparisons in that language, rather than translating them into a different language. TOPIC TRACKING In tracking, the system decides whether stories in a stream belong to predefined topics. Similarity is measured between a topic model and a story, rather than between two stories. The native language hypothesis for tracking predicts better performance if incoming stories are compared in their original language with topic models in that language, and worse performance if translated stories are compared with English topic models. The hypothesis can only be tested indirectly, because Arabic and Mandarin training stories were not available for all tracking topics . In this first set of experiments, we chose to obtain native language training stories from the stream of test stories using topic adaptation, that is, gradual modification of topic models to incorporate test stories that fit the topic particularly well. Adaptation begins with the topic tracking scenario described above in section 2.2, using a single model per topic based on a small set of training stories in English. Each time a story is compared to a topic model to determine whether it should be classed as on-topic, it is also compared to a fixed adaptation threshold ad = 0.5 (not to be confused with the yes/no threshold mentioned in section 2.2.1). If the similarity score is greater than ad , the story is added to the topic set, and the topic model recomputed. For clarity, we use the phrase topic set to refer to the set of stories from which the topic model is built, which grows under adaptation . The training set includes only the original N t training stories for each topic. For cosine similarity, adaptation consists of computing a new centroid for the topic set and pruning to 100 terms. For relevance modeling, a new topic model is computed according to Equation (8). At most 100 stories are placed in each topic set. We have just described global adaptation, in which stories are added to global topic models in English. Stories that originated in Arabic or Mandarin are compared and added in their machine-translated version. Native adaptation differs from global adaptation in making separate topic models for each source language. To decide whether a test story should be added to a native topic set, the test story is compared in its native language with the native model, and added to the native topic set for that language if its similarity score exceeds ad . The English version of the story is also compared to the global topic model, and if its similarity score exceeds ad , it is added to the global topic set. (Global models continue to adapt for other languages which may not yet have a native model, or for smoothing, discussed later.) At the start there are global topic models and native English topic models based on the training stories, but no native Arabic or Mandarin topic models. When there is not yet a native topic model in the story's original language, the translated story is compared to the global topic model. If the similarity exceeds ad , the native topic model is initialized with the untranslated story. Yes/no decisions for topic tracking can then be based on the untranslated story's similarity to the native topic model if one exists. If there is no native topic model yet for that language and topic, the translated story is compared to the global topic model. We have described three experimental conditions: global adapted, native adapted, and a baseline. The baseline, described in Section 2.2, can also be called global unadapted. The baseline uses a single English model per topic based on the small set of training stories. A fourth possible condition, native unadapted is problematic and not included here. There is no straightforward way to initialize native language topic models without adaptation when training stories are provided only in English. Figure 2: DET curves for TDT3 tracking, cosine similarity (above) and relevance models (below), N t =4 training stories, global unadapted baseline, global adapted, and native adapted 405 Table 5: Min(C det ) Norm for TDT3 topic tracking. N t =2 N t =4 Adapted Adapted Baseline Global Native Baseline Global Native Cosine .1501 .1197 .1340 .1238 .1074 .1028 RM .1283 .0892 .0966 .1060 .0818 .0934 The TDT3 tracking results on three conditions, replicated with the two different similarity measures (cosine similarity and relevance modeling) and two different training set sizes (N t =2 and 4) can be seen in Table 5. DET curves for N t =4 are shown in Figure 2, for cosine similarity (above) and relevance modeling (RM) (below). Table 5 shows a robust adaptation effect for cosine and relevance model experiments, and for 2 or 4 training stories. Native and global adaptation are always better (lower cost) than baseline unadapted tracking. In addition, relevance modeling produces better results than cosine similarity. However, results do not show the predicted advantage for native adapted topic models over global adapted topic models. Only cosine similarity, N t =4, seems to show the expected difference (shaded cells), but the difference is very small. The DET curve in Figure 2 shows no sign of a native language effect. Table 6 shows minimum cost figures computed separately for English, Mandarin, and Arabic test sets. Only English shows a pattern similar to the composite results of Table 5 (see the shaded cells). For cosine similarity, there is not much difference between global and native English topic models. For relevance modeling, Native English topic models are slightly worse than global models . Arabic and Mandarin appear to show a native language advantage for all cosine similarity conditions and most relevance model conditions. However, DET curves comparing global and native adapted models separately for English, Arabic, and Mandarin , (Figure 3) show no real native language advantage. Table 6: Min(C det ) Norm for TDT3 topic tracking; breakdown by original story language English N t =2 N t =4 Adapted Adapted Baseline Global Native Baseline Global Native Cosine .1177 .0930 .0977 .0903 .0736 .0713 RM .1006 .0681 .0754 .0737 .0573 .0628 Arabic Cosine .2023 .1654 .1486 .1794 .1558 .1348 RM .1884 .1356 .1404 .1581 .1206 .1377 Mandarin Cosine .2156 .1794 .1714 .1657 .1557 .1422 RM .1829 .1272 .0991 .1286 .0935 .0847 Figure 3: DET curves for TDT3 tracking, cosine similarity, N t =4 training stories, global adapted vs. native adapted breakdown for English, Arabic, and Mandarin In trying to account for the discrepancy between the findings on link detection and tracking, we suspected that the root of the problem was the quality of native models for Arabic and Mandarin . For English, adaptation began with 2 or 4 on-topic models. However, Mandarin and Arabic models did not begin with on-topic stories; they could begin with off-topic models, which should hurt tracking performance. A related issue is data sparseness . When a native topic model is first formed, it is based on one story, which is a poorer basis for tracking than N t stories. In the next three sections we pursue different aspects of these suspicions. In section 5 we perform a best-case experiment, initializing native topic sets with on-topic stories, and smoothing native scores with global scores to address the sparseness problem. If these conditions do not show a native language advantage, we would reject the native language hypothesis. In section 6 we explore the role of the adaptation threshold. In section 7 we compare some additional methods of initializing native language topic models. ON-TOPIC NATIVE CENTROIDS In this section, we consider a best-case scenario, where we take the first N t stories in each language relevant to each topic, to initialize adaptation of native topic models. While this is cheating, and not a way to obtain native training documents in a realistic tracking scenario, it demonstrates what performance can be attained if native training documents are available. More realistic approaches to adapting native topic models are considered in subsequent sections. The baseline and global adapted conditions were carried out as in Section 4, and the native adapted condition was similar except in the way adaptation of native topics began. If there were not yet N t native stories in the topic set for the current test story in its native language, the story was added to the topic set if it was relevant. Once a native topic model had N t stories, we switched to the usual non-cheating mode of adaptation, based on similarity score and adaptation threshold. To address the data sparseness problem, we also smoothed the native similarity scores with the global similarity scores: 406 ) , ( ) 1 ( ) , ( ) , ( S T Sim S T Sim S T Sim global native smooth + = (11) The parameter was not tuned, but set to a fixed value of 0.5. The results can be seen in Table 7. Shaded cell pairs indicate confirmation of the native language hypothesis, where language-specific topic models outperform global models. Table 7: Min(C det ) Norm for TDT3 topic tracking, using N t on-topic native training stories and smoothing native scores N t =2 N t =4 Adapted Adapted Baseline Global Native Baseline Global Native Cosine .1501 .1197 .0932 .1238 .1074 .0758 Rel. .1283 .0892 .0702 .1060 .0818 .0611 Figure 4: DET curve for TDT3 tracking, initializing native adaptation with relevant training stories during adaptation, cosine similarity, N t =4 Figure 4 shows the DET curves for cosine, N t =4 case. When the native models are initialized with on-topic stories, the advantage to native models is clearly seen in the tracking performance. Figure 5: DET curve for TDT3 tracking initializing native adaptation with relevant training stories during adaptation and smoothing, vs. global adaptation, cosine similarity, N t =4, separate analyses for English, Arabic, and Mandarin. DET curves showing results computed separately for the three languages can be seen in Figure 5, for the cosine, N t =4 case. It can be clearly seen that English tracking remains about the same but the Arabic and Mandarin native tracking show a large native language advantage. ADAPTATION THRESHOLD The adaptation threshold was set to 0.5 in the experiments described above without any tuning. The increase in global tracking performance after adaptation shows that the value is at least acceptable . However, an analysis of the details of native adaptation showed that many Arabic and Mandarin topics were not adapting. A summary of some of this analysis can be seen in Table 8. Table 8: Number of topics receiving new stories during native adaptation, breakdown by language Total Topics receiving more stories Similarity N t Topics English Arabic Mandarin 2 36 24 8 11 Cosine 4 30 26 7 9 2 36 36 8 7 Relevance Model 4 30 30 8 5 Fewer than a third of the topics received adapted stories. This means that for most topics, native tracking was based on the global models. In order to determine whether this was due to the adaptation threshold, we performed an experiment varying the adaptation threshold from .3 to .65 in steps of .05. The results can be seen in Figure 6, which shows the minimum cost, min(C Det ) Norm , across the range of adaptation threshold values. Although we see that the original threshold, .5, was not always the optimal value, it is also clear that the pattern we saw at .5 (and in Figure 6) does not change as the threshold is varied, that is tracking with native topic models is not better than tracking with global models. An improperly tuned adaptation threshold was therefore not the reason that the native language hypothesis was not confirmed for tracking. We suspect that different adaptation thresholds may be needed for the different languages, but it would be better to handle this problem by language-specific normalization of similarity scores. 0.06 0.07 0.08 0.09 0.1 0.11 0.12 0.13 0.14 0.15 0.2 0.3 0.4 0.5 0.6 0.7 T hreshold Mi n C o s t Global Nt=2 Global Nt=4 Native Nt=2 Native Nt=4 Relevance Model 0.06 0.07 0.08 0.09 0.1 0.11 0.12 0.13 0.14 0.15 0.2 0.3 0.4 0.5 0.6 0.7 T hreshold Mi n C o s t Global Nt=2 Global Nt=4 Native Nt=2 Native Nt=4 Cosine Similarity Figure 6: Effect of adaptation threshold on min(C Det ) Norm on TDT3 tracking with adaptation. IMPROVING NATIVE TOPIC MODELS In the previous two sections we showed that when native topic models are initialized with language specific training stories that are truly on-topic, then topic tracking is indeed better with native models than with global models. However, in context of the TDT 407 test situation, the way we obtained our language-specific training stories was cheating. In this section we experiment with 2 different "legal" ways to initialize better native language models: (1) Use both global and native models, and smooth native similarity scores with global similarity scores. (2) Initialize native models with dictionary or other translations of the English training stories into the other language. Smoothing was carried out in the native adapted condition according to Equation (11), setting =0.5, without tuning. The comparison with unadapted and globally adapted tracking can be seen in Table 9. The smoothing improves the native topic model performance relative to unsmoothed native topic models (cf. Table 5), and brings the native model performance to roughly the same level as the global. In other words, smoothing improves performance , but we still do not have strong support for the native language hypothesis. This is apparent in Figure 7. Native adapted tracking is not better than global adapted tracking. Table 9: Min(C det ) Norm for TDT3 topic tracking, smoothing native scores with global scores N t =2 N t =4 Adapted Adapted Baseline Global Native Smooth Baseline Global Native Smooth Cosine .1501 .1197 .1125 .1238 .1074 .1010 RM .1283 .0892 .0872 .1060 .0818 .0840 Figure 7: DET curve for TDT3 tracking with smoothing, cosine similarity, N t =4 training stories The final method of initializing topic models for different languages would be to translate the English training stories into the other languages required. We did not have machine translation from English into Arabic or Mandarin available for these experiments . However, we have had success with dictionary translations for Arabic. In [2] we found that dictionary translations from Arabic into English resulted in comparable performance to the machine translations on tracking, and better performance on link detection. Such translated stories would not be "native language" training stories, but might be a better starting point for language-specific adaptation anyway. Training story translations into Arabic used an English/Arabic probabilistic dictionary derived from the Linguistic Data Consor-tium's UN Arabic/English parallel corpus, developed for our cross-language information retrieval work [7]. Each English word has many different Arabic translations, each with a translation probability p(a|e). The Arabic words, but not the English words, have been stemmed according to a light stemming algorithm. To translate an English story, English stop words were removed, and each English word occurrence was replaced by all of its dictionary translations, weighted by their translation probabilities. Weights were summed across all the occurrences of each Arabic word, and the resulting Arabic term vector was truncated to retain only terms above a threshold weight. We translated training stories only into Arabic, because we did not have a method to produce good quality English to Mandarin translation. The results for Arabic can be seen in Table 10. For translation, it makes sense to include an unadapted native condition, labeled translated in the table. Table 10: Min(C det ) Norm for Arabic TDT3 topic tracking, initializing native topic models with dictionary-translated training stories Arabic N t =2 Unadapted Adapted Baseline Translated Global Native Cosine .2023 .2219 .1694 .2209 RM .1884 .1625 .1356 .1613 Arabic N t =4 Cosine .1794 .1640 .1558 1655 RM .1581 .1316 .1206 .1325 Figure 8: DET curve for TDT3 tracking initializing native topics with dictionary-translated training stories, cosine similarity, N t =4, Arabic only The results are mixed. First of all, this case is unusual in that adaptation does not improve translated models. Further analysis revealed that very little adaptation was taking place. Because of this lack of native adaptation, global adaptation consistently outperformed native adaptation here. However, in the unadapted conditions, translated training stories outperformed the global models for Arabic in three of the four cases - cosine N t =4 and relevance models for N t =2 and N t =4 (the shaded baseline-trans-408 lated pairs in Table 10). The DET curve for the cosine N t =4 case can be seen in Figure 8. The native unadapted curve is better (lower) than the global unadapted curve. The translated stories were very different from the test stories, so their similarity scores almost always fell below the adaptation threshold. We believe the need to normalize scores between native stories and dictionary translations is part of the problem, but we also need to investigate the compatibility of the dictionary translations with the native Arabic stories. CONCLUSIONS We have confirmed the native language hypothesis for story link detection. For topic tracking, the picture is more complicated. When native language training stories are available, good native language topic models can be built for tracking stories in their original language. Smoothing the native models with global models improves performance slightly. However, if training stories are not available in the different languages, it is difficult to form native models by adaptation or by translation of training stories, which perform better than the adapted global models. Why should language specific comparisons be more accurate than comparisons based on machine translation? Machine translations are not always good translations. If the translation distorts the meaning of the original story, it is unlikely to be similar to the topic model, particularly if proper names are incorrect, or spelled differently in the machine translations than they are in the English training stories, a common problem in English translations from Mandarin or Arabic. Secondly, even if the translations are correct, the choice of words, and hence the language models, are likely to be different across languages. The second problem could be han-dled by normalizing for source language, as in [12]. But normalization cannot compensate for poor translation. We were surprised that translating the training stories into Arabic to make Arabic topic models did not improve tracking, but again, our dictionary based translations of the topic models were different from native Arabic stories. We intend to try the same experiment with manual translations of the training stories into Arabic and Mandarin. We are also planning to investigate the best way to normalize scores for different languages. When TDT4 relevance judgments are available we intend to replicate some of these experiments on TDT4 data. ACKNOWLEDGMENTS This work was supported in part by the Center for Intelligent Information Retrieval and in part by SPAWARSYSCEN-SD grant number N66001-02-1-8903. Any opinions, findings and conclusions or recommendations expressed in this material are the author (s) and do not necessarily reflect those of the sponsor. REFERENCES [1] Allan, J. Introduction to topic detection and tracking. In Topic detection and tracking: Event-based information organization , J. Allan (ed.): Kluwer Academic Publishers, 1-16 , 2002. [2] Allan, J. Bolivar, A., Connell, M., Cronen-Townsend, S., Feng, A, Feng, F., Kumaran, G., Larkey, L., Lavrenko, V., Raghavan, H. UMass TDT 2003 Research Summary. In Proceedings of TDT 2003 evaluation, unpublished, 2003. [3] Chen, H.-H. and Ku, L. W. An NLP & IR approach to topic detection. In Topic detection and tracking: Event-based information organization, J. Allan (ed.). Boston, MA: Kluwer , 243-264, 2002. [4] Chen, Y.-J. and Chen, H.-H. Nlp and IR approaches to monolingual and multilingual link detection. Presented at Proceedings of 19th International Conference on Computa-tional Linguistics, Taipei, Taiwan, 2002. [5] Fiscus, J. G. and Doddington, G. R. Topic detection and tracking evaluation overview. In Topic detection and tracking: Event-based information organization, J. Allan (ed.). Boston, MA: Kluwer, 17-32, 2002. [6] Krovetz, R. Viewing morphology as an inference process. In Proceedings of SIGIR '93, 191-203, 1993. [7] Larkey, Leah S. and Connell, Margaret E. (2003) Structured Queries, Language Modeling, and Relevance Modeling in Cross-Language Information Retrieval. To appear in Information Processing and Management Special Issue on Cross Language Information Retrieval, 2003. [8] Larkey, L. S., Ballesteros, L., and Connell, M. E. Improving stemming for Arabic information retrieval: Light stemming and co-occurrence analysis. In Proceedings of SIGIR 2002, 275-282, 2002. [9] Lavrenko, V. and Croft, W. B. Relevance-based language models. In Proceedings of SIGIR 2001. New Orleans: ACM, 120-127, 2001. [10] Lavrenko, V. and Croft, W. B. Relevance models in information retrieval. In Language modeling for information retrieval, W. B. Croft and J. Lafferty (eds.). Boston: Kluwer, 11-56, 2003. [11] Lavrenko, V., Allan, J., DeGuzman, E., LaFlamme, D., Pollard , V., and Thomas, S. Relevance models for topic detection and tracking. In Proceedings of the Conference on Human Language Technology, 104-110, 2002. [12] Leek, T., Schwartz, R. M., and Sista, S. Probabilistic approaches to topic detection and tracking. In Topic detection and tracking: Event-based information organization, J. Allan (ed.). Boston, MA: Kluwer, 67-83, 2002. [13] Levow, G.-A. and Oard, D. W. Signal boosting for translin-gual topic tracking: Document expansion and n-best translation . In Topic detection and tracking: Event-based information organization, J. Allan (ed.). Boston, MA: Kluwer, 175-195, 2002. [14] Oard, D. W. Adaptive vector space text filtering for monolingual and cross-language applications. PhD dissertation , University of Maryland, College Park, 1996. http://www.glue.umd.edu/~dlrg/filter/papers/thesis.ps.gz 409

topic models;;classification;crosslingual;native topic models;similarity;story link;topic tracking;native language hypothesis;multilingual topic tracking;multilingual;Arabic;TDT;machine translation

122

Lazy Preservation: Reconstructing Websites by Crawling the Crawlers

Backup of websites is often not considered until after a catastrophic event has occurred to either the website or its webmaster. We introduce "lazy preservation" digital preservation performed as a result of the normal operation of web crawlers and caches. Lazy preservation is especially suitable for third parties; for example, a teacher reconstructing a missing website used in previous classes. We evaluate the effectiveness of lazy preservation by reconstructing 24 websites of varying sizes and composition using Warrick, a web-repository crawler. Because of varying levels of completeness in any one repository, our reconstructions sampled from four different web repositories: Google (44%), MSN (30%), Internet Archive (19%) and Yahoo (7%). We also measured the time required for web resources to be discovered and cached (10-103 days) as well as how long they remained in cache after deletion (7-61 days).

INTRODUCTION "My old web hosting company lost my site in its entirety (duh!) when a hard drive died on them. Needless to say that I was peeved, but I do notice that it is available to browse on the wayback machine... Does anyone have any ideas if I can download my full site?" - A request for help at archive.org [25] Websites may be lost for a number of reasons: hard drive crashes, file system failures, viruses, hacking, etc. A lost website may be restored if care was taken to create a backup beforehand, but sometimes webmasters are negligent in backing up their websites, and in cases such as fire, flooding, or death of the website owner, backups are frequently unavailable . In these cases, webmasters and third parties may turn to the Internet Archive (IA) "Wayback Machine" for help. According to a representative from IA, they have performed over 200 website recoveries in the past year for various individuals . Although IA is often helpful, it is strictly a best-effort approach that performs sporadic, incomplete and slow crawls of the Web (IA is at least 6 months out-of-date [16]). Another source of missing web content is in the caches of search engines (SEs) like Google, MSN and Yahoo that scour the Web looking for content to index. Unfortunately, the SEs do not preserve canonical copies of all the web resources they cache, and it is assumed that the SEs do not keep web pages long after they have been removed from a web server. We define lazy preservation as the collective digital preservation performed by web archives and search engines on behalf of the Web at large. It exists as a preservation service on top of distributed, incomplete, and potentially unreliable web repositories. Lazy preservation requires no individual effort or cost for Web publishers, but it also provides no quality of service guarantees. We explore the effectiveness of lazy preservation by downloading 24 websites of various sizes and subject matter and reconstructing them using a web-repository crawler named Warrick 1 which recovers missing resources from four web repositories (IA, Google, MSN and Yahoo). We compare the downloaded versions of the sites with the reconstructed versions to measure how successful we were at reconstructing the websites. We also measure the time it takes for SEs to crawl and cache web pages that we have created on .com and .edu websites . In June 2005, we created four synthetic web collections consisting of HTML, PDF and images. For 90 days we systematically removed web pages and measured how long they remained cached by the SEs. BACKGROUND AND RELATED WORK The ephemeral nature of the Web has been widely ac-knowledged . To combat the disappearance of web resources, Brewster Kahle's Internet Archive has been archiving the 1 Warrick is named after a fictional forensic scientist with a penchant for gambling. 67 Table 1: Web repository-supported data types Type G Y M IA HTML C C C C Plain text M M M C GIF, PNG, JPG M M R C JavaScript M M C MS Excel M S M C MS PowerPoint M M M C MS Word M M M C PDF M M M C PostScript M S C C = Canonical version is stored M = Modified version is stored (image thumbnails or HTML conversions) R = Stored but not retrievable with direct URL S = Indexed but stored version is not accessible Web since 1996 [4]. National libraries are also actively engaged in archiving culturally important websites [8]. Systems like LOCKSS [24] have been developed to ensure libraries have long-term access to publishers' web content, and commercial systems like Spurl.net and HanzoWeb.com have been developed to allow users to archive selected web resources that they deem important. Other researchers have developed tools for archiving individual websites and web pages. InfoMonitor [7] archives a website's file system and stores the archive remotely. TTA-pache [9] is used to archive requested pages from a particular web server, and iPROXY [23] is used as a proxy server to archive requested pages from a variety of web servers. In many cases these services can be of some value for recovering a lost website, but they are largely useless when backups are inaccessible or destroyed or when a third party wants to reconstruct a website. They also require the webmaster to perform some amount of work in setting up, configuring and monitoring the systems. In regards to commercial search engines, the literature has mostly focused on measuring the amount of content they have indexed (e.g., [15, 18]), relevance of responses to users' queries (e.g., [5, 14]), and ranking of pages (e.g., [28]). Lewandowski et al. [17] studied how frequently Google, MSN and Yahoo updated their cached versions of web pages, but we are unaware of any research that attempts to measure how quickly new resources are added to and removed from commercial SE caches, or research that explores the use of SE caches for reconstructing websites. WEB CRAWLING AND CACHING There are many SEs and web archives that index and store Web content. For them to be useful for website reconstruction , they must at a minimum provide a way to map a given URL to a stored resource. To limit the implemen-tation complexity, we have focused on what we consider to be the four most popular web repositories that meet our minimum criteria. Recent measurements show that Google, MSN and Yahoo index significantly different portions of the Web and have an intersection of less than 45% [15]. Adding additional web repositories like ask.com, gigablast.com, in-cywincy .com and any other web repository that allows direct URL retrieval would likely increase our ability to reconstruct websites. t d t a t r t p TTL c SE cache t m TTL ws t 0 vulnerable replicated endangered unrecoverable Web server Figure 1: Timeline of SE resource acquisition and release Although SEs often publish index size estimates, it is difficult to estimate the number of resources in each SE cache. An HTML web page may consist of numerous web resources (e.g., images, applets, etc.) that may not be counted in the estimates, and not all indexed resources are stored in the SE caches. Google, MSN and Yahoo will not cache an HTML page if it contains a NOARCHIVE meta-tag, and the http Cache-control directives `no-cache' and `no-store' may also prevent caching of resources [1]. Only IA stores web resources indefinitely. The SEs have proprietary cache replacement and removal policies which can only be inferred from observed behavior. All four web repositories perform sporadic and incomplete crawls of websites making their aggregate performance important for website reconstruction. Table 1 shows the most popular types of resources held by the four web repositories. This table is based on our observations when reconstructing websites with a variety of content. IA keeps a canonical version of all web resources, but SEs only keep canonical versions of HTML pages. When adding PDF, PostScript and Microsoft Office (Word, Excel, PowerPoint) resources to their cache, the SEs create HTML versions of the resources which are stripped of all images. SEs also keep only a thumbnail version of the images they cache due to copyright law. MSN uses Picsearch for their image crawling; unfortunately, Picsearch and MSN do not support direct URL queries for accessing these images, so they cannot be used for recovering website images. 3.2 Lifetime of a Web Resource Figure 1 illustrates the life span of a web resource from when it is first made available on a web server to when when it is finally purged from a SE cache. A web resource's time-to -live on the web server (TTL ws ) is defined as the number of days from when the resource is first made accessible on the server ( t 0 ) to when it is removed ( t r ). A new resource is vulnerable until it is discovered by a SE ( t d ) and made available in the SE cache ( t a ). The resource is replicated when it is accessible on the web server and in cache. Once the resource is removed from the web server ( t r ), it becomes endangered since it is only accessible in cache. When a subsequent crawl reveals the resource is no longer available on the web server ( t m ), it will then be purged from cache ( t p ) and be unrecoverable. The period between t a and t p define a resource's time-to-live in the SE cache (TTL c ). A resource is recoverable if it is currently cached (i.e., is replicated or endangered). A recoverable resource can only be recovered during the TTL c period with a probability of P r , the observed number of days that a resource is retrievable from the cache divided by TTL c . 68 It should be noted that the TTL ws and TTL c values of a resource may not necessarily overlap. A SE that is trying to maximize the freshness of its index will try to minimize the difference between TTL ws and TTL c . A SE that is slow in updating its index, perhaps because it obtains crawling data from a third party, may experience late caching where t r &lt; t a . For a website to be lazily preserved, we would like its resources to be cached soon after their appearance on a website (have minimal vulnerability). SEs may also share this goal if they want to index newly discovered content as quickly as possible. Inducing a SE to crawl a website at a specific time is not currently possible. Webmasters may employ various techniques to ensure their websites are crawler-friendly [13, 27] and well connected to the Web. They may even submit their website URLs to SEs or use proprietary mechanisms like Google's Sitemap Protocol [12], but no technique will guarantee immediate indexing and caching of a website. We would also like resources to remain cached long after they have been deleted from the web server (remain endangered ) so they can be recovered for many days after their disappearance. SEs on the other hand may want to minimize the endangered period in order to purge missing content from their index. Just as we have no control as to when a SE crawler will visit, we also have no control over cache eviction policies. 3.3 Web Collection Design In order to obtain measurements for TTL c and other values in Figure 1, we created four synthetic web collections and placed them on websites for which we could obtain crawling data. We deployed the collections in June 2005 at four different locations: 1) www.owenbrau.com, 2) www.cs.odu. edu/ fmccown/lazy/ 3) www.cs.odu.edu/ jsmit/, and 4) www.cs.odu.edu/ mln/lazyp/. The .com website was new and had never been indexed by Google, Yahoo or MSN. The 3 .edu websites had existed for over a year and had been previously crawled by all three SEs. In order for the web collections to be found by the SEs, we placed links to the root of each web collection from the .edu websites, and we submitted owenbrau's base URL to Google, MSN and Yahoo 1 month prior to the experiment. For 90 days we systematically removed resources from each collection. We examined the server web logs to determine when resources were crawled, and we queried Google, MSN and Yahoo daily to determine when the resources were cached. We organized each web collection into a series of 30 update bins (directories) which contained a number of HTML pages referencing the same three inline images (GIF, JPG and PNG) and a number of PDF files. An index.html file (with a single inline image) in the root of the web collection pointed to each of the bins. An index.html file in each bin pointed to the HTML pages and PDF files so a web crawler could easily find all the resources. All these files were static and did not change throughout the 90 day period except the index.html files in each bin which were modified when links to deleted web pages were removed. In all, there were 381 HTML files, 350 PDF files, and 223 images in each web collection. More detail about the organization of the web collections and what the pages and images looked like can be found in [20, 26]. The PDF and HTML pages were made to look like typical web pages with around 120 words per page. The text for each page was randomly generated from a standard English dictionary. By using random words we avoided creating duplicate pages that a SE may reject [6]. Unfortunately, using random words may cause pages to be flagged as spam [10]. Each HTML and PDF page contained a unique identifier (UID) at the top of each page (e.g., `mlnODULPT2 dgrp18 pg18-2-pdf' that included 4 identifiers: the web collection (e.g., `mlnODULPT2' means the `mln' collection), bin number (e.g., `dgrp18' means bin 18), page number and resource type (e.g., `pg18-2-pdf' means page number 2 from bin 18 and PDF resource). The UID contains spaces to allow for more efficient querying of the SE caches. The TTL ws for each resource in the web collection is a function of its bin number b and page number p: TTL ws = b( 90/b - p + 1) (1) 3.4 Daily SE Queries In designing our daily SE queries, care was taken to perform a limited number of daily queries to not overburden the SEs. We could have queried the SEs using the URL for each resource, but this might have led to our resources being cached prematurely; it is possible that if a SE is queried for a URL it did not index that it would add the URL to a list of URLs to be crawled at a later date. This is how IA's advanced search interface handles missing URLs from users' queries. To determine which HTML and PDF resources had been cached, we queried using subsets of the resources' UIDs and looked for cached URLs in the results pages. For example , to find PDF resources from the mln collection, we queried each SE to return the top 100 PDF results from the site www.cs.odu.edu that contain the exact phrase `mlnODULPT2 dgrp18'. 2 It is necessary to divulge the site in the query or multiple results from the site will not be returned . Although this tells the SE on which site the resource is located, it does not divulge the URL of the resource. To query for cached images, we queried for the globally unique filename given to each image. 3.5 Crawling and Caching Observations Although the web server logs registered visits from a variety of crawlers, we report only on crawls from Google, Inktomi (Yahoo) and MSN. 3 Alexa Internet (who provides crawls to IA) only accessed our collection once (induced through our use of the Alexa toolbar). A separate IA robot accessed less than 1% of the collections, likely due to several submissions we made to their Wayback Machine's advanced search interface early in the experiment. Further analysis of the log data can seen in a companion paper [26]. We report only detailed measurements on HTML resources (PDF resources were similar). Images were crawled and cached far less frequently; Google and Picsearch (the MSN Images provider) were the only ones to crawl a significant number of images. The 3 .edu collections had 29% of their images crawled, and owenbrau had 14% of its images crawled. Only 4 unique images appeared in Google Images, all from 2 MSN only allows limiting the results page to 50. 3 Due to a technical mishap beyond our control, we were unable to obtain crawling data for days 41-55 for owebrau and parts of days 66-75 and 97 for the .edu web collections. We were also prevented from making cache queries on days 53, 54, 86 and 87. 69 Table 2: Caching of HTML resources from 4 web collections (350 HTML resources in each collection) Web % URLs crawled % URLs cached t ca T T L c / P r Endangered collection G M Y G M Y G M Y G M Y G M Y fmccown 91 41 56 91 16 36 13 65 47 90 / 0.78 20 / 0.87 35 / 0.57 51 9 24 jsmit 92 31 92 92 14 65 12 66 47 86 / 0.82 20 / 0.91 36 / 0.55 47 7 25 mln 94 33 84 94 14 49 10 65 54 87 / 0.83 21 / 0.90 24 / 0.46 47 8 19 owenbrau 18 0 0 20 0 0 103 N/A N/A 40 / 0.98 N/A N/A 61 N/A N/A Ave 74 26 58 74 11 37 35 66 50 76 / 0.86 20 / 0.89 32 / 0.53 51 8 23 Figure 2: Crawling (top) and caching (bottom) of HTML resources from the mln web collection the mln collection. Google likely used an image duplication detection algorithm to prevent duplicate images from different URLs from being cached. Only one image (from fmccown) appeared in MSN Images. None of the cached images fell out of cache during our experiment. Table 2 summarizes the performance of each SE to crawl and cache 350 HTML resources from each of the four web collections. This table does not include index.html resources which had an infinite T T L ws . We believe there was an error in the MSN query script which caused fewer resources to be found in the MSN cache, but the percentage of crawled URLs provides an upper bound on the number of cached resources; this has little to no effect on the other measurements reported. The three SEs showed equal desire to crawl HTML and PDF resources. Inktomi (Yahoo) crawled 2 times as many resources as MSN, and Google crawled almost 3 times as many resources than MSN. Google was the only SE to crawl and cache any resources from the new owenbrau website. From a preservation perspective, Google out-performed MSN and Yahoo in nearly every category. Google cached the highest percentage of HTML resources (76%) and took only 12 days on average to cache new resources from the edu web collections. On average, Google cached HTML resources for the longest period of time (76 days), consistently provided access to the cached resources (86%), and were the slowest to remove cached resources that were deleted from the web server (51 days). Although Yahoo cached more HTML resources and kept the resources cached for a longer period than MSN, the probability of accessing a resource on any given day was only 53% compared to 89% for MSN. Figure 2 provides an interesting look at the crawling and caching behavior of Google, Yahoo and MSN. These graphs illustrate the crawling and caching of HTML resources from the mln collection; the other two edu collections exhibited similar behavior. The resources are sorted by TTL ws with the longest-living resources appearing on the bottom. The index.html files which were never removed from the web collection have an infinite TTL (`inf'). The red diagonal line indicates the decay of the web collection; on any particular day, only resources below the red line were accessible from the web server. On the top row of Figure 2, blue dots indicate resources that were crawled on a particular day. When resources were requested that had been deleted, the web server responded with a 404 (not found) code represented by green dots above the red line. The bottom row of Figure 2 shows the cached HTML resources (blue) resulting from the crawls. Some pages in Yahoo were indexed but not cached (green). As Figure 2 illustrates, both Google and MSN were quick to make resources available in their cache soon after they were crawled, and they were quick to purge resources from their cache when a crawl revealed the resources were no longer available on the web server. A surprising finding is that many of the HTML resources that were previously purged from Google's cache reappeared on day 102 and remained cached for the remainder of our experiment. The other two edu collections exhibited similar behavior for HTML resources. HTML and PDF resources from owenbrau appeared in the Google cache on day 102 for the first time; these resources had been deleted from the web server 10-20 days before day 102. Manual inspection weeks after the experiment had concluded revealed that the pages remained in Google's cache and fell out months later. Yahoo was very sporadic in caching resources; there was often a lag time of 30 days between the crawl of a resource and its appearance in cache. Many of the crawled resources never appeared in Yahoo's cache. Although Inktomi crawled nearly every available HTML resource on day 10, only half of those resources ever became available in the Yahoo cache. We have observed through subsequent interaction with Yahoo that links to cached content may appear and disappear when performing the same query just a few seconds apart. This likely accounts for the observed cache inconsistency. We have observed from our measurements that nearly all new HTML and PDF resources that we placed on known websites were crawled and cached by Google several days af-70 A A D B C E F G B' C' E F added 20% W ' W changed 33% identical 50% missing 17% Figure 3: Lost website (left), reconstructed website (center), and reconstruction diagram (right) ter they were discovered. Resources on a new website were not cached for months. Yahoo and MSN were 4-5 times slower than Google to acquire new resources, and Yahoo incurs a long transfer delay from Inktomi's crawls into their cache. We have also observed that cached resources are often purged from all three caches as soon as a crawl reveals the resources are missing, but in the case of Google, many HTML resources have reappeared weeks after being removed. Images tend to be largely ignored. Search engines may crawl and cache other websites differ-ently depending on a variety of factors including perceived level of importance (e.g., PageRank) and modification rates. Crawling policies may also be changed over time. This experiment merely provides a glimpse into the current caching behavior of the top three SEs that has not been documented before. Our findings suggest that SEs vary greatly in the level of access they provide to cached resources, and that websites are likely to be reconstructed more successfully if they are reconstructed quickly after being lost. Reconstructions should also be performed several days in a row to ensure maximum access to web repository holdings. In some cases, it may even be beneficial to attempt recovering resources even a month after they have been lost. RECONSTRUCTING WEBSITES We define a reconstructed website to be the collection of recovered resources that share the same URIs as the resources from a lost website or from some previous version of the lost website [19]. The recovered resources may be equivalent to, or very different from, the lost resources. For websites that are composed of static files, recovered resources would be equivalent to the files that were lost. For sites produced dynamically using CGI, PHP, etc., the recovered resources would match the client's view of the resources and would be useful to the webmaster in rebuilding the server-side components. The server-side components are currently not recoverable using lazy preservation (see Section 5). To quantify the difference between a reconstructed website and a lost website, we classify the recovered resources from the website graphs. A website can be represented as a graph G = (V, E) where each resource r i (HTML, PDF, image, etc.), identified by a URI, is a node v i , and there exists a directed edge from v i to v j when there is a hyperlink or reference from r i to r j . The left side of Figure 3 shows a web graph for some website W if we began to crawl it starting at A. Suppose W was lost and reconstructed forming the website W represented in the center of Figure 3. For each resource r i in W we may examine its corresponding resource r i in W that shares the same URI and categorize r i as identical ( r i is byte-for-byte identical to r i ), changed ( r i is not identical to r i ), or missing ( r i could not be found in any web). We would categorize those resources in W that did not share a URI with any resource in W as added ( r i was not a part of the current website but was recovered due to a reference from r j ). Figure 3 shows that resources A, G and E were reconstructed and are identical to their lost versions. An older version of B was found (B') that pointed to G, a resource that does not currently exist in W . Since B' does not reference D, we did not know to recover it (it is possible that G is actually D renamed). An older version of C was found, and although it still references F, F could not be found in any web repository. A measure of change between the lost website W and the reconstructed website W can be described using the following difference vector: difference( W, W ) = R changed |W | , R missing |W | , R added |W | (2) For Figure 3, the difference vector is (2/6, 1/6, 1/5) = (0.333, 0.167, 0.2). The best case scenario would be (0,0,0), the complete reconstruction of a website. A completely unrecoverable website would have a difference vector of (0,1,0). The difference vector for a reconstructed website can be illustrated as a reconstruction diagram as shown on the right side of Figure 3. The changed, identical and missing resources form the core of the reconstructed website. The dark gray portion of the core grows as the percentage of changed resource increases. The hole in the center of the core grows as the percentage of missing resources increases. The added resources appear as crust around the core. This representation will be used later in Table 3 when we report on the websites we reconstructed in our experiments. 4.2 Warrick Operation Warrick, our web-repository crawler, is able to reconstruct a website when given a base URL pointing to where the site used to exist. The web repositories are crawled by issuing queries in the form of URLs to access their stored holdings . For example, Google's cached version of http://foo. edu/page1.html can be accessed like so: http://search. google.com/search?q=cache:http://foo.edu/page1.html. If Google has not cached the page, an error page will be generated . Otherwise the cached page can be stripped of any Google-added HTML, and the page can be parsed for links to other resources from the foo.edu domain (and other domains if necessary). Most repositories require two or more queries to obtain a resource. For each URL, the file extension (if present) is examined to determine if the URL is an image (.png, .gif, .jpg, etc.) or other resource type. All three SEs use a different method for retrieving images than for other resource types. IA has the same interface regardless of the type. We would have better accuracy at determining if a given URL referenced an image or not if we knew the URL's resource MIME type, but this information is not available to us. IA is the first web repository queried by Warrick because it keeps a canonical version of all web resources. When querying for an image URL, if IA does not have the image then Google and Yahoo are queried one at a time until one of them returns an image. Google and Yahoo do not publicize the cached date of their images, so it is not possible to pick the most recently cached image. 71 Table 3: Results of website reconstructions MIME type groupings (orig/recovered) Difference vector Website PR Total HTML Images Other (Changed, Missing, Added) Recon diag Almost identical New recon diag 1. www.eskimo.com/~scs/ 6 719/691 96% 696/669 96% 22/21 95% 1/1 100% (0.011, 0.039, 0.001) 50% 2. www.digitalpreservation.gov 8 414/378 91% 346/329 95% 42/25 60% 26/24 92% (0.097, 0.087, 0.000) 44% 3. www.harding.edu/hr/ 4 73/47 64% 19/19 100% 25/2 8% 29/26 90% (0.438, 0.356, 0.145) 83% 4. www.techlocker.com 4 1216/406 33% 687/149 22% 529/257 49% 0/0 (0.267, 0.666, 0.175) 99% If a non-image resource is being retrieved, again IA is queried first. If IA has the resource and the resource does not have a MIME type of `text/html', then the SEs are not queried since they only store canonical versions of HTML resources. If the resource does have a `text/html' MIME type (or IA did not have a copy), then all three SEs are queried, the cache dates of the resources are compared (if available), and the most recent resource is chosen. Warrick will search HTML resources for URLs to other resources and add them to the crawl frontier (a queue). Resources are recovered in breadth-first order, and reconstruction continues until the frontier is empty. All recovered resources are stored on the local filesystem, and a log is kept of recovered and missing resources. Warrick limits its requests per day to the web repositories based on their published API values (Google, 1000; Yahoo, 5000; MSN, 10,000) or lacking an API, our best guess (IA, 1000). If any repository's limit is exceeded, Warrick will checkpoint and sleep for 24 hours. 4.3 Reconstruction Experiment and Results To gauge the effectiveness of lazy preservation for website reconstruction, we compared the snap-shot of 24 live websites with their reconstructions. We chose sites that were either personally known to us or randomly sampled from dmoz.org. The websites (some were actually subsites) were predominantly English, covered a range of topics, and were from a number of top-level domains. We chose 8 small ( &lt;150 URIs), 8 medium (150-499 URIs) and 8 large ( 500 URIs) websites, and we avoided websites that used robots.txt and Flash exclusively as the main interface. In August 2005 we downloaded all 24 websites by starting at the base URL and following all links and references that that were in and beneath the starting directory, with no limit to the path depth. For simplicity, we restricted the download to port 80 and did not follow links to other hosts within the same domain name. So if the base URL for the website was http://www.foo.edu/bar/, only URLs matching http: //www.foo.edu/bar/* were downloaded. Warrick uses the same default setting for reconstructing websites. Immediately after downloading the websites, we reconstructed five different versions for each of the 24 websites: four using each web repository separately, and one using all web repositories together. The different reconstructions helped to show how effective individual web repositories could reconstruct a website versus the aggregate of all four web repositories. We present 4 of the 24 results of the aggregate reconstructions in Table 3, ordered by percent of recovered URIs. The complete results can be seen in [20]. The `PR' column is Google's PageRank (0-10 with 10 being the most important) for the root page of each website at the time of the experiments. (MSN and Yahoo do not publicly disclose their `importance' metric.) For each website, the total number of resources in the website is shown along with the total number of resources that were recovered and the percentage. Resources are also totalled by MIME type. The difference vector for the website accounts for recovered files that were added. The `Almost identical' column of Table 3 shows the percentage of text-based resources (e.g., HTML, PDF, PostScript , Word, PowerPoint, Excel) that were almost identical to the originals. The last column shows the reconstruction figure for each website if these almost identical resources are moved from the `Changed' category to `Identical' category. We considered two text-based resources to be almost identical if they shared at least 75% of their shingles of size 10. Shingling (as proposed by Broder et al. [3]) is a popular method for quantifying similarity of text documents when word-order is important [2, 11, 21]. We did not use any image similarity metrics. We were able to recover more than 90% of the original resources from a quarter of the 24 websites. For three quarters of the websites we recovered more than half of the resources. On average we were able to recover 68% of the website resources (median=72%). Of those resources recovered, 30% of them on average were not byte-for-byte identical. A majority (72%) of the `changed' text-based files were almost identical to the originals (having 75% of their shingles in common). 67% of the 24 websites had obtained additional files when reconstructed which accounted for 7% of the total number of files reconstructed per website. When all website resources are aggregated together and examined, dynamic pages (those that contained a `?' in the URL) were significantly less likely to be recovered than resources that did not have a query string (11% vs. 73%). URLs with a path depth greater than three were also less likely to be recovered (52% vs. 61%). A chi-square analysis confirms the significance of these findings (p &lt; .001). We were unable to find any correlation between percentage of recovered resources with PageRank or website size. The success of recovering resources based on their MIME type is plotted in Figure 4. The percentage of resources 72 0 25 50 75 100 125 150 175 200 225 html images pdf other ms MIME type groups Nu m b er o f reso u r ces 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Ave # of resources in original websites Aggregate % recon IA % recon Google % recon MSN % recon Yahoo! % recon Figure 4: Recovery success by MIME type that were recovered from the five different website reconstructions we performed (one using all four web repositories , and four using each web repository individually) are shown along with the average number of resources making up the 24 downloaded (or original) websites. A majority (92%) of the resources making up the original websites are HTML and images. We were much more successful at recovering HTML resources than images; we recovered 100% of the HTML resources for 9 of the websites (38%) using all four web repositories. It is likely we recovered fewer images because MSN cannot be used to recover images, and as our caching experiment revealed, images are also much less likely to be cached than other resource types. Figure 4 also emphasizes the importance of using all four web repositories when reconstructing a website. By just using IA or just using Google, many resources will not be recovered. This is further illustrated by Figure 5 which shows the percentage of each web repository's contribution in the aggregate reconstructions (sites are ordered by number of URIs). Although Google was the largest overall contributor to the website reconstructions (providing 44% of the resources) they provided none of the resources for site 17 and provided less than 30% of the resources for 9 of the reconstructions. MSN contributed on average 30% of the resources; IA was third with 19%, and Yahoo was last with a 7% contribution rate. Yahoo's poor contribution rate is likely due to their spotty cache access as exhibited in our caching experiment (Figure 2) and because last-modified datestamps are frequently older than last-cached datestamps (Warrick chooses resources with the most recent datestamps). The amount of time and the number of queries required to reconstruct all 24 websites (using all 4 repositories) is shown in Figure 6. Here we see almost a 1:1 ratio of queries to seconds. Although the size of the original websites gets larger along the x-axis, the number of files reconstructed and the number of resources held in each web repository determine how many queries are performed. In none of our reconstructions did we exceed the daily query limit of any of the web repositories. FUTURE WORK We have made Warrick available on the Web 4 , and it has been used to reconstruct several websites have been lost due to fire, hard-drive crashes, death of the website owner, 4 http://www.cs.odu.edu/ fmccown/warrick/ 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Reconstructed w ebsites C ont r i but i o n Yahoo IA MSN Google Figure 5: Web repositories contributing to each website reconstruction 0 1000 2000 3000 4000 5000 6000 7000 1 2 3 4 5 6 7 8 9 10 11 12 13 14 1516 17 18 19 20 2122 23 24 Reconstructed W ebsites N u m b e r of que r i e s 0 1000 2000 3000 4000 5000 6000 7000 Ti m e ( s e c ) queries time Figure 6: Number of queries performed and time taken to reconstruct websites hacking, and discontinued charitable website hosting [19]. Although the reconstructions have not been complete, individuals are very thankful to have recovered any resources at all when faced with total loss. There are numerous improvements we are making to Warrick including an API for easier inclusion of new web repositories and new methods for discovering more resources within a web repository [19]. We are planning on reconstructing a larger sample from the Web to discover the website characteristics that allow for more effective "lazy recovery". Discovering such characteristics will allow us to create guidelines for webmasters to ensure better lazy preservation of their sites. Our next experiment will take into account rate of change and reconstruction differences over time. We are also interested in recovering the server-side components (CGI programs, databases, etc.) of a lost website. We are investigating methods to inject server-side components into indexable content using erasure codes (popular with RAID systems [22]) so they can be recovered from web repositories when only a subset of pages can be found. A web-repository crawler could be used in the future to safeguard websites that are at risk of being lost. When a website is detected as being lost, a reconstruction could be initiated to preserve what is left of the site. Additionally, websites in countries that are targeted by political censorship could be reconstructed at safe locations. CONCLUSIONS Lazy preservation is a best-effort, wide-coverage digital preservation service that may be used as a last resort when 73 website backups are unavailable. It is not a substitute for digital preservation infrastructure and policy. Web repositories may not crawl orphan pages, protected pages (e.g., robots.txt, password, IP), very large pages, pages deep in a web collection or links influenced by JavaScript, Flash or session IDs. If a web repository will not or cannot crawl and cache a resource, it cannot be recovered. We have measured the ability of Google, MSN and Yahoo to cache four synthetic web collections over a period of four months. We measured web resources to be vulnerable for as little as 10 days and in the worst case, as long as our 90 day test period. More encouragingly, many HTML resources were recoverable for 851 days on average after being deleted from the web server. Google proved to be the most consistent at caching our synthetic web collections. We have also used our web-repository crawler to reconstruct a variety of actual websites with varying success. HTML resources were the most numerous (52%) type of resource in our collection of 24 websites and were the most successfully recoverable resource type (89% recoverable). Images were the second most numerous (40%) resource type, but they were less successfully recovered (53%). Dynamic pages and resources with path depths greater than three were less likely to be recovered. Google was the most frequent source for the reconstructions (44%), but MSN was a close second (30%), followed by IA (19%) and Yahoo (7%). The probability of reconstruction success was not correlated with Google's PageRank or the size of the website. REFERENCES [1] H. Berghel. Responsible web caching. Communications of the ACM, 45(9):1520, 2002. [2] K. Bharat and A. Broder. Mirror, mirror on the web: a study of host pairs with replicated content. In Proceedings of WWW '99, pages 15791590, 1999. [3] A. Z. Broder, S. C. Glassman, M. S. Manasse, and G. Zweig. Syntactic clustering of the Web. Computer Networks & ISDN Systems, 29(8-13):11571166, 1997. [4] M. Burner. Crawling towards eternity: Building an archive of the world wide web. Web Techniques Magazine, 2(5), 1997. [5] F. Can, R. Nuray, and A. B. Sevdik. Automatic performance evaluation of web search engines. Info. Processing & Management, 40(3):495514, 2004. [6] J. Cho, N. Shivakumar, and H. Garcia-Molina. Finding replicated web collections. In Proceedings of SIGMOD '00, pages 355366, 2000. [7] B. F. Cooper and H. Garcia-Molina. Infomonitor: Unobtrusively archiving a World Wide Web server. International Journal on Digital Libraries, 5(2):106119, April 2005. [8] M. Day. Collecting and preserving the World Wide Web. 2003. http: //library.wellcome.ac.uk/assets/WTL039229.pdf. [9] C. E. Dyreson, H. Lin, and Y. Wang. Managing versions of web documents in a transaction-time web server. In Proceedings of WWW '04, pages 422432, 2004. [10] D. Fetterly, M. Manasse, and M. Najork. Spam, damn spam, and statistics: using statistical analysis to locate spam web pages. In Proceedings of WebDB '04, pages 16, 2004. [11] D. Fetterly, M. Manasse, M. Najork, and J. Wiener. A large-scale study of the evolution of web pages. In Proceedings of WWW '03, pages 669678, 2003. [12] Google Sitemap Protocol, 2005. http://www.google. com/webmasters/sitemaps/docs/en/protocol.html. [13] Google webmaster help center: Webmaster guidelines, 2006. http://www.google.com/support/webmasters/ bin/answer.py?answer=35769. [14] M. Gordon and P. Pathak. Finding information on the World Wide Web: the retrieval effectiveness of search engines. Inf. Process. Manage., 35(2):141180, 1999. [15] A. Gulli and A. Signorini. The indexable web is more than 11.5 billion pages. In Proceedings of WWW '05, pages 902903, May 2005. [16] Internet Archive FAQ: How can I get my site included in the Archive?, 2006. http://www.archive.org/about/faqs.php. [17] D. Lewandowski, H. Wahlig, and G. Meyer-Beautor. The freshness of Web search engine databases. Journal of Information Science, 32(2):131148, Apr 2006. [18] F. McCown, X. Liu, M. L. Nelson, and M. Zubair. Search engine coverage of the OAI-PMH corpus. IEEE Internet Computing, 10(2):6673, Mar/Apr 2006. [19] F. McCown and M. L. Nelson. Evaluation of crawling policies for a web-repository crawler. In Proceedings of HYPERTEXT '06, pages 145156, 2006. [20] F. McCown, J. A. Smith, M. L. Nelson, and J. Bollen. Reconstructing websites for the lazy webmaster. Technical report, Old Dominion University, 2005. http://arxiv.org/abs/cs.IR/0512069. [21] A. Ntoulas, J. Cho, and C. Olston. What's new on the Web? The evolution of the Web from a search engine perspective. In Proceedings of WWW '04, pages 112, 2004. [22] J. S. Plank. A tutorial on Reed-Solomon coding for fault-tolerance in RAID-like systems. Software: Practice and Experience, 27(9):9951012, 1997. [23] H. C. Rao, Y. Chen, and M. Chen. A proxy-based personal web archiving service. SIGOPS Operating Systems Review, 35(1):6172, 2001. [24] V. Reich and D. S. Rosenthal. LOCKSS: A permanent web publishing and access system. D-Lib Magazine, 7(6), 2001. [25] A. Ross. Internet Archive forums: Web forum posting. Oct 2004. http://www.archive.org/iathreads/ post-view.php?id=23121. [26] J. A. Smith, F. McCown, and M. L. Nelson. Observed web robot behavior on decaying web subsites. D-Lib Magazine, 12(2), Feb 2006. [27] M. Weideman and M. Mgidana. Website navigation architectures and their effect on website visibility: a literature survey. In Proceedings of SAICSIT '04, pages 292296, 2004. [28] J. Zhang and A. Dimitroff. The impact of webpage content characteristics on webpage visibility in search engine results (part I). Information Processing & Management, 41(3):665690, 2005. 74

Search engines (SEs);cached resources;web repositories;recovery;reconstruction;crawling;caching;lazy preservation;search engine;digital preservation

123

Learning Concepts from Large Scale Imbalanced Data Sets Using Support Cluster Machines

This paper considers the problem of using Support Vector Machines (SVMs) to learn concepts from large scale imbalanced data sets. The objective of this paper is twofold. Firstly, we investigate the effects of large scale and imbalance on SVMs. We highlight the role of linear non-separability in this problem. Secondly, we develop a both practical and theoretical guaranteed meta-algorithm to handle the trouble of scale and imbalance. The approach is named Support Cluster Machines (SCMs). It incorporates the informative and the representative under-sampling mechanisms to speedup the training procedure. The SCMs differs from the previous similar ideas in two ways, (a) the theoretical foundation has been provided, and (b) the clustering is performed in the feature space rather than in the input space. The theoretical analysis not only provides justification , but also guides the technical choices of the proposed approach. Finally, experiments on both the synthetic and the TRECVID data are carried out. The results support the previous analysis and show that the SCMs are efficient and effective while dealing with large scale imbalanced data sets.

INTRODUCTION In the context of concept modelling, this paper considers the problem of how to make full use of the large scale annotated data sets. In particular, we study the behaviors of Support Vector Machines (SVMs) on large scale imbalanced data sets, not only because its solid theoretical foundations but also for its empirical success in various applications. 1.1 Motivation Bridging the semantic gap has been becoming the most challenging problem of Multimedia Information Retrieval (MIR). Currently, there are mainly two types of methods to bridge the gap [8]. The first one is relevance feedback which attempts to capture the user's precise needs through iterative feedback and query refinement. Another promising direction is concept modelling. As noted by Hauptmann [14], this splits the semantic gap between low level features and user information needs into two, hopefully smaller gaps: (a) mapping the low-level features into the intermediate semantic concepts and (b) mapping these concepts into user needs. The automated image annotation methods for CBIR and the high level feature extraction methods in CBVR are all the efforts to model the first mapping. Of these methods, supervised learning is one of the most successful ones. An early difficulty of supervised learning is the lack of annotated training data. Currently, however, it seems no longer a problem. This is due to both the techniques developed to leverage surrounding texts of web images and the large scale collaborative annotation. Actually, there is an underway effort named Large Scale Concept Ontology for Multimedia Understanding (LSCOM), which intends to annotate 1000 concepts in broadcast news video [13]. The initial fruits of this effort have been harvested in the practice of TRECVID hosted by National Institute of Standards and Technology (NIST) [1]. In TRECVID 2005, 39 concepts are annotated by multiple participants through web collaboration, and ten of them are used in the evaluation. The available large amount of annotated data is undoubt-edly beneficial to supervised learning. However, it also brings out a novel challenge, that is, how to make full use of the data while training the classifiers. On the one hand, the annotated data sets are usually in rather large scale. The de-441 velopment set of TRECVID 2005 includes 74523 keyframes. The data set of LSCOM with over 1000 annotated concepts might be even larger. With all the data, the training of SVMs will be rather slow. On the other hand, each concept will be the minority class under one-against-all strategy . Only a small portion of the data belong to the concept, while all the others are not (In our case, the minority class always refers to the positive class). The ratio of the positive examples and the negative ones is typically below 1 : 100 in TRECVID data. These novel challenges have spurred great interest in the communities of data mining and machine learning[2, 6, 21, 22, 29]. Our first motivation is to investigate the effects of large scale and imbalance on SVMs. This is critical for correct technical choices and development. The second objective of this paper is to provide a practical as well as theoretical guaranteed approach to addressing the problem. 1.2 Our Results The major contribution of this paper can be summarized as follows: 1. We investigate the effects of large scale and imbalance on SVMs and highlight the role of linear non-separability of the data sets. We find that SVMs has no difficulties with linear separable large scale imbalanced data. 2. We establish the relations between the SVMs trained on the centroids of the clusters and the SVMs obtained on the original data set. We show that the difference between their optimal solutions are bounded by the perturbation of the kernel matrix. We also prove the optimal criteria for approximating the original optimal solutions. 3. We develop a meta-algorithm named Support Cluster Machines (SCMs). A fast kernel k-means approach has been employed to partition the data in the feature space rather than in the input space. Experiments on both the synthetic data and the TRECVID data are carried out. The results support the previous analysis and show that the SCMs are efficient and effective while dealing with large scale imbalanced data sets. 1.3 Organization The structure of this paper is as follows. In Section 2 we give a brief review of SVMs and kernel k-means. We discuss the effects of the large scale imbalanced data on SVMs in Section 3. We develop the theoretical foundations and present the detailed SCMs approach in Section 4. In Section 5 we carry out experiments on both the synthetic and the TRECVID data sets. Finally, we conclude the paper in Section 6. PRELIMINARIES Here, we present a sketch introduction to the soft-margin SVMs for the convenience of the deduction in Section 4. For a binary classification problem, given a training data set D of size n D = {(x i , y i ) |x i R N , y i {1, -1}}, where x i indicates the training vector of the ith sample and y i indicates its target value, and i = 1, . . . , n. The classification hyperplane is defined as w, (x) + b = 0, where ( ) is a mapping from R N to a (usually) higher dimension Hilbert space H, and , denotes the dot product in H. Thus, the decision function f(x) is f (x) = sign( w, (x) + b). The SVMs aims to find the hyperplane with the maximum margin between the two classes, i.e., the optimal hyperplane. This can be obtained by solving the following quadratic optimization problem min w,b, 1 2 w 2 + C n i=1 i subject to y i ( w, (x i ) + b) 1 i (1) i 0, i = 1, . . . , n. With the help of Lagrange multipliers, the dual of the above problem is min G() = 1 2 T Q - e T subject to 0 i C, i = 1, . . . , n (2) T y = 0, where is a vector with components i that are the Lagrange multipliers, C is the upper bound, e is a vector of all ones, and Q is an n n positive semi-definite matrix, Q ij = y i y j (x i ), (x j ) . Since the mapping ( ) only appears in the dot product, therefore, we need not know its explicit form. Instead, we define a kernel K( , ) to calculate the dot product, i.e., K(x i , x j ) = (x i ), (x j ) . The matrix K with components K(x i , x j ) is named Gram Matrix (or kernel matrix). With kernel K , , we can implicitly map the training data from input space to a feature space H. 2.2 Kernel k -means and Graph Partitioning Given a set of vectors x 1 , . . . , x n , the standard k-means algorithm aims to find clusters 1 , . . . , k that minimize the objective function J( { c } k c=1 ) = k c=1 x i c x i - m c 2 , (3) where { c } k c=1 denotes the partitioning of the data set and m c = xic x i | c | is the centroid of the cluster c . Similar to the idea of nonlinear SVMs, the k-means can also be performed in the feature space with the help of a nonlinear mapping ( ), which results in the so-called kernel k-means J( { c } k c=1 ) = k c=1 x i c (x i ) - m c 2 , (4) where m c = xic (x i ) | c | . If we expand the Euclidean distance (x i ) - m c 2 in the objective function, we can find that the image of x i only appears in the form of dot product . Thus, given a kernel matrix K with the same meaning 442 in SVMs, we can compute the distance between points and centroids without knowing explicit representation of (x i ). Recently, an appealing alternative, i.e., the graph clustering has attracted great interest. It treats clustering as a graph partition problem. Given a graph G = ( V, E, A), which consists of a set of vertices V and a set of edges E such that an edge between two vertices represents their similarity. The affinity matrix A is |V||V| whose entries represent the weights of the edges. Let links( V 1 , V 2 ) be the sum of the edge weights between the nodes in V 1 and V 2 , that is links( V 1 , V 2 ) = iV 1 ,jV 2 A ij . Ratio association is a type of graph partitioning objective which aims to maximize within-cluster association relative to the size of the cluster RAssoc(G) = max V 1 ,...,V k k c=1 links( V c , V c ) |V c | . (5) The following theorem establishes the relation between kernel k-means and graph clustering [10]. With this result, we can develop some techniques to handle the difficulty of storing the large kernel matrix for kernel k-means. Theorem 1. Given a data set, we can construct a weighted graph G = ( V, E, A), by treating each sample as a node and linking an edge between each other. If we define the edge weight A ij = K(x i , x j ), that is, A = K, the minimization of (4) is equivalent to the maximization of (5). THE EFFECTS OF LARGE SCALE IM-BALANCED DATA ON SVMS There are two obstacles yielded by large scale. The first one is the kernel evaluation, which has been intensively discussed in the previous work. The computational cost scales quadratically with the data size. Furthermore, it is impossible to store the whole kernel matrix in the memory for common computers. The decomposition algorithms (e.g., SMO) have been developed to solve the problem [20, 22]. The SMO-like algorithms actually transform the space load to the time cost, i.e., numerous iterations until convergence. To reduce or avoid the kernel reevaluations, various efficient caching techniques are also proposed [16]. Another obstacle caused by large scale is the increased classification difficulty , that is, the more probable data overlapping. We can not prove it is inevitable but it typically happens. Assume we will draw n randomly chosen numbers between 1 to 100 from a uniform distribution, our chances of drawing a number close to 100 would improve with increasing values of n, even though the expected mean of the draws is invariant [2]. The checkerboard experiment in [29] is an intuitive example . This is true especially for the real world data, either because of the weak features (we mean features that are less discriminative) or because of the noises. With the large scale data, the samples in the overlapping area might be so many that the samples violating KKT conditions become abundant. This means the SMO algorithm might need more iterations to converge. Generally, the existing algorithmic approaches have not been able to tackle the very large data set. Whereas, the under-sampling method, e.g., active learning, is possible. With unlabelled data, active learning selects a well-chosen subset of data to label so that reduce the labor of manual annotations [24]. With large scale labelled data, active learning can also be used to reduce the scale of training data [21]. The key issue of active learning is how to choose the most "valuable" samples. The informative sampling is a popular criterion. That is, the samples closest to the boundary or maximally violating the KKT conditions (the misclassified samples) are preferred [24, 26]. Active learning is usually in an iterative style. It requires an initial (usually random selected) data set to obtain the estimation of the boundary. The samples selected in the following iterations depend on this initial boundary. In addition, active learning can not work like the decomposition approach which stops until all the samples satisfy the KKT conditions. This imply a potential danger, that is, if the initial data are selected improperly, the algorithm might not be able to find the suitable hyperplane . Thus, another criterion, i.e., representative, must be considered. Here, "representative" refers to the ability to characterize the data distribution. Nguyen et al. [19] show that the active learning method considering the representative criterion will achieve better results. Specifically for SVMs, pre-clustering is proposed to estimate the data distribution before the under-sampling [31, 3, 30]. Similar ideas of representative sampling appear in [5, 12]. 3.2 The Imbalanced Data The reason why general machine learning systems suffer performance loss with imbalanced data is not yet clear [23, 28], but the analysis on SVMs seems relatively straightforward . Akbani et al. have summarized three possible causes for SVMs [2]. They are, (a) positive samples lie further from the ideal boundary, (b) the weakness of the soft-margin SVMs, and (c) the imbalanced support vector ratio. Of these causes, in our opinion, what really matters is the second one. The first cause is pointed out by Wu et al. [29]. This situation occurs when the data are linearly separable and the imbalance is caused by the insufficient sampling of the minority class. Only in this case does the "ideal" boundary make sense. As for the third cause, Akbani et al. have pointed out that it plays a minor role because of the constraint T y = 0 on Lagrange multipliers [2]. The second cause states that the soft-margin SVMs has inherent weakness for handling imbalanced data. We find that it depends on the linear separability of the data whether the imbalance has negative effects on SVMs. For linearly separable data, the imbalance will have tiny effects on SVMs, since all the slack variables of (1) tend to be zeros (, unless the C is so small that the maximization of the margin dominates the objective). In the result, there is no contradiction between the capacity of the SVMs and the empirical error . Unfortunately, linear non-separable data often occurs. The SVMs has to achieve a tradeoff between maximizing the margin and minimizing the empirical error. For imbalanced data, the majority class outnumbers the minority one in the overlapping area. To reduce the overwhelming errors of misclassifying the majority class, the optimal hyperplane will inevitably be skew to the minority. In the extreme, if C is not very large, SVMs simply learns to classify everything as negative because that makes the "margin" the largest, with zero cumulative error on the abundant negative examples . The only tradeoff is the small amount of cumulative 443 error on the few positive examples, which does not count for much. Several variants of SVMs have been adopted to solve the problem of imbalance. One choice is the so-called one-class SVMs, which uses only positive examples for training. Without using the information of the negative samples, it is usually difficult to achieve as good result as that of binary SVMs classifier [18]. Using different penalty constants C + and C for the positive and negative examples have been reported to be effective [27, 17]. However, Wu et al. point out that the effectiveness of this method is limited [29]. The explanation of Wu is based on the KKT condition T y = 0, which imposes an equal total influence from the positive and negative support vectors. We evaluate this method and the result shows that tuning C + C does work (details refer to Section 5). We find this also depends on the linear separability of the data whether this method works. For linearly separable data, tuning C + C has little effects, since the penalty constants are useless with the zero-valued slack variables. However, if the data are linearly non-separable, tuning C + C does change the position of separating hyperplane. The method to modify the kernel matrix is also proposed to improve SVMs for imbalanced data [29]. A possible drawback of this type approach is its high computational costs. OVERALL APPROACH The proposed approach is named Support Cluster Machines (SCMs). We first partition the negative samples into disjoint clusters, then train an initial SVMs model using the positive samples and the representatives of the negative clusters. With the global picture of the initial SVMs, we can approximately identify the support vectors and non-support vectors. A shrinking technique is then used to remove the samples which are most probably not support vectors. This procedure of clustering and shrinking are performed itera-tively several times until some stop criteria satisfied. With such a from coarse-to-fine procedure, the representative and informative mechanisms are incorporated. There are four key issues in the meta-algorithm of SCMs: (a) How to get the partition of the training data, (b) How to get the representative for each cluster, (c) How to safely remove the non-support vector samples, (d) When to stop the iteration procedure. Though similar ideas have been proposed to speed-up SVMs in [30, 3, 31], no theoretical analysis of this idea has been provided. In the following, we present an in-depth analysis for this type of approaches and attempt to improve the algorithm under the theoretical guide. 4.1 Theoretical Analysis Suppose { c } k c=1 is a partition of the training set that the samples within the same cluster have the same class label. If we construct a representative u c for each cluster c , we can obtain two novel models of SVMs. The first one is named Support Cluster Machines (SCMs). It treats each representative as a sample, thus the data size is reduced from n to k. This equals to the classification of the clusters. That is where the name SCMs come from. The new training set is D = {(u c , y c ) |u c R N , y c {1, -1}, c = 1, . . . , k}, in which y c equals the labels of the samples within c . We define the dual problem of support cluster machines as min G ( ) = 1 2 T Q - e T subjectto 0 i | i |C, i = 1, . . . , k (6) T y = 0, where is a vector of size k with components i corresponding to u i , | i |C is the upper bound for i , e is a k dimension vector of all ones, and Q is an k k positive semi-definite matrix, Q ij = y i y j (u i ), (u j ) . Another one is named Duplicate Support Vector Machines (DSVMs). Different from SCMs, it does not reduce the size of training set. Instead, it replace each sample x i with the representative of the cluster that x i belongs to. Thus, the samples within the same cluster are duplicate. That is why it is named DSVMs. The training set is ~ D = {(~x i , ~ y i ) |x i D, if x i c , ~ x i = u c and ~ y i = y i }, and the corresponding dual problem is defined as min ~ G() = 1 2 T ~ Q - e T subjectto 0 i C, i = 1, . . . , n (7) T y = 0, where ~ Q is is an n n positive semi-definite matrix, ~ Q ij = ~ y i ~ y j (~ x i ), (~ x j ) . We have the following theorem that states (6) is somehow equivalent to (7): Theorem 2. With the above definitions of the SCMs and the DSVMs, if and are their optimal solutions respectively , the relation G ( ) = ~ G( ) holds. Furthermore, any R k satisfying { c = x i c i , c = 1, . . . , k} is the optimal solution of SCMs. Inversely, any R n satisfying { x i c i = c , c = 1, . . . , k} and the constraints of (7) is the optimal solution of DSVMs. The proof is in Appendix A. Theorem 2 shows that solving the SCMs is equivalent to solving a quadratic programming problem of the same scale as that of the SVMs in (2). Comparing (2) and (7), we can find that only the Hessian matrix is different. Thus, to estimate the approximation from SCMs of (6) to SVMs of (2), we only need to analyze the stability of the quadratic programming model in (2) when the Hessian matrix varies from Q to ~ Q. Daniel has presented a study on the stability of the solution of definite quadratic programming, which requires that both Q and ~ Q are positive definite [7]. However, in our situation, Q is usually positive definite and ~ Q is not (because of the duplications). We develop a novel theorem for this case. If define = Q - ~ Q , where denotes the Frobenius norm of a matrix, the value of measure the size of the perturbations between Q and ~ Q. We have the following theorem: Theorem 3. If Q is positive definite and = Q - ~ Q , let and ~ be the optimal solutions to (2) and (7) respectively , we have ~ ~ mC G( ~ ) - G( ) (m 2 + ~ m 2 )C 2 2 444 where is the minimum eigenvalue of Q, m and ~ m indicate the numbers of the support vectors for (2) and (7) respectively . The proof is in Appendix B. This theorem shows that the approximation from (2) to (7) is bounded by . Note that this does not mean that with minimal we are sure to get the best approximate solution. For example, adopting the support vectors of (1) to construct ~ Q will yield the exact optimal solution of (2) but the corresponding are not necessarily minimum. However, we do not know which samples are support vectors beforehand. What we can do is to minimize the potential maximal distortion between the solutions between (2) and (7). Now we consider the next problem, that is, given the partition { c } k c=1 , what are the best representatives {u c } k c=1 for the clusters in the sense of approximating Q? In fact, we have the following theorem: Theorem 4. Given the partition { c } k c=1 , the {u c } k c=1 satisfying (u c ) = x i c (x i ) | c | , c = 1, . . . , k (8) will make = Q - ~ Q minimum. The proof is in Appendix C. This theorem shows that, given the partition, (u c ) = m c yields the best approximation between ~ Q and Q. Here we come to the last question, i.e., what partition { c } k c=1 will make = Q - ~ Q minimum. To make the problem more clearly, we expand 2 as Q - ~ Q 2 = k h=1 k l=1 x i h x j l ( (x i ), (x j ) - m h , m l ) 2 . (9) There are approximately k n /k! types of such partitions of the data set. An exhaustive search for the best partition is impossible. Recalling that (9) is similar to (4), we have the following theorem which states their relaxed equivalence. Theorem 5. The relaxed optimal solution of minimizing (9) and the relaxed optimal solution of minimizing (4) are equivalent. The proof can be found in Appendix D. Minimizing amounts to find a low-rank matrix approximating Q. Ding et al. have pointed out the relaxed equivalence between kernel PCA and kernel k-means in [11]. Note that minimizing (9) is different from kernel PCA in that it is with an additional block-wise constant constraint. That is, the value of ~ Q ij must be invariant with respect to the cluster h containing ~ x i and the cluster l containing ~ x j . With Theorem 5 we know that kernel k-means is a suitable method to obtain the partition of data. According to the above results, the SCMs essentially finds an approximate solution to the original SVMs by smoothing the kernel matrix K (or Hessian matrix Q). Fig.1 illustrates the procedure of smoothing the kernel matrix via clustering. Hence, by solving a smaller quadratic programming problem , the position of separating hyperplane can be roughly determined. -5 0 5 10 15 -4 -2 0 2 4 6 8 10 12 14 16 50 100 150 200 20 40 60 80 100 120 140 160 180 200 50 100 150 200 20 40 60 80 100 120 140 160 180 200 50 100 150 200 20 40 60 80 100 120 140 160 180 200 (a) (b) (c ) (d) Figure 1: (a) 2D data distribution, (b) the visualization of the kernel matrix Q, (c) the kernel matrix Q by re-ordering the entries so that the samples belonging to the same cluster come together, (d) the approximate kernel matrix ~ Q obtained by replacing each sample with the corresponding centroid. 4.2 Kernel-based Graph Clustering In the previous work, k-means [30], BIRCH [31] and PDDP [3] have been used to obtain the partition of the data. None of them performs clustering in the feature space, though the SVMs works in the feature space. This is somewhat unnatural . Firstly, recalling that the kernel K( , ) usually implies an implicitly nonlinear mapping from the input space to the feature space, the optimal partition of input space is not necessarily the optimal one of feature space. Take k-means as an example, due to the fact that the squared Euclidean distance is used as the distortion measure, the clusters must be separated by piece-wise hyperplanes (i.e., voronoi diagram). However, these separating hyperplanes are no longer hyperplanes in the feature space with nonlinear mapping ( ). Secondly, the k-means approach can not capture the complex structure of data. As shown in Fig.2, the negative class is in a ring-shape in the input space. If the k-means is used, the centroids of positive and negative class might overlap. Whereas in the feature space, the kernel k-means might get separable centroids. Several factors limit the application of kernel k-means to large scale data. Firstly, it is almost impossible to store the whole kernel matrix K in the memory, e.g., for n = 100 000, we still need 20 gigabytes memory taking the symmetry into account. Secondly, the kernel k-means relies heavily on an effective initialization to achieve good results, and we do not have such a sound method yet. Finally, the computational cost of the kernel k-means might exceeds that of SVMs, and therefore, we lose the benefits of under-sampling. Dhillon et al. recently propose a multilevel kernel k-means method [9], which seems to cater to our requirements. The approach is based on the equivalence between graph clustering and kernel k-means. It incorporates the coarsening and initial partitioning phases to obtain a good initial clustering. Most importantly, the approach is extremely efficient. It can handle a graph with 28,294 nodes and 1,007,284 edges in several seconds. Therefore, here we adopt this approach. The detailed description can be found in [9]. In the following, we focus on how to address the difficulty of storing large scale kernel matrix. Theorem 1 states that kernel k-means is equivalent to a type of graph clustering. Kernel k-means focuses on grouping data so that their average distance from the centroid is minimum ,while graph clustering aims to minimizing the average pair-wise distance among the data. Central grouping and pair-wise grouping are two different views of the same approach. From the perspective of pair-wise grouping, we can expect that two samples with large distance will not belong to the same cluster in the optimal solution. Thus, 445 1 x 2 x 1 z 2 z 3 z Positive Class Negative Class Figure 2: The left and right figures show the data distribution of input space and feature space respectively . The two classes are indicated by squares and circles. Each class is grouped into one cluster, and the solid mark indicates the centroid of the class. we add the constraint that two samples with distance large enough are not linked by an edge, that is, transforming the dense graph to a sparse graph. This procedure is the common practice in spectral clustering or manifold embedding. Usually, two methods have been widely used for this purpose , i.e., k-nearest neighbor and -ball. Here, we adopt the -ball approach. Concretely, the edges with weight A ij &lt; is removed from the original graph, in which the parameter is pre-determined. By transforming a dense graph into a sparse graph, we only need store the sparse affinity matrix instead of the original kernel matrix. Nevertheless, we have to point out that the time complexity of constructing sparse graph is O(n 2 ) for data set with n examples, which is the efficiency bottleneck of the current implementation. With the sparse graph, each iteration of the multilevel kernel k-means costs O(ln ) time, where ln is the number of nonzero entries in the kernel matrix. 4.3 Support Cluster Machines According to Theorem 4, choosing the centroid of each cluster as representative will yield the best approximation. However, the explicit form of ( ) is unknown. We don't know the exact pre-images of {m c } k c=1 , what we can get are the dot products between the centroids by m h , m l = 1 | h || l | x i h x j l (x i ), (x j ) , which requires O(n 2 ) costs. Then the pre-computed kernel SVMs can be used. The pre-computed kernel SVMs takes the kernel matrix K as input, and save the indices of support vectors in the model [15]. To classify the incoming sample x, we have to calculate the dot product between x and all the samples in the support clusters, e.g., c (If m c is a support vector, we define the cluster c as support cluster.) x, m c = 1 | c | x i c x, x i . We need another O(nm) costs to predict all the training samples if there are m samples in support clusters. This is unacceptable for large scale data. To reduce the kernel reevaluation, we adopt the same method as [3], i.e., selecting a pseudo-center for each cluster as the representative u c = arg min x i c (x i ) - 1 | c | x j c (x j ) 2 , 1 x Positive class Negative class 2 x 1 x Positive class Negative class 2 x (a) (b) Figure 3: (a) Each class is grouped into one cluster, (b) each class is grouped into two clusters. The solid mark represents the centroid of the corresponding class. The solid lines indicate the support hyperplanes yielded by SCMs and the dot lines indicate the true support hyperplanes. which can be directly obtained by u c = arg max x i c x j c (x i ), (x j ) . (10) Thus, the kernel evaluation within training procedure requires O( k c=1 | c | 2 + k 2 ) time, which be further reduced by probabilistic speedup proposed by Smola [25]. The kernel evaluation of predicting the training samples is reduced from O(nm) to O(ns), where s indicates the number of support clusters. 4.4 Shrinking Techniques With the initial SCMs, we can remove the samples that are not likely support vectors. However, there is no theoretical guarantee for the security of the shrinking. In Fig. 3, we give a simple example to show that the shrinking might not be safe. In the example, if the samples outside the margin between support hyperplanes are to be removed, the case (a) will remove the true support vectors while the case (b) will not. The example shows that the security depends on whether the hyperplane of SCMs is parallel to the true separating hyperplane. However, we do not know the direction of true separating hyperplane before the classification. Therefore, what we can do is to adopt sufficient initial cluster numbers so that the solution of SCMs can approximate the original optimal solution enough. Specifically for large scale imbalanced data, the samples satisfying the following condition will be removed from the training set: | w, (x) + b| &gt; , (11) where is a predefined parameter. 4.5 The Algorithm Yu [31] and Boley [3] have adopted different stop criteria . In Yu et al.'s approach, the algorithm stops when each cluster has only one sample. Whereas, Boley et al. limit the maximum iterations by a fixed parameter. Here, we propose two novel criteria especially suitable for imbalanced data. The first one is to stop whenever the ratio of positive and negative samples is relatively imbalanced. Another choice is the Neyman-Pearson criterion, that is, minimizing the total error rate subject to a constraint that the miss rate of positive class is less than some threshold. Thus, once the 446 miss rate of positive class exceeds some threshold, we stop the algorithm. The overall approach is illustrated in Algorithm 1. With large scale balanced data, we carry out the data clustering for both classes separately. Whereas with imbalanced data, the clustering and shrinking will only be conducted on the majority class. The computation complexity is dominated by kernel evaluation. Therefore, it will not exceed O((n ) 2 + (n + ) 2 ), where n and n + indicate the number of negative and positive examples respectively. Algorithm 1: Support Cluster Machines Input : Training data set D = D + D Output : Decision function f repeat 1 { + c , m + c } k + c=1 =KernelKMeans( D + ) 2 { c , m c } k c=1 =KernelKMeans( D ) 3 D = {m + c } k + c=1 {m + c } k c=1 4 f =SVMTrain( D ) 5 f ( D) =SVMPredict(f , D) 6 D = D + D =Shrinking (f ( D)); 7 until stop criterion is true 8 EXPERIMENTS The experiments on both the synthetic and the TRECVID data are carried out. The experiments on synthetic data are used to analyze the effects of large scale and imbalance on SVMs and the experiments on TRECVID data serve to evaluate the effectiveness and efficiency of SCMs. The multilevel kernel graph partitioning code graclus [9] is adopted for data clustering and the well-known LibSVM software [15] is used in our experiments. All our experiments are done in a Pentium 4 3.00GHz machine with 1G memory. 5.1 Synthetic Data Set We generate two-dimensional data for the convenience of observation. Let x is a random variable uniformly distributed in [0, ]. The data are generated by D + = {(x, y)|y =sin(x)-+0.7[rand(0, 1)-1], x [0, ]} D = {(x, y)|y =- sin(x)+1+0.7rand(0, 1), x [0, ]}, where rand(0, 1) generates the random numbers uniformly distributed between 0 and 1, and is a parameter controlling the overlapping ratio of the two classes. Fig. 4 and Fig. 5 show some examples of the synthetic data. We use the linear kernel function in all the experiments on synthetic data. 5.1.1 The Effects of Scale We generate two types of balanced data, i.e., n + = n , but one ( D 1 = D( = 1.5)) is linearly separable and the Table 1: The effects of scale and overlapping on the time costs of training SVMs (in seconds). n + + n 200 2000 4000 8000 20000 40000 80000 time( D 1 ) 0.01 0.03 0.04 0.07 0.23 0.63 1.32 time( D 2 ) 0.02 0.70 3.24 14.01 58.51 201.07 840.60 0 0.5 1 1.5 2 2.5 3 3.5 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 Positive class Negative class 0 0.5 1 1.5 2 2.5 3 3.5 -1.5 -1 -0.5 0 0.5 1 1.5 2 Positive class Negative class (a) (b) Figure 4: (a) example of non-overlapped balanced data sets, (b) example of overlapped balanced data sets. 0 0.5 1 1.5 2 2.5 3 3.5 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 Positive class Negative class 0 0.5 1 1.5 2 2.5 3 3.5 -1.5 -1 -0.5 0 0.5 1 1.5 2 Positive class Negative class (a) (b) Figure 5: (a) example of non-overlapped imbalanced data sets, (b) example of overlapped imbalanced data sets. other ( D 2 = D( = 0.6)) is not, as shown in Fig.4. We observe the difference of the behaviors of time costs for D 1 and D 2 when the scale increases. With the same parameter settings, the time costs of optimizing the objective for D 1 and D 2 are shown in Table 1, from which we can get two conclusions, (a) time costs increase with the scale, and (b) in the same scale, the linearly non-separable data will cost more time to converge. 5.1.2 The Effects of Imbalance We generate two types of imbalanced data, i.e., n + n , but one ( D 1 = D( = 1.5)) is linearly separable and the other ( D 2 = D( = 0.6)) is not, as shown in Fig.5. We observe the difference of the effects of imbalance for linearly separable data D 1 and linearly non-separable D 2 . For the space limitation, we will not describe the detailed results here but only present the major conclusions. For linearly separable data, SVMs can find the non-skew hyperplane if C is not too small. In this situation, tuning C + C is meaningless . For linearly non-separable data, the boundary will be skew to positive class if C + = C . In this case, increasing C + C dose "push" the skewed separating hyperplane to the negative class. For both D 1 and D 2 , if the C is too small, underfitting occurs, that is, the SVMs simply classify all the samples into negative class. 5.2 TRECVID Data Set 5.2.1 Experimental Setup In this section, we evaluate the proposed approach on the high level feature extraction task of TRECVID [1]. Four concepts, including "car","maps","sports" and "waterscape", are chosen to model from the data sets. The development data of TRECVID 2005 are employed and divided into training set and validation set in equal size. The detailed statis-447 Table 2: The details of the training set and validation set of TRECVID 2005. Concept |D train | |D val | Positive Negative Positive Negative Car 1097 28881 1097 28881 Maps 296 30462 296 30463 Sports 790 29541 791 29542 Waterscape 293 30153 293 30154 tics of the data is summarized in Table 2. In our experiments , the global 64-dimension color autocorrelogram feature is used to represent the visual content of each image. Conforming to the convention of TRECVID, average precision (AP) is chosen as the evaluation criterion. Totally five algorithms have been implemented for comparison: Whole All the negative examples are used Random Random sampling of the negative examples Active Active sampling of the negative examples SCMs I SCMs with k-means in the input space SCMs SCMs with kernel k-means In the Active method, we firstly randomly select a subset of negative examples. With this initial set, we train an SVMs model and use this model to classify the whole training data set. Then the maximally misclassified negative examples are added to the training set. This procedure iterates until the ratio between the negative and the positive examples exceeding five. Since both the Random and Active methods depend on the initial random chosen data set, we repeat each of them for ten times and calculate their average performances for comparison. Both SCMs I and SCMs methods adopt the Gaussian kernel during the SVMs classification. The only difference is that SCMs I performs data clustering with k-means in the input space while SCMs with k-means in the feature space. 5.2.2 Parameter Settings Currently, the experiments focus on the comparative performance between the different approaches based on the the same parameter settings. Therefore, some of the parameters are heuristically determined and might not be optimal. The current implementation of SCMs involves the following parameter settings: (a) Gaussian kernel is adopted and the parameters are selected via cross-validation, furthermore, the kernel function of kernel k-means clustering is adopted the same as that of SVMs, (b) the threshold for transforming dense graphs to sparse ones is experimentally determined as = 0.6, (c) the parameter of shrinking technique is experimentally chosen as = 1.3, (d) for SCMs, the data are imbalanced for each concept, we only carry out data clustering for negative classes, therefore, k + always equals |D + | and k is always chosen as |D | 10 , (e) we stop the iteration of SCMs when the number of the negative examples are not more than the five times of that of the positive examples. 5.2.3 Experiment Results The average performance and time costs of the various approaches are in Table 3 and Table 4 respectively. We can see that both the Random and Active methods use fewer time than the others, but their performances are not as good as the others. Furthermore, the SCMs achieves Table 3: The average performance of the approaches on the chosen concepts, measured by average precision . Concept Whole Random Active SCMs I SCMs Car 0.196 0.127 0.150 0.161 0.192 Maps 0.363 0.274 0.311 0.305 0.353 Sports 0.281 0.216 0.253 0.260 0.283 Waterscape 0.269 0.143 0.232 0.241 0.261 Table 4: The average time costs of the approaches on the chosen concepts (in seconds). Concept Whole Random Active SCMs I SCMs Car 4000.2 431.0 1324.6 1832.0 2103.4 Maps 402.6 35.2 164.8 234.3 308.5 Sports 1384.5 125.4 523.8 732.5 812.7 Waterscape 932.4 80.1 400.3 504.0 621.3 the comparable performance with that of Whole while uses fewer time costs. Note that SCMs I also achieves satisfying results. This might be due to the Gaussian kernels, in which e - x-y 2 is monotonic with x -y 2 . Therefore, the order of the pair-wise distances is the same for both the input space and feature space, which perhaps leads to similar clustering results. CONCLUSIONS In this paper, we have investigated the effects of scale and imbalance on SVMs. We highlight the role of data overlapping in this problem and find that SVMs has no difficulties with linear separable large scale imbalanced data. We propose a meta-algorithm named Support Cluster Machines (SCMs) for effectively learning from large scale and imbalanced data sets. Different from the previous work, we develop the theoretical justifications for the idea and choose the technical component guided by the theoretical results. Finally, experiments on both the synthetic and the TRECVID data are carried out. The results support the previous analysis and show that the SCMs are efficient and effective while dealing with large scale imbalanced data sets. However, as a pilot study, there is still some room for improvement . Firstly, we have not incorporated the caching techniques to avoid the kernel reevaluations. Therefore, we have to recalculate the dot product on line whenever it is re-quired . Secondly, the parameters within the algorithms are currently selected heuristically, which depend on the tradeoff of efficiency and accuracy. ACKNOWLEDGMENTS We would like to thank the anonymous reviewers for their insightful suggestions. We also thank Dr. Chih-Jen Lin for the code of libSVM, Brian J. Kulis for the code of graclus and National Institute of Standards and Technology for providing the TRECVID data sets. Finally, special thanks go to Dr. Ya-xiang Yuan for his helpful discussions on optimization theory. REFERENCES [1] TREC Video Retrieval. National Institute of Standards and Technology, http://www-nlpir.nist.gov/projects/trecvid/. 448 [2] R. Akbani, S. Kwek, and N. Japkowicz. Applying Support Vector Machines to Imbalanced Datasets. In Proceedings of ECML'04, pages 3950, 2004. [3] D. Boley and D. Cao. Training Support Vector Machine using Adaptive Clustering. In Proceeding of 2004 SIAM International Conference on Data Mining, April 2004. [4] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, New York, NY, USA, 2004. [5] K. Brinker. Incorporating Diversity in Active Learning with Support Vector Machines. In Proceedings of ICML'03, pages 5966, 2003. [6] N. V. Chawla, N. Japkowicz, and A. Kotcz. Editorial: Special Issue on Learning from Imbalanced Data Sets. SIGKDD Explor. Newsl., 6(1):16, 2004. [7] J. W. Daniel. Stability of the Solution of Definite Quadratic Programs. Mathematical Programming, 5(1):4153, December 1973. [8] R. Datta, J. Li, and J. Z. Wang. Content-based Image Retrieval: Approaches and Trends of the New Age. In Proceedings of ACM SIGMM workshop on MIR'05, pages 253262, 2005. [9] I. Dhillon, Y. Guan, and B. Kulis. A Fast Kernel-based Multilevel Algorithm for Graph Clustering. In Proceeding of ACM SIGKDD'05, pages 629634, 2005. [10] I. S. Dhillon, Y. Guan, and B. Kulis. A Unified View of Graph Partitioning and Weighted Kernel k-means. Technical Report TR-04-25, The University of Texas at Austin, Department of Computer Sciences, June 2004. [11] C. Ding and X. He. K-means clustering via principal component analysis. In Proceedings of ICML'04, pages 2936, 2004. [12] K.-S. Goh, E. Y. Chang, and W.-C. Lai. Multimodal Concept-dependent Active Learning for Image Retrieval. In Proceedings of ACM MM'04, pages 564571, 2004. [13] A. G. Hauptmann. Towards a Large Scale Concept Ontology for Broadcast Video. In Proceedings of CIVR'04, pages 674675, 2004. [14] A. G. Hauptmann. Lessons for the Future from a Decade of Informedia Video Analysis Research. In Proceedings of CIVR'05, pages 110, 2005. [15] C.-W. Hsu, C.-C. Chang, and C.-J. Lin. A Practical Guide to Support Vector Classification. 2005. available at http://www.csie.ntu.edu.tw/~cjlin/libsvm/. [16] T. Joachims. Making Large-scale Support Vector Machine Learning Practical. Advances in kernel methods: support vector learning, pages 169184, 1999. [17] Y. Lin, Y. Lee, and G. Wahba. Support Vector Machines for Classification in Nonstandard Situations. Machine Learning, 46(1-3):191202, 2002. [18] L. M. Manevitz and M. Yousef. One-class SVMs for Document Classification. Journal of Machine Learning Research, 2:139154, 2002. [19] H. T. Nguyen and A. Smeulders. Active Learning Using Pre-clustering. In Proceedings of ICML'04, pages 7986, 2004. [20] E. Osuna, R. Freund, and F. Girosi. An Improved Training Algorithm for Support Vector Machines. In IEEE Workshop on Neural Networks and Signal Processing, September 1997. [21] D. Pavlov, J. Mao, and B. Dom. Scaling-Up Support Vector Machines Using Boosting Algorithm. In Proceeding of ICPR'00, volume 2, pages 22192222, 2000. [22] J. C. Platt. Fast Training of Support Vector Machines using Sequential Minimal Optimization. Advances in kernel methods: support vector learning, pages 185208, 1999. [23] R. C. Prati, G. E. A. P. A. Batista, and M. C. Monard. Class Imbalances versus Class Overlapping: an Analysis of a Learning System Behavior. In Proceedings of the MICAI 2004, pages 312321, 2004. [24] G. Schohn and D. Cohn. Less is More: Active Learning with Support Vector Machines. In Proceddings of ICML'00, pages 839846, 2000. [25] A. J. Smola and B. Sch okopf. Sparse Greedy Matrix Approximation for Machine Learning. In Proceedings of ICML'00, pages 911918, 2000. [26] S. Tong and E. Chang. Support Vector Machine Active Learning for Image Retrieval. In Proceedings of ACM MM'01, pages 107118, 2001. [27] K. Veropoulos, N. Cristianini, and C. Campbell. Controlling the Sensitivity of Support Vector Machines. In Proceedings of IJCAI'99, 1999. [28] G. M. Weiss and F. J. Provost. Learning When Training Data are Costly: The Effect of Class Distribution on Tree Induction. Journal of Artificial Intelligence Research (JAIR), 19:315354, 2003. [29] G. Wu and E. Y. Chang. KBA: Kernel Boundary Alignment Considering Imbalanced Data Distribution. IEEE Transactions on Knowledge and Data Engineering, 17(6):786795, 2005. [30] Z. Xu, K. Yu, V. Tresp, X. Xu, and J. Wang. Representative Sampling for Text Classification Using Support Vector Machines. In Proceedings of ECIR'03, pages 393407, 2003. [31] H. Yu, J. Yang, J. Han, and X. Li. Making SVMs Scalable to Large Data Sets using Hierarchical Cluster Indexing. Data Min. Knowl. Discov., 11(3):295321, 2005. APPENDIX A. PROOF OF THEOREM 2 Firstly, we define ^ which satisfies ^ c = x i c i , c = 1, . . . , k. It is easy to verify that ^ is a feasible solution of SCMs. Secondly, we define satisfying i = c | c | if x i c , i = 1, . . . , n. It is easy to verify that is a feasible solution of DSVMs. According to the relation of D and ~ D, we can obtain the following equation 1 2 n i=1 n j=1 i y i (~ x i ), (~ x j ) j y j n i=1 i = 1 2 k h=1 k l=1 ^ h y h (u h ), (u l ) ^ l y l k h=1 ^ h , which means ~ G( ) = G ( ^ ). Similarly, we can get ~ G( ) = G ( ). Using the fact that and are the optimal solu-449 tions to SCMs and DSVMs respectively, we have G ( ) G ( ^ ) and ~ G( ) ~ G( ). Thus, the equation G ( ) = ~ G( ) holds. For any R k satisfying { c = x i c i , c = 1, . . . , k}, we know it is a feasible solution to SCMs and G ( ) = ~ G( ) = G ( ) holds, which means is the optimal solution of SCMs. Similarly, for any R n satisfying { x i c i = c , c = 1, . . . , k} and the constraints of (7), we have ~ G() = G ( ) = ~ G( ), which means is the optimal solution of DSVMs. B. PROOF OF THEOREM 3 Note that the feasible regions of (2) and (7) are the same. By the fact that and ~ are optimal solutions to (2) and (7) respectively, we know that ( ~ ) T G( ) 0 (12) ( - ~ ) T ~ G( ~ ) 0 (13) hold, where the gradients G() = Q - e and ~ G() = ~ Q - e. (14) Adding (12) and (13) and then a little arrangement yields ( ~ ) T [ ~ G( ~ ) - ~ G( )] (~ ) T [ G( ) - ~ G( )]. Substituting (14) in the above inequality, we get ( ~ ) T ~ Q( ~ ) (~ ) T (Q - ~ Q) . (15) Adding ( ~ ) T (Q - ~ Q)( ~ ) to the both sides of (15), we have ( ~ ) T Q( ~ ) (~ ) T (Q - ~ Q) ~ . (16) If &gt; 0 is the smallest eigenvalue of Q, we have ~ 2 (~ ) T Q( ~ ) ( ~ ) T (Q - ~ Q) ~ ~ Q - ~ Q ~ and ~ ~ mC. Using (16) we get ~ ~ mC . Now we turn to prove the second result. is the optimal solution of (2), therefore, 0 G(~ ) - G( ) is obvious. Meanwhile, we have G( ~ ) - G( )= 1 2 ( ~ ) T (Q - ~ Q) ~ + ~ G( ~ ) - G( ) 12(~ ) T (Q - ~ Q) ~ + ~ G( ) - G( ) = 1 2 ( ~ ) T (Q - ~ Q) ~ - 12( ) T (Q - ~ Q) 12 Q - ~Q ~ 2 + 1 2 Q ~ Q 2 (m 2 + ~ m 2 )C 2 2 C. PROOF OF THEOREM 4 Expanding to be the explicit function of {(u c ) } k c=1 , we get 2 = YKY - ~ Y ~ K ~ Y 2 , in which Y and ~ Y denote diagonal matrices whose diagonal elements are y 1 , . . . , y n and ~ y 1 , . . . , ~ y n respectively. Using the fact that Y equals to ~ Y, we have 2 = Y(K - ~ K)Y 2 . Since Y only change the signs of the elements of K - ~ K by Y(K - ~ K)Y, we have 2 = K - ~ K 2 = k h=1 k l=1 x i h x j l ( (x i ), (x j ) (u h ), (u l ) ) 2 . It is a biquadratic function of {(u c ) } k c=1 . Therefore, this is an unconstrained convex optimization problem [4]. The necessary and sufficient condition for {u c } k c=1 to be optimal is 2 ( {(u c ) } k c=1 ) = 0. We can verify that (u c ) = xic (x i ) | c | , c = 1, . . . , k satisfies the condition that the gradient is zero. D. PROOF OF THEOREM 5 We define a n k matrix Z as Z ic = 1 | c | if x i c 0 otherwise . We can see that Z captures the disjoint cluster memberships. There is only one non-zero entry in each row of Z and Z T Z = I k holds (I k indicates the identity matrix). Suppose is the matrix of the images of the samples in feature space, i.e., = [(x 1 ), . . . , (x n )]. We can verify that the matrix ZZ T consists of the mean vectors of the clusters containing the corresponding sample. Thus, the 2 can be written as Q - ~ Q 2 = T - (ZZ T ) T ZZ T 2 . Using the fact that trace(A T A) = A 2 F , trace(A + B) = trace(A) + trace(B) and trace(AB) = trace(BA), we have 2 = trace(( T ) T T - (Z T T Z)(Z T T Z)). Since trace(( T ) T T ) is constant, minimizing is equivalent to maximizing J 1 = trace((Z T T Z)(Z T T Z)). (17) With similar procedure, we can see that minimizing J( { c } k c=1 ) amounts to maximizing J 2 = trace(Z T T Z). (18) Matrix K = T is a symmetric matrix. Let 1 . . . , n 0 denote its eigenvalues and (v 1 , . . . , v n ) be the corresponding eigenvectors. Matrix H = Z T T Z is also a symmetric matrix. Let 1 . . . , k 0 denote its eigenvalues. According to Poincar e Separation Theorem, we know the relations i i , i = 1, . . . , k hold. Therefore , we have J 2 = k i=1 i k i=1 i . Similarly, we have J 1 = k i=1 2 i k i=1 2 i . In both cases, the equations hold when Z = (v 1 , . . . , v k )R, where R is an arbitrary k k orthonormal matrix. Actually, the solution to maximizing J 2 is just the well-known theorem of Ky Fan (the Theorem 3.2. of [11]). Note that the optimal Z might no longer conforms to the definition of Z ic = 1 | c | if x i c 0 otherwise , but it is still a orthonormal matrix. That is why it is called a relaxed optimal solution. 450

Support Vector Machines;concept modelling;Concept Modelling;Imbalance;Support Vector Machines (SVMs);Large Scale;Clustering;imbalanced data;kernel k-means;support cluster machines (SCMs);TRECVID;meta-algorithm;large scale data;shrinking techniques;clusters;Kernel k-means

124

Learning Query Languages of Web Interfaces

This paper studies the problem of automatic acquisition of the query languages supported by a Web information resource . We describe a system that automatically probes the search interface of a resource with a set of test queries and analyses the returned pages to recognize supported query operators. The automatic acquisition assumes the availability of the number of matches the resource returns for a submitted query. The match numbers are used to train a learning system and to generate classification rules that recognize the query operators supported by a provider and their syntactic encodings. These classification rules are employed during the automatic probing of new providers to determine query operators they support. We report on results of experiments with a set of real Web resources.

INTRODUCTION Searching for relevant information is a primary activity on the Web. Often, people search for information using general-purpose search engines, such as Google or Yahoo!, which collect and index billions of Web pages. However, there exists an important part of the Web that remains unavailable for centralized indexing. This so-called "hidden" part of the Web includes the content of local databases and document collections accessible through search interfaces offered by various small- and middle-sized Web sites, including company sites, university sites, media sites, etc. According to the study conducted by BrightPlanet in 2000 [6], the size of the Hidden Web is about 400 to 550 times larger than the commonly defined (or "Visible") World Wide Web. This surprising discovery has fed new research on collecting and organizing the Hidden Web resources [1, 2, 15, 17, 19]. Commercial approaches to the Hidden Web are usually in the shape of Yahoo!-like directories which organize local sites belonging to specific domains. Some important examples of such directories are InvisibleWeb[1] and BrightPlanet[2] whose gateway site, CompletePlanet[3], is a directory as well as a meta-search engine. For each database incorporated into its search, the meta-search engine is provided with a manually written "wrapper", a software component that specifies how to submit queries and extract query answers embedded into HTML-formatted result pages. Similar to the Visible Web, search resources on the Hidden Web are highly heterogeneous. In particular, they use different document retrieval models, such as Boolean or vector-space models. They allow different operators for the query formulation and, moreover, the syntax of supported operators can vary from one site to another. Conventionally, query languages are determined manually; reading the help pages associated with a given search interface, probing the interface with sample queries and checking the result pages is often the method of choice. The manual acquisition of Web search interfaces has important drawbacks. First, the manual approach is hardly scalable to thousands of search resources that compose the Hidden Web. Second, the manual testing of Web resources with probe queries is often error-prone due to the inability to check results. Third, cases of incorrect or incomplete help pages are frequent. Operators that are actually supported by an engine may not be mentioned in the help pages, and conversely, help pages might mention operators that are not supported by the engine. To overcome the shortcomings of the manual approach, we address the problem of acquiring the query languages of Web resources in an automatic manner. We develop a system that automatically probes a resource's search interface with a set of test queries and analyses the returned pages to recognize supported query operators. The automatic acquisition assumes the availability of the number of matches the resource returns for a submitted query. The match numbers are used to train a learning system and to generate classification rules that recognize the query operators supported by a provider and their syntactic encodings. New technologies surrounding the XML syntax standard, in particular Web Services [18], establish a new basis for automatic discovery and information exchange and are becoming widely employed in corporate applications. However, this has yet to happen for thousands of public information providers. The question of when and how they will move toward open cooperation using Web Service technologies remains widely open [4]. Instead, the query-probing approach for acquiring supported operators does not assume any cooperation of Web providers; its only requirement is that they 1114 2004 ACM Symposium on Applied Computing provide an accessible interface and allow queries to be run. This paper is organized as follows. In Section 2 we discuss the heterogeneity of Web interfaces; we formalize the problem and show its connection with the concept of learning by querying in Section 3. In Section 4 we design a classifier system for the automatic acquisition of a query language and investigate different aspects of the system. In Section 6 we review the prior art; in Section 5 we present experimental results to illustrate the performance of our system. Section 7 discusses open issues and Section 8 concludes the paper. QUERYING WEB RESOURCES Web resources vary considerably in the ways they retrieve relevant documents. In the theory of information retrieval, there exist at least five basic retrieval models, but only three of these models are visible on the Web, namely the Boolean, the extended Boolean and the vector-space models. In the Boolean query model, a query is a condition, which documents either do or do not satisfy, with the query result being a set of documents. In the vector-space model, a query is a list of terms, and documents are assigned a score according to how similar they are to the query. The query result is a ranked list of documents. A document in the query result might not contain all query terms. Finally, the extended Boolean model combines the advantages of both the Boolean and the vector-space query model. In this model, keywords can be preceded by special characters (like + and - ) requiring an obligatory presence or absence of a given keyword in a document. For example, the query +information +provider will retrieve all documents containing both keywords and rank them according to some similarity function. Analysis of information providers suggests that the majority of providers adopt one of the three basic models. Moreover , beyond query answers, many resources report the number of documents in their collections matching the query. If a resource deploys the (extended) Boolean model, the match number shows how many documents match the query. In the case of the vector-space model, the match number refers to documents containing at least one query term, thus being equivalent to the Boolean disjunction. In the following, we develop an approach for automatic determination of query operators by reasoning on submitted queries and corresponding match numbers. Though this approach excludes resources that do not report match numbers , other ways of automatic detection of query operators appear even more problematic and difficult to implement. A method based on downloading answer documents and verifying the query against their content often fails, either for legal reasons, when the content of documents is unavailable or password-protected, or for technical reasons, when a query matches millions of documents and downloading even a part of them requires prohibitive time and network resources. 2.1 Query language model A query language of a Web provider includes a set of basic operators and the way of combining the operators to get complex queries. Basic operators have different arities, in particular, the default term processing and the unary and binary operators. The default processing refers primarily to case sensitivity in this paper, but we could also refer to whether the query term is treated as a complete word or as a substring in a possible matching document. Unary operators include the Stem-operator, which replaces a query term with its lexem; binary operators include the Boolean operators (conjunction), (disjunction), and (negation) 1 and the operator P hrase which requires the adjacency of all terms in a document. Some other operators, like substring matching or word proximity operators have been studied in various systems, however the six query operators mentioned above are by far the ones most frequently supported by Web interfaces. In the following, we develop a method to cope with the operator set O = {Case, Stem, , , , P hrase}. Issues relevant to the possible extension of set O with other operators are delegated to Section 7. 2.2 Query interpretation Web providers are queried by filling their search forms with query strings. CGI or JavaScript code linked to the query form interprets the query strings according to certain rules. These rules allow syntactic encodings for the supported query operators. If correctly interpreted, the query is executed on the document collection before a (full or partial ) answer is reported to the user. Unfortunately, the same query operator may be encoded differently by different providers. For example, the Boolean conjunction is often encoded as A AND B , A B , or +A +B , where A and B are query terms. Worse, two providers can interpret the same query string differently. For example, query string A B can be interpreted as a Boolean conjunction , Boolean disjunction, or P hrase. Example 1. To illustrate the problem, consider the query string q = Casablanca AND Bogart . On Google, AND is interpreted as the Boolean conjunction, that is, i Google ( Casablanca AND Bogart ) = Casablanca Bogart . As a result, query q matches 24,500 pages at Google, as op-posed to 551,000 for query q 1 = Casablanca and 263,000 for q 2 = Bogart . On the Internet Movie Database (IMDB) (http://www.imdb.com/.), AND is taken literally and all terms in a query are implicitly OR-connected. Therefore, the IMDB interprets query q as follows: i IM DB ( Casablanca AND Bogart ) = Casablanca AND Bogart . The query returns 12,020 matches documents on IMDB, as op-posed to only 22 for q 1 = Casablanca and 4 for q 2 = Bogart . If we investigate an unknown query language, then Example 1 shows that observing match numbers for probe queries can provide a good insight into the supported operators. However, no definitive decision appears possible from the three queries above q, q 1 , q 2 . An accurate decision on supported operators/syntaxes will require probing the provider with other queries and comparing all match numbers in order to confirm or reject various hypotheses. Example 2. As in Example 1, let us compare match numbers for the queries q= Casablanca AND Bogart , q 1 = Casablanca , and q 2 = Bogart . For Google, the fact that q matches less documents than any of q 1 and q 2 , favors the Conjunction-hypotheses , but is still insufficient to exclude other hypotheses , like that of P hrase. Probing Google with query q 3 = Bogart AND Casablanca returns the same number of matched documents as q. This (most likely) discards the P hrase-hypothesis , but not the hypothesis Casablanca AND 1 Negation is a binary operator in Web query languages and its interpretation is given by 'AND NOT', that is, A B is a synonym for A B (the latter using the unary ). 1115 Bogart . To exclude this one, even more queries should be sent to Google, like q 4 = Casablanca AND , and so on. Sim-ilarly in IMDB, the fact that query q matches more documents than q 1 and q 2 suggests that q is processed as a disjunction , but it can not tell whether AND is taken literally or ignored. A deeper analysis requires further probing IMDB with, for example, queries q 4 = Casablanca AND or q 5 = Casablanca Bogart to compare their match numbers to the ones of previous queries and decide about the AND . Our approach to the automatic acquisition of Web query languages formalizes and generalizes the idea described in Examples 1 and 2. We build a learning system that trains a number of classifiers with data from manually annotated sites to automatically determine supported operators and their syntaxes at a new site. The training data from annotated sites includes an ensemble of test queries together with the corresponding match numbers. PROBLEM DEFINITION Assume an information provider P supports some or all query operators in O; these operators form a set O P , O P O and allow us to compose a set of complex queries Q(O P ). For any operator o i O P , P accepts one or more syntactical encodings, s i1 , s i2 , . . .. The set {s ij } of accepted syntaxes for o i O P is denoted S i . The interpretation I P of operator set O P is defined as I P = {(o i , s ij )|o i O P , s ij S i } = {(o i , S i )|o i O P }. Interpretation I P is monovalued if each operator has at most one syntax, i.e, |S i | = 1 for all o i O P . I P is multivalued, if it allows multiple syntaxes for at least one operator, i.e., o i O P such that |S i | &gt; 1. In Google, the Boolean conjunction can be encoded by both AND and (whitespace). Therefore, for any query terms A and B, both query strings A B and A AND B are interpreted as A B. I Google contains (, AND ) and (, ) and is a multivalued interpretation. We distinguish between ambiguous and unambiguous interpretations . A pair of distinct operator encodings (o i , s ij ) and (o k , s kl ) is ambiguous if the two operators have the same syntax: o i = o k but s ij = s kl . An interpretation I P is ambiguous , if it contains at least one ambiguous pair of encodings . An interpretation I is unambiguous, if for any pair of encodings (o i , s ij ) and (o k , s kl ) in I, o i = o k s ij = s kl . Ambiguous interpretations can be observed with Web providers that interpret query strings dynamically, when the final decision depends on results of the query execution with different retrieval models 2 . However, the major part of Web providers interpret query strings unambiguously and our method copes with unambiguous interpretations only. Further discussion on ambiguous interpretations is in Section 7. Like with the query operators, we select the most frequent syntaxes on the Web, S = { Default 3 , , , AND , + , OR , NOT , - , "" (quote marks)}. Like set O, these syntaxes have been selected after verification of hundreds of Web providers. Set S is easily extendable to alternative syntaxes, like ones employed by non-English providers. For 2 Citeseer at http://citeseer.nj.nec.com/cs is an example of ambiguous interpretation. By default, it interprets A B as a conjunction; however if A B matches zero documents , the query is interpreted as disjunction. 3 'Default' refers to the absence of any syntax; it assumes the processing of plain terms. example, French providers may use ET for the Boolean conjunction and OU for the disjunction. The theoretical framework for the query language acquisition is derived from the learning of an unknown concept by querying [5]. Assume that provider P supports the basic operators in O; complex queries composed from the basic operators form a set Q(O). For the document collection at P , query q Q(O) constrains a subset P (q) of documents matching q. An abstract query q Q(O) is mapped into a textual string with a mapping M : O 2 S that defines (possibly multiple) syntaxes for operators in O. The mapping of a complex query q is denoted m(q), the set of mapped queries is denoted Q(S) = Q(M (O)). The sets O and S are assumed to be known, whereas the mapping M is unknown. We are given an oracle that can be queried with a mapped query m(q) Q(S) on the size of subset P (q), oracle(m(q)) = |P (q)|. By observing the or-acle's responses to queries, the learning system should produce a hypothesis on the mapping M , which should be as close as possible to the correct one. The identification of the mapping M may be simple under certain circumstances. Below we show an example of reconstruction when O P includes a particular subset of operators and the oracle is noiseless. Example 3. Let O include the three Boolean operators (, and ) and P hrase. Then, for a given syntax set S, any unambiguous mapping M : O 2 S can be exactly identified if the oracle is noise-less 4 . In such a case, subset sizes returned by the oracle fit the Boolean logic on sets.Indeed, when querying the oracle with terms A and B and syntaxes from S, the disjunction is distinguishable from other operators by the fact that it constrains bigger subsets in a collection than any of terms does: |A B| |A|, |A B| |B| (1) Furthermore, among three other operators, the conjunction is recognized by its commutativity: |A B| = |B A| (2) Finally, the difference between negation and phrases is detected by the basic equation linking three Boolean operators: |A B| = |AB| + |A B| + |BA| (3) Sizes of subsets constrained by the Boolean operators satisfy the disequation (1) and equations (2), (3) for any pair of A and B, so one can easily design a learning system that exactly identifies an unambiguous mapping M after only a few probing queries. Unfortunately, easy identification of the mapping M is rather an exception on the real Web, where few if any of the assumptions made in Example 3 become true. First, any change in the operator set O p makes the exact reconstruction less obvious. If the conjunction and/or disjunction are not supported, then the size of A B (or A B) is unavailable and equation (3) cannot help distinguish negation from phrases. In cases like this, the identification of supported syntaxes requires an analysis of the semantic correlation between query terms A and B and guessing on their co-occurrence in (unknown) document collections. 4 Oracle noiseless assumes the pure Boolean logics, with no query preprocessing, like the stopword removal. 1116 Second, Web query interfaces that play the role of oracles and return sizes of subsets constrained by queries m(q) Q(S) are rarely noiseless. When probing interfaces with test queries, the match numbers may violate equations (2) and (3). Most violations happen because converting query strings into queries on collections hides the stop-word removal and term stemming. It is not clear, whether queries like A AND B are interpreted as one (A is a stopword), two, or three terms. Moreover, for the performance reasons, real match numbers are often replaced by their estimations which are calculated using various collection statistics [13], without the real retrieval of documents matching the query. LEARNING SYSTEM To automatically determine supported query operators, we reduce the overall problem to a set of classification tasks, where each task is associated with recognizing a specific query operator or syntax, and where some standard learning algorithms like SVM, k-nearest neighbors or decision trees can be applied. To build the classifiers, we collect and annotate a set of Web providers. We develop a set of test queries and probe all selected providers with the test queries. We train the classifiers with query matches for test queries. For any new provider, we first probe it with the test queries. Query matches returned by the provider upon test queries are used to automatically classify operators and syntaxes and produce an unambiguous interpretation for P . To achieve a good level of classification accuracy, we investigate different aspects of the learning system including the target function, probe queries, data preparation, and feature encoding and selection. 4.1 Target function Due to the multivalued relationships between query operators and syntaxes, the target function for our learning system has two alternatives, one for the direct mapping M and the other one for the inverted mapping M -1 : T 1 : O 2 S . T 1 targets the unknown mapping M ; it assigns zero or more syntaxes to each operator in O. T 1 builds a multi-value classifier for every o i O, or alternatively, a set of binary classifiers for all valid combinations (o i , s j ), o i O, s j S(o i ). T 2 : S O. T 2 targets the inverted mapping M -1 ; it assigns at most one operator to every syntax s j S. Either target function gets implemented as a set of classifiers , operator classifiers for T 1 or syntax classifiers for T 2 . Classifiers are trained with match numbers for probe queries from annotated providers. For a new provider P , either function produces a hypothesis I T (P ) that approximates the real interpretation I P . The major difference between T 1 and T 2 is that the former can produce ambiguous interpretations , while the output of T 2 is always unambiguous. Indeed, two operator classifiers with T 1 can output the same syntax leading to ambiguity, while each classifier in T 2 outputs at most, one operator for one syntax. In experiments we tested both functions, though when building the learning system we put an emphasis on T 2 , which is free of ambiguity. To build syntax classifiers for the target function T 2 , we should consider beyond "good" classification cases for the operators in O and include some "real-world" cases where providers process syntaxes in S literally or simply ignore them. For certain providers, it is difficult to find any valid interpretation. In the learning system, we extend the set of possible interpretations of syntaxes in S by three more cases, O = O{Ignored, Literal, U nknown}. Syntaxes in S have different alternatives for their interpretation; below we revisit some syntaxes and report possible matches in O as they are specified in the learning system. Default : Case sensitivity for query terms: possible values are case-insensitive (Case) or case-sensitive (Literal). * : This unary operator can be interpreted as Stem, when i(A*) = Stem(A), Ignored when i(A*) = i(A), and Literal, when A* is accepted as one term. : Whitespace is often a default for another syntax in S. Three possible interpretations include the Boolean conjunction when i( A B )= A B, the Boolean disjunction when i( A B )= A B, and P hrase when i( A B )= P hrase (A,B). AND : Three alternatives here are the conjunction when i( A AND B )= A B, Ignored, when AND is ignored and the interpretation goes with the whitespace meaning, i( A AND B )= i( A B )= M -1 (' ') (A, B), and Literal when i( A AND B )= M -1 ( ) (A, AND ,B). " " (Quote marks): Two possible interpretations are P hrase, when i( "A B" )= P hrase(A,B), and Ignore when quote marks are ignored and terms are interpreted with the whitespace, i( "A B" )= i( A B ) = M -1 ( ) (A, B). A similar analysis is done for the syntaxes + , OR , NOT and - . Additionally, all syntaxes for binary operators can be labeled as U nknown. 4.2 Probing with test queries To train syntax classifiers for T 2 , we collect data from annotated sites by probing their interfaces and extracting the match numbers. Probing has a fairly low cost, but requires a certain policy when selecting terms for test queries to provide meaningful data for the learning. We define a set R of model queries that contain syntaxes in S and parameter terms A and B, which are later bound with real terms. We form the set R by first selecting well-formed queries that contain all syntaxes we want to classify. Second, we add queries that are term permutations of previously selected queries, for example the permutation B A for query A B . Finally, we add model queries that are not well-formed, but appear helpful for building accurate classification rules. Below, the set R of model queries is illustrated using the pair of terms A and B; model queries are split into three groups containing one, two or three words: One word queries: A , B ,UpperCase(A), A* , Stem(A). Two word queries: A B , B A , "A B" , "B A" , +A +B , +B +A , A -B , A AND , A OR , A NOT . Three word queries: A AND B , B AND A , A OR B , B OR A , A NOT B , B NOT A . In total, the set R is composed of 22 model queries, all in lower case, except UpperCase (A), which is an upper case of term A. Six queries in R are permutations of other queries 1117 and three queries are (purposely) not well-formed. These queries A AND , A OR , A NOT are selected to help detect Literal-cases for AND , OR , NOT . Probe queries are obtained from the model queries by replacing parameters A and B with specific query terms, like knowledge and retrieval . These 22 probe queries form a probe package denoted R A,B . For a provider P , probe queries together with corresponding match numbers form the elementary feature set F 0 A,B = {(m(q i ), oracle(P (q i ))), w(q i ) R A,B }. Query terms are selected from a generic English vocabulary with all standard stopwords excluded. One site can be probed with one or more probe packages, all packages using different term pairs (A,B). To probe the sites with test queries, we bind model queries in R with query terms. To obtain meaningful training data, query terms should not be common stopwords, such as and or the . As the term co-occurrence in a provider's document collection is unknown, we select pairs with different degrees of semantic correlation. Here, the term pairs fall into three categories: C 1 : terms that form a phrase (such as A= information and B= retrieval ); C 2 : terms that do not form a phrase but occur in the same document ( knowledge and wireless ); C 3 : terms that rarely occur in the same document (such as cancer and wireless ). These three categories can be expressed through term co-occurrence in some generic document collection P G . We re-use our query probing component to establish criteria for term selection for the three categories. A pair of terms (A, B) is in category C 1 (phrase co-occurrence) if the match number for P hrase(A, B) is comparable with the conjunction A B, that is |P G (P hrase(A,B))| |P G (AB)| &gt; , for some threshold 0 &lt; &lt; 1. A term pair (A, B) is in category C 2 (high co-occurrence ) if the terms are not co-occurred in a phrase, but their conjunction is comparable with either A or B, |P G (AB)| min{|P G (A)|,|P G (B)|} &gt; , for some 0 &lt; &lt; 1. If pair (A,B) does not fit the conditions for categories C 1 and C 2 , then it is in category C 3 (low co-occurrence). For our experiments, we have selected Google as generic document collection G and set the values of and both to 0.01. 4.3 Elementary features Match numbers for probe queries in F 0 A,B represent elementary features that can be directly used to train classifiers . Unfortunately, this often leads to poor results. The reason is that Web resources considerably differ in size and, therefore, the query matches from different resources are of different magnitude and thus hardly comparable. A query may match millions of documents on Google, but only a few at a small local resource. To leverage the processing of query matches from resources of different size, we develop two alternative methods for the feature encoding. In the first approach, we normalize the query matches in F 0 by the maximum number of matches for the two basic queries A and B . We thus obtain features F 1 with values mostly between 0 and 1 (except for queries related to the Boolean disjunction). The second approach F 2 to the feature encoding, uses the "less-equal-greater"-relationship between any two probe queries in a probe package. This produces a three-value feature for each pair of test queries. 4.4 Feature selection The refinement of raw features produces l=22 refined real value features with F 1 and l(l-1) 2 = 231 three-value features with F 2 . The basic approach is to train each classifier with the entire feature set F 0 , F 1 or F 2 . However, because of the noise in the data, building accurate classifiers may require a lot of training data. To control the amount of training data and enhance the quality of classification rules, we proceed with two methods of feature selection. First, we distinguish between relevant and irrelevant features for a given classifier and remove irrelevant ones. Second, beyond the direct feature filtering, we use prior knowledge and classify new syntaxes using previously classified ones. Removing irrelevant features. The definition of relevant features requires establishing syntactical dependencies between model queries in R and semantic relationships between syntaxes in S. Model query r i R syntactically depends on model query r j if r i includes syntaxes present in r j . Syntaxes s i and s j in S are semantically related if they can be interpreted with the same operator in O. We define the relevant feature set F i for syntax s i as containing three parts, F S(s i ) = F S i = F S 0 i + F S 1 i + F S 2 i . F S 0 i simply contains all model queries r j R that involve syntax s i , for example F S 0 ( AND )= { A AND B , B AND A , A AND }. Next, F S 1 i contains model queries for syntactically dependent syntaxes. Actually, F S 1 i contains the two model queries A B and B A for all binary syntaxes. Finally, F S 2 1 contains the model queries for semantically related syntaxes. For example, F S 2 ( AND ) = F S 0 ( + ), and vice versa, F S 2 ( + )=F S 0 ( AND ). Use of prior knowledge. Beyond removing irrelevant features, it is possible to benefit from the dependencies between syntaxes established in Section 4.1. For example, the Literal-cases for OR and AND depend on the interpretation of whitespaces. The classification of AND as Literal becomes simpler when the system already knows that, for example, is interpreted as conjunction. To use the prior knowledge, we alter the training and classification process. We impose an order on the syntaxes in S. When training or using syntax classifiers, we use the classification results of previous syntaxes. We convert the syntax set in the ordered list S O = (Default , , , "" , AND , + , OR ', NOT , - ) and impose the order on how the classifiers are trained and used for the classification. In the prior knowledge approach, the feature set used to train the classifier for syntax s i S O will include the classifications of all s j preceding s i in S O . Removing irrelevant features and using prior knowledge are two independent methods for feature selection and can be applied separately or together. This allows us to consider four feature selection methods for training classifiers and classifying new sites: 1. Full feature set, F f s i = F , where F is a selected feature encoding, F 0 , F 1 or F 2 ; 2. Relevant feature set, Rf s i = F S i ; 3. Prior knowledge features, P Kf s i =F M -1 (s j ), j &lt; i. 4. Relevant prior knowledge feature set RP Kf s i = F S i M -1 (s j ), j &lt; i. EXPERIMENTAL EVALUATION To run experiments, we collected and annotated 36 Web sites with search interfaces. All sites report the match num-1118 bers for user queries and unambiguously interpret their query languages. Selected sites represent a wide spectrum of supported operator sets. For each site, we annotated all supported operators and their syntaxes. For the extraction of the match numbers from HTML pages we used the Xerox IWrap wrapper toolkit [7, 12]. Out of 36 providers, only 4 support monovalued interpretations; in the other 32 cases, at least one operator has two or more syntaxes. Figure 1: T 1 and T 2 target functions. Figure 2: Three feature encodings for DT, KNN and SVM. 5.1 Experimental framework In all experiments we estimate the classification accuracy for the individual operators in O (with T 1 ) and the syntaxes in S (with T 2 ). We also estimate the mean accuracy for the target functions T 1 and T 2 . Experiments are conducted using the cross-validation method. 36 annotated sites are split into N =9 groups, S 1 , S 2 ,. . . , S N . We run N experiments; in experiment i, classifiers are trained with the groups S 1 ,. . . ,S i-1 , S i+1 ,. . . ,S N and then tested with sites from group S i . Accuracy (precision) values over N experiments are averaged for each operator/syntax classifier and form the individual accuracies. The average of individual accuracies over O/S gives the mean accuracy. We test the learning system by varying the system parameters introduced in Section 4. We train and test classifiers with three different learning algorithms: decision trees from Borgelt's package (DT), k-nearest neighbors algorithm (KNN), and support vector machines (SVM) 5 . The following list recalls the test parameters and possible options. 1. Target function: T 1 includes |O |=9 operator classifiers ; multivalued interpretations are implemented as classifications with subsets of |O |. For T 2 , the system includes |S|=9 syntax classifiers. 2. Feature encoding: The three different feature encodings (see Section 4.3) include the raw match numbers given by F 0 , the normalized match numbers given by F 1 , and three-value feature comparison given by F 2 . 3. Feature selection: The four methods presented in Section 4.4 include Ffs (full feature set), Rfs (relevant feature set), PKfs (prior knowledge feature set) and RPKfs (relevant prior knowledge feature set). 4. Term selection: We test three term selection categories , C 1 , C 2 and C 3 introduced in Section 4.2. Additionally , we test the mixture of the three categories, when three term pairs are used to probe a site, i.e. one term pair from each category C 1 , C 2 and C 3 . Experiments have been run for all parameter combinations ; most combinations achieve mean accuracy superior to 60%. The four system parameters appear to be uncorrelated in their impact on the classification results. To figure out the most interesting ones, we determine overall "winners" for each parameter, except for the learning algorithm. The winners are T 2 target function, F 2 feature encoding, and M ixed term selection. RP Kf s feature selection behaves best for DT and KNN and P Kf s feature selection is the winner for SVM. We report more detail below. 5.2 Experimental Results Decision trees are induced by the Borgelt's software; they are then pruned using the confidence level (p=0.5) pruning method. In SVM, linear kernel functions have been used. For the KNN method, we report results for k=3 which behaves better that k=1, 5 and 10. Because the implemen-tation of the KNN algorithm cannot process large sets of features, we were not able to test the Ffs and PKfs feature selection methods. All three learning algorithms show a similar performance. 3NN slightly outperforms DT and SVM for the "winner" combination (86.42% against 84.26% and 79.74%), however it is often less accurate with other parameter combinations. 5 Available at http://fuzzy.cs.uni-magdeburg.de/ borgelt/software.html, http://www.dontveter.com/ nnsoft/nnsoft.html, http://svmlight.joachims.org/, respectively. 1119 Target functions and feature selection. The target functions T 1 and T 2 implement alternative approaches to the query language acquisition; T 1 uses operator classifiers while T 2 uses syntax classifiers. As seen in Section 4, T 2 has an advantage over T 1 because it avoids multivalued classification and outputs only unambiguous interpretations, while the output of T 1 should be further tested for unambiguity. Thus we have built the learning system for T 2 . Series of experiments conducted with T 1 and T 2 confirm the superiority of T 2 . As operator classifiers in T 1 are trained indepen-dently , their combined output does not guarantee unambiguity . Unlike T 2 , high accuracy of individual classifiers may not be translated into global good accuracy, because one misclassification may produce an ambiguous interpretation and undermine the good performance of other classifiers. In practice, we test the output of operator classifiers of T 1 and discard those that form ambiguous interpretations. This gives a 2% to 10% drop in the mean accuracy. Figure 1 plots mean accuracies for T 1 and T 2 for all feature selection methods (with fixed F 2 feature encoding and M ixed term selection) and the three learning methods (only Rfs and RPKfs could be measured for 3NN algorithm). Within feature selection methods, keeping relevant features spurs the performance of DT and 3NN better than the prior knowledge, with their combination being the winner. For SVM, instead, adding prior knowledge to the full feature set is the best choice. In the following, all reported experiments refer to the target function T 2 . Feature encoding. Previous figures compared the mean accuracies. We unfold the mean value and plot individual accuracies for the syntaxes in S. Figure 2 plots accuracy values for the three feature encoding methods (for T 2 -RP Kf sM ixed combination for DT and 3NN and T 2 -P Kf s-M ixed combination for SVM). As the figure shows, the pair-wise comparison F 2 performs best with respect to the raw and normalized match numbers. Term selection. We complete the analysis of system parameters by testing four methods of term selection. They include categories C 1 , C 2 and C 3 and M ixed. Figure 3 plots mean accuracies for all learning algorithms and four term selection methods, giving M ixed as the winner. Figure 3: Four term selection methods for DT, KNN, and SVM. 5.3 Bias in training data Among the syntaxes in S, all methods show only little difference for the unary operators Def ault and Case. Among the syntaxes for binary operators, certain ( , AND and + ) are easier to detect than others ( " " , OR and NOT ). However, this phenomenon is not linked to the nature of the operators or their syntaxes, but rather can be explained by the bias in training data. In Table 1, we unfold the individual accuracies and show results for each case (s, o), s S, o O in the annotated data. Each non-empty cell in Table 1 reports the occurrence of the case (in brackets) and its classification accuracy. We can observe a definitive bias of high accuracy for more frequent cases; instead, rare cases have a very low accuracy. This explains good results for , AND and + , where occurrences are fairly split between two main cases. For other syntaxes instead, the high error ratio for rare cases decreases the individual accuracy. RELATED WORK The Hidden Web has emerged as a research trend and different research groups have started to address various problems relevant to organizing Hidden Web resources [15, 14, 17, 19, 20]. One focus is crawling; [17] presents a task-specific and human-assisted approach to the crawling of the Hidden Web. The crawler searches for Hidden Web resources relevant to a specific domain. The identification is achieved by selecting domain-specific keywords from a fuzzy set and assigning them to elements of HTML forms; the resource is judged relevant if it returns valid search results. Another important task is the classification of Hidden Web resources. [15, 14] and [19] have developed approaches to this problem based on query probing. Moreover, [15] makes use of the number of documents matching a submitted conjunction query, as does our approach. Instead of query languages, they use match numbers to reason about the relevance of a provider for a given category. Originally, the query probing has been used for the automatic language model discovery in [9], it probed documents in a collection to analyze the content of the result pages. [14] extends the work in [15] to the problem of database selection by computing content summaries based on probing . Once query language interfaces are understood, meaningful query transformation becomes possible. [11] describes one way of transforming a front-end query into subsuming queries that are supported by the sources and a way to filter out incorrectly returned documents. In [16], interaction with online-vendors is automated. In the close domain of meta-searching, the declaration of a search resource's query features is often coupled with methods of converting/translating meta-search queries into the resource's native queries. Considerable research effort has been devoted to minimizing the possible overheads of query translation when the meta-search and search resource differ in supporting basic query features [11]. In all these methods, the manual discovery of the query features is assumed. In information mediation systems that query Web resources to materialize views on hidden data [20], one approach is to reconstruct an image of a resource's database. Because of a restricted Web interface, a critical problem is the entire or partial reconstruction of the database image without the unnecessary overload of the Web servers. [8] builds efficient query covers that are accessible through nearest-neighbor interfaces for the specific domain of spatial databases. OPEN QUESTIONS The experiments have raised a number of open questions that require further research. Below we list some of them. Stopwords. In tests, common English stopwords were excluded from probing. However, the set of stopwords is 1120 Operators Syntaxes default "" AND OR + NOT Case 97.6(20) Stemming 41.1(9) Conjunction 92.9(15) 95.2(27) 100(16) Disjunction 100(19) 87.8(17) Negation 91.0(15) 94.9(26) Phrase 0(1) 90.4(28) Literal 100(16) 79.6(7) 91.7(11) 69.2(14) Ignored 79.9(27) 18.5(4) 3.7(1) 55.6(4) 100(19) 25.0(4) 81.0(7) Unknown 0(1) 4.1(4) 0(1) 19.4(4) 0(1) 0(3) 0(3) Table 1: Classification accuracy and occurrence for all syntax+interpretation cases (DT,T 2 ,F 2 ,RPKfs,M ixed). often domain-dependent; this should be taken into account when generating test queries. A more difficult case is when a resource treats terms as stopwords in a certain context. For example, Google accepts the term "WWW" when it is queried alone and ignores it when the term is conjuncted with other terms. Such query-dependent treatment of stopwords is considered as noise in the current system. Acquiring other operators. We have addressed the set of most frequently used query operators. Other operators defined by existing document retrieval models, like proximity operators, can be added to the operator set and processed in a similar manner. Two remarks concerning less frequent operators are that their syntactical encodings may vary even more than for Boolean operators, and, more im-portantly , finding sufficient training data to build reliable classifiers may be technically difficult. Query composition. The next issue is the manner in which basic query operators are combined to form complex queries. The most frequent manner on the Web is the use of parentheses or a certain operator priority. How to detect this remains an open problem at this point. Ambiguous interpretations. Recognizing ambiguous interpretations is the most difficult problem. One example is Citeseer, which interprets whitespaces as conjunction by default, but switches to disjunction if the conjunction query matches no documents. Some other Web providers behave in the same or a similar manner. We will need to extend the learning system to include a possibility of triggering the retrieval model as a function of the oracle answers. CONCLUSION We have addressed the problem of automatic recognition of operators and syntaxes supported by query languages of Web resources. We have developed a machine learning approach based on reformulation of the entire problem as a set of classification problems. By introducing various refined solutions for the target function, feature encoding, and feature selection, we have achieved 86% mean accuracy for the set of the most frequent operators and syntaxes. Further improvement in the accuracy is possible with better preparation of annotated sites, but this is limited because of the complexity of the a-priori unknown operator composition and the noise produced by the hidden query preprocessing. REFERENCES [1] The InvisibleWeb, http://www.invisibleweb.com/. [2] BrightPlanet, http://www.brightplanet.com/. [3] CompletePlanet, http://www.completeplanet.com/. [4] G. Alonso. Myths around web services. IEEE Bulletin on Data Engineering, 25(4):39, 2002. [5] D. Angluin. Queries and concept learning. Machine Learning, 2(4):319342, 1987. [6] M. K. Bergman. The Deep Web: Surfacing hidden value. Journal of Electronic Publishing, 7(1), 2001. [7] D. Bredelet and B. Roustant. Java IWrap: Wrapper induction by grammar learning. Master's thesis, ENSIMAG Grenoble, 2000. [8] S. Byers, J. Freire, and C. T. Silva. Efficient acquisition of web data through restricted query interfaces. In Proc. WWW Conf., China, May 2001. [9] J. P. Callan, M. Connell, and A. Du. Automatic discovery of language models for text databases. In Proc. ACM SIGMOD Conf., pp. 479490, June 1999. [10] C.-C. K. Chang and H Garcia-Molina. Approximate query translation across heterogeneous information sources. In Proc. VLDB Conf., pp. 566577, Cairo, Egypt, September 2000. [11] C.-C. K. Chang, H. Garcia-Molina, and A. Paepcke. Boolean query mapping across heterogeneous information sources. IEEE TKDE, 8(4):515521, 1996. [12] B. Chidlovskii. Automatic repairing of web wrappers by combining redundant views. In Proc. of the IEEE Intern. Conf. Tools with AI, USA, November 2002. [13] L. Gravano, H. Garcia-Molina, and A. Tomasic. Gloss: Text-source discovery over the internet. ACM TODS, 24(2):229264, 1999. [14] P. G. Ipeirotis and L. Gravano. Distributed search over the hidden web: Hierarchical database sampling and selection. In Proc. VLDB Conf., pp. 394405, Hong Kong, China, August 2002. [15] P. G. Ipeirotis, L. Gravano, and M. Sahami. Probe, count, and classify: Categorizing hidden-web databases. In Proc. ACM SIGMOD Conf., pp. 6778, Santa Barbara, CA, USA, May 2001. [16] M. Perkowitz, R. B. Doorenbos, O. Etzioni, and D. S. Weld. Learning to understand information on the internet: An example-based approach. Journal of Intelligent Information Systems, 8(2):133153, 1997. [17] S. Raghavan and H. Garcia-Molina. Crawling the hidden web. In Proc. VLDB Conf., pp. 129138, Rome, Italy, September 2001. [18] D. Tsur. Are web services the next revolution in e-commerce? In Proc. VLDB Conf., pp. 614617, Rome, Italy, September 2001. [19] W. Wang, W. Meng, and C. Yu. Concept hierarchy based text database categorization. In Proc. Intern. WISE Conf., pp. 283290, China, June 2000. [20] R. Yerneni, C. Li, H. Garcia-Molina, and J. Ullman. Computing capabilities of mediators. In Proc. ACM SIGMOD Conf., pp. 443454, PA, USA, June 1999. 1121

query operators;automatic acquisition;learning;hidden web;search interface;web resources;machine learning;search engine;query languages;Hidden Web;web interfaces

125

Learning Spatially Variant Dissimilarity (SVaD) Measures

Clustering algorithms typically operate on a feature vector representation of the data and find clusters that are compact with respect to an assumed (dis)similarity measure between the data points in feature space. This makes the type of clusters identified highly dependent on the assumed similarity measure. Building on recent work in this area, we formally define a class of spatially varying dissimilarity measures and propose algorithms to learn the dissimilarity measure automatically from the data. The idea is to identify clusters that are compact with respect to the unknown spatially varying dissimilarity measure. Our experiments show that the proposed algorithms are more stable and achieve better accuracy on various textual data sets when compared with similar algorithms proposed in the literature.

INTRODUCTION Clustering plays a major role in data mining as a tool to discover structure in data. Object clustering algorithms operate on a feature vector representation of the data and find clusters that are compact with respect to an assumed (dis)similarity measure between the data points in feature space. As a consequence, the nature of clusters identified by a clustering algorithm is highly dependent on the assumed similarity measure. The most commonly used dissimilarity measure, namely the Euclidean metric, assumes that the dissimilarity measure is isotropic and spatially invariant, and it is effective only when the clusters are roughly spherical and all of them have approximately the same size, which is rarely the case in practice [8]. The problem of finding non-spherical clusters is often addressed by utilizing a feature weighting technique. These techniques discover a single set of weights such that relevant features are given more importance than irrelevant features. However, in practice, each cluster may have a different set of relevant features. We consider Spatially Varying Dissimilarity (SVaD) measures to address this problem. Diday et. al. [4] proposed the adaptive distance dynamic clusters (ADDC) algorithm in this vain. A fuzzified version of ADDC, popularly known as the Gustafson-Kessel (GK) algorithm [7] uses a dynamically updated covariance matrix so that each cluster can have its own norm matrix. These algorithms can deal with hyperelliposoidal clusters of various sizes and orientations. The EM algorithm [2] with Gaussian probability distributions can also be used to achieve similar results. However, the above algorithms are computationally expensive for high-dimensional data since they invert covariance matrices in every iteration. Moreover, matrix inversion can be unstable when the data is sparse in relation to the dimensionality. One possible solution to the problems of high computation and instability arising out of using covariance matrices is to force the matrices to be diagonal, which amounts to weighting each feature differently in different clusters. While this restricts the dissimilarity measures to have axis parallel isometry, the weights also provide a simple interpretation of the clusters in terms of relevant features, which is important in knowledge discovery. Examples of such algorithms are SCAD and Fuzzy-SKWIC [5, 6], which perform fuzzy clustering of data while simultaneously finding feature weights in individual clusters. In this paper, we generalize the idea of the feature weighting approach to define a class of spatially varying dissimilarity measures and propose algorithms that learn the dissimilarity measure automatically from the given data while performing the clustering. The idea is to identify clusters inherent in the data that are compact with respect to the unknown spatially varying dissimilarity measure. We compare the proposed algorithms with a diagonal version of GK (DGK) and a crisp version of SCAD (CSCAD) on a variety of data sets. Our algorithms perform better than DGK and CSCAD, and use more stable update equations for weights than CSCAD. The rest of the paper is organized as follows. In the next section, we define a general class of dissimilarity measures 611 Research Track Poster and formulate two objective functions based on them. In Section 3, we derive learning algorithms that optimize the objective functions. We present an experimental study of the proposed algorithms in Section 4. We compare the performance of the proposed algorithms with that of DGK and CSCAD. These two algorithms are explained in Appendix A. Finally, we summarize our contributions and conclude with some future directions in Section 5. SPATIALLY VARIANT DISSIMILARITY (SVAD) MEASURES We first define a general class of dissimilarity measures and formulate a few objective functions in terms of the given data set. Optimization of the objective functions would result in learning the underlying dissimilarity measure. 2.1 SVaD Measures In the following definition, we generalize the concept of dissimilarity measures in which the weights associated with features change over feature space. Definition 2.1 We define the measure of dissimilarity of x from y 1 to be a weighted sum of M dissimilarity measures between x and y where the values of the weights depend on the region from which the dissimilarity is being measured . Let P = {R 1 , . . . , R K } be a collection of K regions that partition the feature space, and w 1 , w 2 , . . ., and w K be the weights associated with R 1 , R 2 , . . ., and R K , respectively. Let g 1 , g 2 , . . ., and g M be M dissimilarity measures. Then, each w j , j = 1, . . . , K, is an M -dimensional vector where its l-th component, w jl is associated with g l . Let W denote the K-tuple (w 1 , . . . , w K ) and let r be a real number. Then, the dissimilarity of x from y is given by: f W (x, y) = M l=1 w r jl g l (x, y), if y R j . (1) We refer to f W as a Spatially Variant Dissimilarity (SVaD) measure. Note that f W need not be symmetric even if g i are symmetric . Hence, f W is not a metric. Moreover, the behavior of f W depends on the behavior of g i . There are many ways to define g i . We list two instances of f W . Example 2.1 (Minkowski) Let d be the feature space and M = d. Let a point x d be represented as (x 1 , . . . , x d ). Then, when g i (x, y) = |x i - y i | p for i = 1, . . . , d, and p 1, the resulting SVaD measure, f M W is called Minkowski SVaD (MSVaD) measure. That is, f M W (x, y) = d l=1 w r jl |x l - y l | p , if y R j . (2) One may note that when w 1 = = w K and p = 2, f M W is the weighted Euclidean distance. When p = 2, we call f M W a Euclidean SVaD (ESVaD) measure and denote it by f E W . 1 We use the phrase "dissimilarity of x from y" rather than "dissimilarity between x and y" because we consider a general situation where the dissimilarity measure depends on the location of y. As an example of this situation in text mining, when the dissimilarity is measured from a document on `terrorism' to a document x, a particular set of keywords may be weighted heavily whereas when the dissimilarity is measured from a document on `football' to x, a different set of keywords may be weighted heavily. Example 2.2 (Cosine) Let the feature space be the set of points with l 2 norm equal to one. That is, x 2 = 1 for all points x in feature space. Then, when g l (x, y) = (1/d - x l y l ) for l = 1, . . . , d, the resulting SVaD measure f C W is called a Cosine SVaD (CSVaD) measure: f C W (x, y) = d i=1 w r jl (1/d - x l y l ), if y R j . (3) In the formulation of the objective function below, we use a set of parameters to represent the regions R 1 , R 2 , . . ., and R K . Let c 1 , c 2 , . . ., and c K be K points in feature space. Then y R j iff f W (y, c j ) &lt; f W (y, c i ) for i = j. (4) In the case of ties, y is assigned to the region with the lowest index. Thus, the K-tuple of points C = (c 1 , c 2 , . . . , c K ) defines a partition in feature space. The partition induced by the points in C is similar in nature to a Voronoi tessellation. We use the notation f W,C whenever we use the set C to parameterize the regions used in the dissimilarity measure. 2.2 Objective Function for Clustering The goal of the present work is to identify the spatially varying dissimilarity measure and the associated compact clusters simultaneously. It is worth mentioning here that, as in the case of any clustering algorithm, the underlying assumption in this paper is the existence of such a dissimilarity measure and clusters for a given data set. Let x 1 , x 2 , . . ., and x n be n given data points. Let K be a given positive integer. Assuming that C represents the cluster centers, let us assign each data point x i to a cluster R j with the closest c j as the cluster center 2 , i.e., j = arg min l f W,C (x i , c l ). (5) Then, the within-cluster dissimilarity is given by J (W, C) = K j=1 x i R j M l=1 w r jl g l (x i , c j ). (6) J (W, C) represents the sum of the dissimilarity measures of all the data points from their closest centroids. The objective is to find W and C that minimize J (W, C). To avoid the trivial solution to J (W, C), we consider a normalization condition on w j , viz., M l=1 w jl = 1. (7) Note that even with this condition, J (W, C) has a trivial solution: w jp = 1 where p = arg min l x i R j g l (x i , c j ), and the remaining weights are zero. One way to avoid convergence of w j to unit vectors is to impose a regularization condition on w j . We consider the following two regularization measures in this paper: (1) Entropy measure: M l=1 w jl log(w jl ) and (2) Gini measure: M l=1 w 2 jl . 2 We use P = {R 1 , R 2 , . . . , R K } to represent the corresponding partition of the data set as well. The intended interpretation (cluster or region) would be evident from the context. 612 Research Track Poster ALGORITHMS TO LEARN SVAD MEASURES The problem of determining the optimal W and C is similar to the traditional clustering problem that is solved by the K-Means Algorithm (KMA) except for the additional W matrix. We propose a class of iterative algorithms similar to KMA. These algorithms start with a random partition of the data set and iteratively update C, W and P so that J (W, C) is minimized. These iterative algorithms are instances of Alternating Optimization (AO) algorithms. In [1], it is shown that AO algorithms converge to a local optimum under some conditions. We outline the algorithm below before actually describing how to update C, W and P in every iteration. Randomly assign the data points to K clusters. REPEAT Update C: Compute the centroid of each cluster c j . Update W : Compute the w jl j, l. Update P: Reassign the data points to the clusters. UNTIL (termination condition is reached). In the above algorithm, the update of C depends on the definition of g i , and the update of W on the regularization terms. The update of P is done by reassigning the data points according to (5). Before explaining the computation of C in every iteration for various g i , we first derive update equations for W for various regularization measures. 3.1 Update of Weights While updating weights, we need to find the values of weights that minimize the objective function for a given C and P. As mentioned above, we consider the two regularization measures for w jl and derive update equations. If we consider the entropy regularization with r = 1, the objective function becomes: J EN T (W, C) = K j=1 x i R j M l=1 w jl g l (x i , c j ) + K j=1 j M l=1 w jl log(w jl ) + K j=1 j M l=1 w jl - 1 . (8) Note that j are the Lagrange multipliers corresponding to the normalization constraints in (7), and j represent the relative importance given to the regularization term relative to the within-cluster dissimilarity. Differentiating J EN T (W, C) with respect to w jl and equating it to zero, we obtain w jl = exp -( j + x i Rj g l ( x i , c j )) j - 1 . Solving for j by substituting the above value of w jl in (7) and substituting the value of j back in the above equation, we obtain w jl = exp x i R j g l (x i , c j )/ j M n=1 exp x i R j g n (x i , c j )/ j . (9) If we consider the Gini measure for regularization with r = 2, the corresponding w jl that minimizes the objective function can be shown to be w jl = 1/( j + x i R j g l (x i , c j )) M n=1 (1/( j + x i R j g n (x i , c j ))) . (10) In both cases, the updated value of w jl is inversely related Algorithm Update Equations Acronyms P C W EEnt (5) (11) (9) EsGini (5) (11) (10) CEnt (5) (12) (9) CsGini (5) (12) (10) Table 1: Summary of algorithms. to x i R j g l (x i , c j ). This has various interpretations based on the nature of g l . For example, when we consider the ESVaD measure, w jl is inversely related to the variance of l-th element of the data vectors in the j-th cluster. In other words, when the variance along a particular dimension is high in a cluster, then the dimension is less important to the cluster. This popular heuristic has been used in various contexts (such as relevance feedback) in the literature [9]. Similarly, when we consider the CSVaD measure, w jl is directly proportional to the correlation of the j-th dimension in the l-th cluster. 3.2 Update of Centroids Learning ESVaD Measures: Substituting the ESVaD measure in the objective function and solving the first order necessary conditions, we observe that c jl = 1 |R j | x i R j x il (11) minimizes J ESV AD (W, C). Learning CSVaD Measures: Let x il = w jl x il , then using the Cauchy-Swartz inequality, it can be shown that c jl = 1 |R j | x i R j x il (12) maximizes x i R j d l=1 w jl x il c jl . Hence, (12) also minimizes the objective function when CSVaD is used as the dissimilarity measure. Table 1 summarizes the update equations used in various algorithms. We refer to this set of algorithms as SVaD learning algorithms. EXPERIMENTS In this section, we present an experimental study of the algorithms described in the previous sections. We applied the proposed algorithms on various text data sets and compared the performance of EEnt and EsGini with that of K-Means, CSCAD and DGK algorithms. The reason for choosing the K-Means algorithm (KMA) apart from CSCAD and DGK is that it provides a baseline for assessing the advantages of feature weighting. KMA is also a popular algorithm for text clustering. We have included a brief description of CSCAD and DGK algorithms in Appendix A. Text data sets are sparse and high dimensional. We consider standard labeled document collections and test the proposed algorithms for their ability to discover dissimilarity measures that distinguish one class from another without actually considering the class labels of the documents. We measure the success of the algorithms by the purity of the regions that they discover. 613 Research Track Poster 4.1 Data Sets We performed our experiments on three standard data sets: 20 News Group, Yahoo K1, and Classic 3. These data sets are described below. 20 News Group 3 : We considered different subsets of 20 News Group data that are known to contain clusters of varying degrees of separation [10]. As in [10], we considered three random samples of three subsets of the 20 News Group data. The subsets denoted by Binary has 250 documents each from talk.politics.mideast and talk.politics.misc. Multi5 has 100 documents each from comp.graphics, rec.motorcycles, rec.sport.baseball, sci.space, and talk.politics.mideast. Finally , Multi10 has 50 documents each from alt.atheism, comp. sys.mac.hardware, misc.forsale, rec.autos, rec.sport.hockey, sci.crypt, sci.electronics, sci.med, sci.space, and talk.politics. gun. It may be noted that Binary data sets have two highly overlapping classes. Each of Multi5 data sets has samples from 5 distinct classes, whereas Multi10 data sets have only a few samples from 10 different classes. The size of the vocabulary used to represent the documents in Binary data set is about 4000, Multi5 about 3200 and Multi10 about 2800. We observed that the relative performance of the algorithms on various samples of Binary, Multi5 and Multi10 data sets was similar. Hence, we report results on only one of them. Yahoo K1 4 : This data set contains 2340 Reuters news articles downloaded from Yahoo in 1997. There are 494 from Health, 1389 from Entertainment, 141 from Sports, 114 from Politics, 60 from Technology and 142 from Business. After preprocessing, the documents from this data set are represented using 12015 words. Note that this data set has samples from 6 different classes. Here, the distribution of data points across the class is uneven, ranging from 60 to 1389. Classic 3 5 : Classic 3 data set contains 1400 aerospace systems abstracts from the Cranfield collection, 1033 medical abstracts from the Medline collection and 1460 information retrieval abstracts from the Cisi collection, making up 3893 documents in all. After preprocessing, this data set has 4301 words. The points are almost equally distributed among the three distinct classes. The data sets were preprocessed using two major steps. First, a set of words (vocabulary) is extracted and then each document is represented with respect to this vocabulary. Finding the vocabulary includes: (1) elimination of the standard list of stop words from the documents, (2) application of Porter stemming 6 for term normalization, and (3) keeping only the words which appear in at least 3 documents. We represent each document by the unitized frequency vector. 4.2 Evaluation of Algorithms We use the accuracy measure to compare the performance of various algorithms. Let a ij represent the number of data points from class i that are in cluster j. Then the accuracy of the partition is given by j max i a ij /n where n is the total number of data points. It is to be noted that points coming from a single class need not form a single cluster. There could be multiple 3 http://www-2.cs.cmu.edu/afs/cs.cmu.edu/project/theo-20/www/data/news20 .tar.gz 4 ftp://ftp.cs.umn.edu/dept/users/boley/PDDPdata/doc-K 5 ftp://ftp.cs.cornell.edu/pub/smart 6 http://www.tartarus.org/~martin/PorterStemmer/ Iteration 0 1 2 3 4 5 J (W, C) 334.7 329.5 328.3 328.1 327.8 Accuracy 73.8 80.2 81.4 81.6 82 82 Table 2: Evolution of J (W, C) and Accuracies with iterations when EEnt applied on a Multi5 data. clusters in a class that represent sub-classes. We study the performance of SVaD learning algorithms for various values of K, i.e., the number of clusters. 4.3 Experimental Setup In our implementations, we have observed that the proposed algorithms, if applied on randomly initialized centroids , show unstable behavior. One reason for this behavior is that the number of parameters that are estimated in feature-weighting clustering algorithms is twice as large as that estimated by the traditional KMA. We, therefore, first estimate the cluster centers giving equal weights to all the dimensions using KMA and then fine-tune the cluster centers and the weights using the feature-weighting clustering algorithms. In every iteration, the new sets of weights are updated as follows. Let w n (t+1) represent the weights com-puted using one of (9), (10), (14) or (15) in iteration (t + 1) and w(t) the weights in iteration t. Then, the weights in iteration (t + 1) are w(t + 1) = (1 - (t))w(t) + (t)w n (t + 1), (13) where (t) [0, 1] decreases with t. That is, (t) = (t 1 ), for a given constant [0, 1]. In our experiments, we observed that the variance of purity values for different initial values of (0) and above 0.5 is very small. Hence, we report the results for (0) = 0.5 and = 0.5. We set the value of j = 1. It may be noted that when the documents are represented as unit vectors, KMA with the cosine dissimilarity measure and Euclidean distance measure would yield the same clusters . This is essentially the same as Spherical K-Means algorithms described in [3]. Therefore, we consider only the weighted Euclidean measure and restrict our comparisons to EEnt and EsGini in the experiments. Since the clusters obtained by KMA are used to initialize all other algorithms considered here, and since the results of KMA are sensitive to initialization, the accuracy numbers reported in this section are averages over 10 random initializations of KMA. 4.4 Results and Observations 4.4.1 Effect of SVaD Measures on Accuracies In Table 2, we show a sample run of EEnt algorithm on one of the Multi5 data sets. This table shows the evolution of J (W, C) and the corresponding accuracies of the clusters with the iterations. The accuracy, shown at iteration 0, is that of the clusters obtained by KMA. The purity of clusters increases with decrease in the value of the objective function defined using SVaD measures. We have observed a similar behavior of EEnt and EsGini on other data sets also. This validates our hypothesis that SVaD measures capture the underlying structure in the data sets more accurately. 614 Research Track Poster 4.4.2 Comparison with Other Algorithms Figure 1 to Figure 5 show average accuracies of various algorithms on the 5 data sets for various number of clusters . The accuracies of KMA and DGK are very close to each other and hence, in the figures, the lines corresponding to these algorithms are indistinguishable. The lines corresponding to CSCAD are also close to that of KMA in all the cases except Class 3. General observations: The accuracies of SVaD algorithms follow the trend of the accuracies of other algorithms. In all our experiments, both SVaD learning algorithms improve the accuracies of clusters obtained by KMA. It is observed in our experiments that the improvement could be as large as 8% in some instances. EEnt and EsGini consis-tently perform better than DGK on all data sets and for all values of K. EEnt and EsGini perform better than CSCAD on all data sets excepts in the case of Classic 3 and for a few values of K. Note that the weight update equation of CSCAD (15) may result in negative values of w jl . Our experience with CSCAD shows that it is quite sensitive to initialization and it may have convergence problems. In contrast, it may be observed that w jl in (9) and (10) are always positive. Moreover , in our experience, these two versions are much less sensitive to the choice of j . Data specific observations: When K = 2, EEnt and EsGini could not further improve the results of KMA on the Binary data set. The reason is that the data set contains two highly overlapping classes. However, for other values of K, they marginally improve the accuracies. In the case of Multi5, the accuracies of the algorithms are non-monotonic with K. The improvement of accuracies is large for intermediate values of K and small for extreme values of K. When K = 5, KMA finds relatively stable clusters. Hence, SVaD algorithms are unable to improve the accuracies as much as they did for intermediate values of K. For larger values of K, the clusters are closely spaced and hence there is little scope for improvement by the SVaD algorithms. Multi10 data sets are the toughest to cluster because of the large number of classes present in the data. In this case, the accuracies of the algorithms are monotonically increasing with the number of clusters. The extent of improvement of accuracies of SVaD algorithms over KMA is almost constant over the entire range of K. This reflects the fact that the documents in Multi10 data set are uniformly distributed over feature space. The distribution of documents in Yahoo K1 data set is highly skewed. The extent of improvements that the SVaD algorithms could achieve decrease with K. For higher values of K, KMA is able to find almost pure sub-clusters, resulting in accuracies of about 90%. This leaves little scope for improvement. The performance of CSCAD differs noticeably in the case of Classic 3. It performs better than the SVaD algorithms for K = 3 and better than EEnt for K = 9. However, for larger values of K, the SVaD algorithms perform better than the rest. As in the case of Multi5, the improvements of SVaD algorithms over others are significant and consistent. One may recall that Multi5 and Classic 3 consist of documents from distinct classes. Therefore, this observation implies that when there are distinct clusters in the data set, KMA yields confusing clusters when the number of clusters is over-Figure 1: Accuracy results on Binary data. Figure 2: Accuracy results on Multi5 data. specified. In this scenario, EEnt and EsGini can fine-tune the clusters to improve their purity. SUMMARY AND CONCLUSIONS We have defined a general class of spatially variant dissimilarity measures and proposed algorithms to learn the measure underlying a given data set in an unsupervised learning framework. Through our experiments on various textual data sets, we have shown that such a formulation of dissimilarity measure can more accurately capture the hidden structure in the data than a standard Euclidean measure that does not vary over feature space. We have also shown that the proposed learning algorithms perform better than other similar algorithms in the literature, and have better stability properties. Even though we have applied these algorithms only to text data sets, the algorithms derived here do not assume any specific characteristics of textual data sets. Hence, they Figure 3: Accuracy results on Multi10 data. 615 Research Track Poster Figure 4: Accuracy results on Yahoo K1 data. Figure 5: Accuracy results on Classic 3 data. are applicable to general data sets. Since the algorithms perform better for larger K, it would be interesting to investigate whether they can be used to find subtopics of a topic. Finally, it will be interesting to learn SVaD measures for labeled data sets. REFERENCES [1] J. C. Bezdek and R. J. Hathaway. Some notes on alternating optimization. In Proceedings of the 2002 AFSS International Conference on Fuzzy Systems. Calcutta, pages 288300. Springer-Verlag, 2002. [2] A. P. Dempster, N. M. Laird, and Rubin. Maximum likelihood from incomplete data via the EM algorithm. Journal Royal Statistical Society B, 39(2):138, 1977. [3] I. S. Dhillon and D. S. Modha. Concept decompositions for large sparse text data using clustering. Machine Learning, 42(1):143175, January 2001. [4] E. Diday and J. C. Simon. Cluster analysis. In K. S. Fu, editor, Pattern Recognition, pages 4794. Springer-Verlag, 1976. [5] H. Frigui and O. Nasraoui. Simultaneous clustering and attribute discrimination. In Proceedings of FUZZIEEE, pages 158163, San Antonio, 2000. [6] H. Frigui and O. Nasraoui. Simultaneous categorization of text documents and identification of cluster-dependent keywords. In Proceedings of FUZZIEEE, pages 158163, Honolulu, Hawaii, 2001. [7] D. E. Gustafson and W. C. Kessel. Fuzzy clustering with the fuzzy covariance matrix. In Proccedings of IEEE CDC, pages 761766, San Diego, California, 1979. [8] R. Krishnapuram and J. Kim. A note on fuzzy clustering algorithms for Gaussian clusters. IEEE Transactions on Fuzzy Systems, 7(4):453461, Aug 1999. [9] Y. Rui, T. S. Huang, and S. Mehrotra. Relevance feedback techniques in interactive content-based image retrieval. In Storage and Retrieval for Image and Video Databases (SPIE), pages 2536, 1998. [10] N. Slonim and N. Tishby. Document clustering using word clusters via the information bottleneck method. In Proceedings of SIGIR, pages 208215, 2000. APPENDIX A. OTHER FEATURE WEIGHTING CLUSTERING TECHNIQUES A.1 Diagonal Gustafson-Kessel (DGK) Gustafson and Kessel [7] associate each cluster with a different norm matrix. Let A = (A 1 , . . . , A k ) be the set of k norm matrices associated with k clusters. Let u ji is the fuzzy membership of x i in cluster j and U = [u ji ]. By restricting A j s to be diagonal and u ji {0, 1}, we can reformulate the original optimization problem in terms of SVaD measures as follows: min C,W J DGK (C, W ) = k j=1 x i R j M l=1 w jl g l (x i , c j ), subject to l w jl = j . Note that this problem can be solved using the same AO algorithms described in Section 3. Here, the update for C and P would remain the same as that discussed in Section 3. It can be easily shown that when j = 1, j, w jl = M m=1 x i R j g m (x i , c j ) 1/M x i R j g l (x i , c j ) (14) minimize J DGK for a given C. A.2 Crisp Simultaneous Clustering and Attribute Discrimination (CSCAD) Frigui et. al. in [5, 6], considered a fuzzy version of the feature-weighting based clustering problem (SCAD). To make a fair comparison of our algorithms with SCAD, we derive its crisp version and refer to it as Crisp SCAD (CSCAD). In [5, 6], the Gini measure is used for regularization. If the Gini measure is considered with r = 1, the weights w jl that minimize the corresponding objective function for a given C and P, are given by w jl = 1 M + 1 2 j 1 M M n=1 x i R j g n (x i , c j ) x i R j g l (x i , c j ) . (15) Since SCAD uses the weighted Euclidean measure, the update equations of centroids in CSCAD remain the same as in (11). The update equation for w jl in SCAD is quite similar to (15). One may note that, in (15), the value of w jl can become negative. In [5], a heuristic is used to estimate the value j in every iteration and set the negative values of w jl to zero before normalizing the weights. 616 Research Track Poster

Clustering;feature weighting;spatially varying dissimilarity (SVaD);Learning Dissimilarity Measures;clustering;dissimilarity measure

126

Learning the Unified Kernel Machines for Classification

Kernel machines have been shown as the state-of-the-art learning techniques for classification. In this paper, we propose a novel general framework of learning the Unified Kernel Machines (UKM) from both labeled and unlabeled data. Our proposed framework integrates supervised learning, semi-supervised kernel learning, and active learning in a unified solution. In the suggested framework, we particularly focus our attention on designing a new semi-supervised kernel learning method, i.e., Spectral Kernel Learning (SKL), which is built on the principles of kernel target alignment and unsupervised kernel design. Our algorithm is related to an equivalent quadratic programming problem that can be efficiently solved. Empirical results have shown that our method is more effective and robust to learn the semi-supervised kernels than traditional approaches. Based on the framework, we present a specific paradigm of unified kernel machines with respect to Kernel Logistic Regresions (KLR), i.e., Unified Kernel Logistic Regression (UKLR). We evaluate our proposed UKLR classification scheme in comparison with traditional solutions. The promising results show that our proposed UKLR paradigm is more effective than the traditional classification approaches.

INTRODUCTION Classification is a core data mining technique and has been actively studied in the past decades. In general, the goal of classification is to assign unlabeled testing examples with a set of predefined categories. Traditional classification methods are usually conducted in a supervised learning way, in which only labeled data are used to train a predefined classification model. In literature, a variety of statistical models have been proposed for classification in the machine learning and data mining communities. One of the most popular and successful methodologies is the kernel-machine techniques , such as Support Vector Machines (SVM) [25] and Kernel Logistic Regressions (KLR) [29]. Like other early work for classification, traditional kernel-machine methods are usually performed in the supervised learning way, which consider only the labeled data in the training phase. It is obvious that a good classification model should take advantages on not only the labeled data, but also the unlabeled data when they are available. Learning on both labeled and unlabeled data has become an important research topic in recent years. One way to exploit the unlabled data is to use active learning [7]. The goal of active learning is to choose the most informative example from the unlabeled data for manual labeling. In the past years, active learning has been studied for many classification tasks [16]. Another emerging popular technique to exploit unlabeled data is semi-supervised learning [5], which has attracted a surge of research attention recently [30]. A variety of machine-learning techniques have been proposed for semi-supervised learning, in which the most well-known approaches are based on the graph Laplacians methodology [28, 31, 5]. While promising results have been popularly reported in this research topic, there is so far few comprehensive semi-supervised learning scheme applicable for large-scale classification problems. Although supervised learning, semi-supervised learning and active learning have been studied separately, so far there is few comprehensive scheme to combine these techniques effectively together for classification tasks. To this end, we propose a general framework of learning the Unified Kernel Machines (UKM) [3, 4] by unifying supervised kernel-machine learning, semi-supervised learning, unsupervised kernel design and active learning together for large-scale classification problems. The rest of this paper is organized as follows. Section 2 reviews related work of our framework and proposed solutions. Section 3 presents our framework of learning the unified ker-187 Research Track Paper nel machines. Section 4 proposes a new algorithm of learning semi-supervised kernels by Spectral Kernel Learning (SKL). Section 5 presents a specific UKM paradigm for classification , i.e., the Unified Kernel Logistic Regression (UKLR). Section 6 evaluates the empirical performance of our proposed algorithm and the UKLR classification scheme. Section 7 sets out our conclusion. RELATED WORK Kernel machines have been widely studied for data classification in the past decade. Most of earlier studies on kernel machines usually are based on supervised learning. One of the most well-known techniques is the Support Vector Machines, which have achieved many successful stories in a variety of applications [25]. In addition to SVM, a series of kernel machines have also been actively studied, such as Kernel Logistic Regression [29], Boosting [17], Regularized Least-Square (RLS) [12] and Minimax Probability Machines (MPM) [15], which have shown comparable performance with SVM for classification. The main theoretical foundation behind many of the kernel machines is the theory of regularization and reproducing kernel Hilbert space in statistical learning [17, 25]. Some theoretical connections between the various kernel machines have been explored in recent studies [12]. Semi-supervised learning has recently received a surge of research attention for classification [5, 30]. The idea of semi-supervised learning is to use both labeled and unlabeled data when constructing the classifiers for classification tasks. One of the most popular solutions in semi-supervised learning is based on the graph theory [6], such as Markov random walks [22], Gaussian random fields [31], Diffusion models [13] and Manifold learning [2]. They have demonstrated some promising results on classification. Some recent studies have begun to seek connections between the graph-based semi-supervised learning and the kernel machine learning. Smola and Kondor showed some theoretical understanding between kernel and regularization based on the graph theory [21]. Belkin et al. developed a framework for regularization on graphs and provided some analysis on generalization error bounds [1]. Based on the emerging theoretical connections between kernels and graphs, some recent work has proposed to learn the semi-supervised kernels by graph Laplacians [32]. Zhang et al. recently provided a theoretical framework of unsupervised kernel design and showed that the graph Laplacians solution can be considered as an equivalent kernel learning approach [27]. All of the above studies have formed the solid foundation for semi-supervised kernel learning in this work. To exploit the unlabeled data, another research attention is to employ active learning for reducing the labeling efforts in classification tasks. Active learning, or called pool-based active learning, has been proposed as an effective technique for reducing the amount of labeled data in traditional supervised classification tasks [19]. In general, the key of active learning is to choose the most informative unlabeled examples for manual labeling. A lot of active learning methods have been proposed in the community. Typically they measure the classification uncertainty by the amount of disagreement to the classification model [9, 10] or measure the distance of each unlabeled example away from the classification boundary [16, 24]. FRAMEWORK OF LEARNING UNIFIED KERNEL MACHINES In this section, we present the framework of learning the unified kernel machines by combining supervised kernel machines , semi-supervised kernel learning and active learning techniques into a unified solution. Figure 1 gives an overview of our proposed scheme. For simplicity, we restrict our discussions to classification problems. Let M(K, ) denote a kernel machine that has some underlying probabilistic model, such as kernel logistic regressions (or support vector machines). In general, a kernel machine contains two components, i.e., the kernel K (either a kernel function or simply a kernel matrix), and the model parameters . In traditional supervised kernel-machine learning , the kernel K is usually a known parametric kernel function and the goal of the learning task is usually to determine the model parameter . This often limits the performance of the kernel machine if the specified kernel is not appropriate. To this end, we propose a unified scheme to learn the unified kernel machines by learning on both the kernel K and the model parameters together. In order to exploit the unlabeled data, we suggest to combine semi-supervised kernel learning and active learning techniques together for learning the unified kernel machines effectively from the labeled and unlabeled data. More specifically, we outline a general framework of learning the unified kernel machine as follows. Figure 1: Learning the Unified Kernel Machines Let L denote the labeled data and U denote the unlabeled data. The goal of the unified kernel machine learning task is to learn the kernel machine M(K , ) that can classify the data effectively. Specifically, it includes the following five steps: Step 1. Kernel Initialization The first step is to initialize the kernel component K 0 of the kernel machine M(K 0 , 0 ). Typically, users can specify the initial kernel K 0 (function or matrix) with a stanard kernel. When some domain knowledge is ava-iable , users can also design some kernel with domain knowledge (or some data-dependent kernels). Step 2. Semi-Supervised Kernel Learning The initial kernel may not be good enough to classify the data correctly. Hence, we suggest to employ 188 Research Track Paper the semi-supervised kernel learning technique to learn a new kernel K by engaging both the labeled L and unlabled data U available. Step 3. Model Parameter Estimation When the kernel K is known, to estimate the parameters of the kernel machines based on some model assumption , such as Kernel Logistic Regression or Support Vector Machines, one can simply employ the standard supervised kernel-machine learning to solve the model parameters . Step 4. Active Learning In many classification tasks, labeling cost is expensive. Active learning is an important method to reduce human efforts in labeling. Typically, we can choose a batch of most informative examples S that can most effectively update the current kernel machine M(K, ). Step 5. Convergence Evaluation The last step is the convergence evaluation in which we check whether the kernel machine is good enough for the classification task. If not, we will repeat the above steps until a satisfied kernel machine is acquired. This is a general framework of learning unified kernel machines . In this paper, we focus our main attention on the the part of semi-supervised kernel learning technique, which is a core component of learning the unified kernel machines. SPECTRAL KERNEL LEARNING We propose a new semi-supervised kernel learning method, which is a fast and robust algorithm for learning semi-supervised kernels from labeled and unlabeled data. In the following parts, we first introduce the theoretical motivations and then present our spectral kernel learning algorithm. Finally, we show the connections of our method to existing work and justify the effectiveness of our solution from empirical observations . 4.1 Theoretical Foundation Let us first consider a standard supservisd kernel learning problem. Assume that the data (X, Y ) are drawn from an unknown distribution D. The goal of supervised learning is to find a prediction function p(X) that minimizes the following expected true loss: E (X,Y )D L(p(X), Y ), where E (X,Y )D denotes the expectation over the true underlying distribution D. In order to achieve a stable esimia-tion , we usually need to restrict the size of hypothesis function family. Given l training examples (x 1 ,y 1 ),. . .,(x l ,y l ), typically we train a predition function ^ p in a reproducing Hilbert space H by minimizing the empirical loss [25]. Since the reproducing Hilbert space can be large, to avoid over-fitting problems, we often consider a regularized method as follow: ^ p = arg inf pH 1 l l i=1 L(p(x i ), y i ) + ||p|| 2 H , (1) where is a chosen positive regularization parameter. It can be shown that the solution of (1) can be represented as the following kernel method: ^ p(x) = l i=1 ^ i k(x i , x) = arg inf R l 1 n l i=1 L (p(x i ), y i ) + l i,j=1 i j k(x i , x j ) , where is a parameter vector to be estimated from the data and k is a kernel, which is known as kernel function . Typically a kernel returns the inner product between the mapping images of two given data examples, such that k(x i , x j ) = (x i ), (x j ) for x i , x j X . Let us now consider a semi-supervised learning setting. Given labeled data {(x i , y i ) } l i=1 and unlabeled data {x j } n j=l+1 , we consider to learn the real-valued vectors f R m by the following semi-supervised learning method: ^ f = arg inf f R 1 n n i=1 L(f i , y i ) + f K -1 f , (2) where K is an m m kernel matrix with K i,j = k(x i , x j ). Zhang et al. [27] proved that the solution of the above semi-supervised learning is equivelent to the solution of standard supervised learning in (1), such that ^ f j = ^ p(x j ) j = 1, . . . , m. (3) The theorem offers a princple of unsuperivsed kernel design : one can design a new kernel k( , ) based on the unlabeled data and then replace the orignal kernel k by k in the standard supervised kernel learning. More specifically, the framework of spectral kernel design suggests to design the new kernel matrix K by a function g as follows: K = n i=1 g( i )v i v i , (4) where ( i , v i ) are the eigen-pairs of the original kernel matrix K, and the function g( ) can be regarded as a filter function or a transformation function that modifies the spectra of the kernel. The authors in [27] show a theoretical justification that designing a kernel matrix with faster spectral decay rates should result in better generalization performance, which offers an important pricinple in learning an effective kernel matrix. On the other hand, there are some recent papers that have studied theoretical principles for learning effective kernel functions or matrices from labeled and unlabeled data. One important work is the kernel target alignment, which can be used not only to assess the relationship between the feature spaces by two kernels, but also to measure the similarity between the feature space by a kernel and the feature space induced by labels [8]. Specifically, given two kernel matrices K 1 and K 2 , their relationship is defined by the following score of alignment: Definition 1. Kernel Alignment: The empirical alignment of two given kernels K 1 and K 2 with respect to the sample set S is the quantity: ^ A(K 1 , K 2 ) = K 1 , K 2 F K 1 , K 1 F K 2 , K 2 F (5) 189 Research Track Paper where K i is the kernel matrix induced by the kernel k i and , is the Frobenius product between two matrices, i.e., K 1 , K 2 F = n i,j=1 k 1 (x i , x j )k 2 (x i , x j ). The above definition of kernel alignment offers a principle to learn the kernel matrix by assessing the relationship between a given kernel and a target kernel induced by the given labels. Let y = {y i } l i=1 denote a vector of labels in which y i {+1, -1} for binary classification. Then the target kernel can be defined as T = yy . Let K be the kernel matrix with the following structure K = K tr K trt K trt K t (6) where K ij = (x i ), (x j ) , K tr denotes the matrix part of "train-data block" and K t denotes the matrix part of "test-data block." The theory in [8] provides the principle of learning the kernel matrix, i.e., looking for a kernel matrix K with good generalization performance is equivalent to finding the matrix that maximizes the following empirical kernel alignment score: ^ A(K tr , T ) = K tr , T F K tr , K tr F T, T F (7) This principle has been used to learn the kernel matrices with multiple kernel combinations [14] and also the semi-supervised kernels from graph Laplacians [32]. Motivated by the related theorecial work, we propose a new spectral kernel learning (SKL) algorithm which learns spectrals of the kernel matrix by obeying both the principle of unsupervised kernel design and the principle of kernel target alignment. 4.2 Algorithm Assume that we are given a set of labeled data L = {x i , y i } l i=1 , a set of unlabeled data U = {x i } n i=l+1 , and an initial kernel matrix K. We first conduct the eigen-decomposition of the kernel matrix: K = n i=1 i v i v i , (8) where ( i , v i ) are eigen pairs of K and are assumed in a decreasing order, i.e., 1 2 . . . n . For efficiency consideration, we select the top d eigen pairs, such that K d = d i=1 i v i v i K , (9) where the parameter d n is a dimension cutoff factor that can be determined by some criteria, such as the cumulative eigen energy. Based on the principle of unsupervised kernel design, we consider to learn the kernel matrix as follows K = d i=1 i v i v i , (10) where i 0 are spectral coefficients of the new kernel matrix . The goal of spectral kernel learning (SKL) algorithm is to find the optimal spectral coefficients i for the following optimization max K, ^ A( K tr , T ) (11) subject to K = d i=1 i v i v i trace( K) = 1 i 0, i C i+1 , i = 1 . . . d - 1 , where C is introduced as a decay factor that satisfies C 1, v i are top d eigen vectors of the original kernel matrix K, K tr is the kernel matrix restricted to the (labeled) training data and T is the target kernel induced by labels. Note that C is introduced as an important parameter to control the decay rate of spectral coefficients that will influence the overall performance of the kernel machine. The above optimization problem belongs to convex optimization and is usually regarded as a semi-definite programming problem (SDP) [14], which may not be computation-ally efficient. In the following, we turn it into a Quadratic Programming (QP) problem that can be solved much more efficiently. By the fact that the objective function (7) is invariant to the constant term T, T F , we can rewrite the objective function into the following form K tr , T F K tr , K tr F . (12) The above alignment is invariant to scales. In order to remove the trace constraint in (11), we consider the following alternative approach. Instead of maximizing the objective function (12) directly, we can fix the numerator to 1 and then minimize the denominator. Therefore, we can turn the optimization problem into: min K tr , K tr F (13) subject to K = d i=1 i v i v i K tr , T F = 1 i 0, i C i+1 , i = 1 . . . d - 1 . This minimization problem without the trace constraint is equivalent to the original maximization problem with the trace constraint. Let vec(A) denote the column vectorization of a matrix A and let D = [vec(V 1,tr ) . . . vec(V d,tr )] be a constant matrix with size of l 2 d, in which the d matrices of V i = v i v i are with size of l l. It is not difficult to show that the above problem is equivalent to the following optimization min ||D|| (14) subject to vec(T ) D = 1 i 0 i C i+1 , i = 1 . . . d - 1 . Minimizing the norm is then equivalent to minimizing the squared norm. Hence, we can obtain the final optimization 190 Research Track Paper 0 5 10 15 20 25 30 0.4 0.5 0.6 0.7 0.8 0.9 1 Dimension (d) Cumulative Energy (a) Cumulative eigen energy 0 5 10 15 20 25 30 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Dimension (d) Scaled Coefficient Original Kernel SKL (C=1) SKL (C=2) SKL (C=3) (b) Spectral coefficients Figure 2: Illustration of cumulative eigen energy and the spectral coefficients of different decay factors on the Ionosphere dataset. The initial kernel is a linear kernel and the number of labeled data is 20. 0 10 20 30 40 50 0.65 0.7 0.75 0.8 0.85 0.9 0.95 Dimension (d) Accuracy K Origin K Trunc K Cluster K Spectral (a) C=1 0 5 10 15 20 25 30 35 40 45 50 0.65 0.7 0.75 0.8 0.85 0.9 0.95 Dimension (d) Accuracy K Origin K Trunc K Cluster K Spectral (b) C=2 0 5 10 15 20 25 30 35 40 45 50 0.65 0.7 0.75 0.8 0.85 0.9 0.95 Dimension (d) Accuracy K Origin K Trunc K Cluster K Spectral (c) C=3 Figure 3: Classification performance of semi-supervised kernels with different decay factors on the Ionosphere dataset. The initial kernel is a linear kernel and the number of labeled data is 20. problem as min D D subject to vec(T ) D = 1 i 0 i C i+1 , i = 1 . . . d - 1 . This is a standard Quadratic Programming (QP) problem that can be solved efficiently. 4.3 Connections and Justifications The essential of our semi-supervised kernel learning method is based on the theories of unsupervised kernel design and kernel target alignment. More specifically, we consider a dimension-reduction effective method to learn the semi-supervised kernel that maximizes the kernel alignment score. By examining the work on unsupervised kernel design, the following two pieces of work can be summarized as a special case of spectral kernel learning framework: Cluster Kernel This method adopts a "[1. . . ,1,0,. . . ,0]" kernel that has been used in spectral clustering [18]. It sets the top spectral coefficients to 1 and the rest to 0, i.e., i = 1 for i d 0 for i &gt; d . (15) For a comparison, we refer to this method as "Cluster kernel" denoted by K Cluster . Truncated Kernel Another method is called the truncated kernel that keeps only the top d spectral coefficients i = i for i d 0 for i &gt; d , (16) where i are the top eigen values of an initial kernel. We can see that this is exactly the method of kernel principal component analysis [20] that keeps only the d most significant principal components of a given kernel. For a comparison, we denote this method as K Trunc . 191 Research Track Paper 0 5 10 15 20 25 30 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Dimension (d) Scaled Coefficient Original Kernel SKL (C=1) SKL (C=2) SKL (C=3) (a) Spectral coefficients 0 10 20 30 40 50 0.7 0.75 0.8 0.85 0.9 0.95 1 Dimension (d) Accuracy K Origin K Trunc K Cluster K Spectral (b) C=1 0 10 20 30 40 50 0.7 0.75 0.8 0.85 0.9 0.95 1 Dimension (d) Accuracy K Origin K Trunc K Cluster K Spectral (c) C=2 Figure 4: Example of Spectral coefficients and performance impacted by different decay factors on the Ionosphere dataset. The initial kernel is an RBF kernel and the number of labeled data is 20. 0 10 20 30 40 50 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 Dimension (d) Accuracy K Origin K Trunc K Cluster K Spectral (a) C=1 0 5 10 15 20 25 30 35 40 45 50 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 Dimension (d) Accuracy K Origin K Trunc K Cluster K Spectral (b) C=2 0 5 10 15 20 25 30 35 40 45 50 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 Dimension (d) Accuracy K Origin K Trunc K Cluster K Spectral (c) C=3 Figure 5: Classification performance of semi-supervised kernels with different decay factors on the Heart dataset. The initial kernel is a linear kernel and the number of labeled data is 20. In our case, in comparison with semi-supervised kernel learning methods by graph Laplacians, our work is similar to the approach in [32], which learns the spectral transformation of graph Laplacians by kernel target alignment with order constraints. However, we should emphasize two important differences that will explain why our method can work more effectively. First, the work in [32] belongs to traditional graph based semi-supervised learning methods which assume the kernel matrix is derived from the spectral decomposition of graph Laplacians. Instead, our spectral kernel learning method learns on any initial kernel and assume the kernel matrix is derived from the spectral decomposition of the normalized kernel. Second, compared to the kernel learning method in [14], the authors in [32] proposed to add order constraints into the optimization of kernel target alignment [8] to enforce the constraints of graph smoothness. In our case, we suggest a decay factor C to constrain the relationship of spectral coefficients in the optimization that can make the spectral coefficients decay faster. In fact, if we ignore the difference of graph Laplacians and assume that the initial kernel in our method is given as K L -1 , we can see that the method in [32] can be regarded as a special case of our method when the decay factor C is set to 1 and the dimension cut-off parameter d is set to n. 4.4 Empirical Observations To argue that C = 1 in the spectral kernel learning algorithm may not be a good choice for learning an effective kernel, we illustrate some empirical examples to justifiy the motivation of our spectral kernel learning algorithm. One goal of our spectral kernel learning methodology is to attain a fast decay rate of the spectral coefficients of the kernel matrix. Figure 2 illustrates an example of the change of the resulting spectral coefficients using different decay factors in our spectral kernel learning algorithms. From the figure, we can see that the curves with larger decay factors (C = 2, 3) have faster decay rates than the original kernel and the one using C = 1. Meanwhile, we can see that the cumulative eigen energy score converges to 100% quickly when the number of dimensions is increased. This shows that we may use much small number of eigen-pairs in our semi-supervised kernel learning algorithm for large-scale problems. To examine more details in the impact of performance with different decay factors, we evaluate the classification 192 Research Track Paper performance of spectral kernel learning methods with different decay factors in Figure 3. In the figure, we compare the performance of different kernels with respect to spectral kernel design methods. We can see that two unsupervised kernels, K Trunc and K Cluster , tend to perform better than the original kernel when the dimension is small. But their performances are not very stable when the number of dimensions is increased. For comparison, the spectral kernel learning method achieves very stable and good performance when the decay factor C is larger than 1. When the decay factor is equal to 1, the performance becomes unstable due to the slow decay rates observed from our previous results in Figure 3. This observation matches the theoretical justification [27] that a kernel with good performance usually favors a faster decay rate of spectral coefficients. Figure 4 and Figure 5 illustrate more empirical examples based on different initial kernels, in which similar results can be observed. Note that our suggested kernel learning method can learn on any valid kernel, and different initial kernels will impact the performance of the resulting spectral kernels. It is usually helpful if the initial kernel is provided with domain knowledge. UNIFIED KERNEL LOGISTIC REGRESSION In this section, we present a specific paradigm based on the proposed framework of learning unified kernel machines. We assume the underlying probabilistic model of the kernel machine is Kernel Logistic Regression (KLR). Based on the UKM framework, we develop the Unified Kernel Logistic Regression (UKLR) paradigm to tackle classification tasks. Note that our framework is not restricted to the KLR model, but also can be widely extended for many other kernel machines, such as Support Vector Machine (SVM) and Regularized Least-Square (RLS) classifiers. Similar to other kernel machines, such as SVM, a KLR problem can be formulated in terms of a stanard regularized form of loss+penalty in the reproducing kernel Hilbert space (RKHS): min f H K 1 l l i=1 ln(1 + e -y i f (x i ) ) + 2 ||f || 2 H K , (17) where H K is the RKHS by a kernel K and is a regularization parameter. By the representer theorem, the optimal f (x) has the form: f (x) = l i=1 i K(x, x i ) , (18) where i are model parameters. Note that we omit the constant term in f (x) for simplified notations. To solve the KLR model parameters, there are a number of available techniques for effective solutions [29]. When the kernel K and the model parameters are available , we use the following solution for active learning, which is simple and efficient for large-scale problems. More specifically , we measure the information entropy of each unlabeled data example as follows H(x; , K) = N C i=1 p(C i |x)log(p(C i |x)) , (19) Algorithm: Unified Kernel Logistic Regresssion Input K 0 : Initial normalized kernel L: Set of labeled data U: Set of unlabeled data Repeat Spectral Kernel Learning K Spectral Kernel(K 0 , L, U ); KLR Parameter Estimation KLR Solver(L, K); Convergence Test If (converged), Exit Loop; Active Learning x max xU H(x; , K) L L {x }, U U - {x } Until converged. Output UKLR = M(K, ). Figure 6: The UKLR Algorithm. where N C is the number of classes and C i denotes the i th class and p(C i |x) is the probability of the data example x belonging to the i th class which can be naturally obtained by the current KLR model (, K). The unlabeled data examples with maximum values of entropy will be considered as the most informative data for labeling. By unifying the spectral kernel learning method proposed in Section 3, we summarize the proposed algorithm of Unified Kernel Logistic Regression (UKLR) in Figure 6. In the algorithm, note that we can usually initialize a kernel by a standard kernel with appropriate parameters determined by cross validation or by a proper deisgn of the initial kernel with domain knowledge. EXPERIMENTAL RESULTS We discuss our empirical evaluation of the proposed framework and algorithms for classification. We first evaluate the effectiveness of our suggested spectral kernel learning algorithm for learning semi-supervised kernels and then compare the performance of our unified kernel logistic regression paradigm with traditional classification schemes. 6.1 Experimental Testbed and Settings We use the datasets from UCI machine learning repository 1 . Four datasets are employed in our experiments. Table 1 shows the details of four UCI datasets in our experiments . For experimental settings, to examine the influences of different training sizes, we test the compared algorithms on four different training set sizes for each of the four UCI datasets. For each given training set size, we conduct 20 random trials in which a labeled set is randomly sampled 1 www.ics.uci.edu/ mlearn/MLRepository.html 193 Research Track Paper Table 1: List of UCI machine learning datasets. Dataset #Instances #Features #Classes Heart 270 13 2 Ionosphere 351 34 2 Sonar 208 60 2 Wine 178 13 3 from the whole dataset and all classes must be present in the sampled labeled set. The rest data examples of the dataset are then used as the testing (unlabeled) data. To train a classifier, we employ the standard KLR model for classification. We choose the bounds on the regularization parameters via cross validation for all compared kernels to avoid an unfair comparison. For multi-class classification, we perform one-against-all binary training and testing and then pick the class with the maximum class probability. 6.2 Semi-Supervised Kernel Learning In this part, we evaluate the performance of our spectral kernel learning algorithm for learning semi-supervised kernels . We implemented our algorithm by a standard Matlab Quadratic Programming solver (quadprog). The dimension-cut parameter d in our algorithm is simply fixed to 20 without further optimizing. Note that one can easily determine an appropriate value of d by examining the range of the cumulative eigen energy score in order to reduce the com-putational cost for large-scale problems. The decay factor C is important for our spectral kernel learning algorithm. As we have shown examples before, C must be a positive real value greater than 1. Typically we favor a larger decay factor to achieve better performance. But it must not be set too large since the too large decay factor may result in the overly stringent constraints in the optimization which gives no solutions. In our experiments, C is simply fixed to constant values (greater than 1) for the engaged datasets. For a comparison, we compare our SKL algorithms with the state-of-the-art semi-supervised kernel learning method by graph Laplacians [32], which is related to a quadrati-cally constrained quaratic program (QCQP). More specifically , we have implemented two graph Laplacians based semi-supervised kernels by order constraints [32]. One is the order-constrained graph kernel (denoted as "Order") and the other is the improved order-constrained graph kernel (denoted as "Imp-Order"), which removes the constraints from constant eigenvectors. To carry a fair comparison, we use the top 20 smallest eigenvalues and eigenvectors from the graph Laplacian which is constructed with 10-NN un-weighted graphs. We also include three standard kernels for comparisons. Table 2 shows the experimental results of the compared kernels (3 standard and 5 semi-supervised kernels) based on KLR classifiers on four UCI datasets with different sizes of labeled data. Each cell in the table has two rows: the upper row shows the average testing set accruacies with standard errors; and the lower row gives the average run time in seconds for learning the semi-supervised kernels on a 3GHz desktop computer. We conducted a paired t-test at significance level of 0.05 to assess the statistical significance of the test set accuracy results. From the experimental results, we found that the two order-constrained based graph kernels perform well in the Ionosphere and Wine datasets, but they do not achieve important improvements on the Heart and Sonar datasets. Among all the compared kernels, the semi-supervised kernels by our spectral kernel learning algorithms achieve the best performances. The semi-supervised kernel initialized with an RBF kernel outperforms other kernels in most cases. For example, in Ionosphere dataset, an RBF kernel with 10 initial training examples only achieves 73.56% test set accuracy, and the SKL algorithm can boost the accuracy significantly to 83.36%. Finally, looking into the time performance, the average run time of our algorithm is less than 10% of the previous QCQP algorithms. 6.3 Unified Kernel Logistic Regression In this part, we evaluate the performance of our proposed paradigm of unified kernel logistic regression (UKLR). As a comparison, we implement two traditional classification schemes: one is traditional KLR classification scheme that is trained on randomly sampled labeled data, denoted as "KLR+Rand." The other is the active KLR classification scheme that actively selects the most informative examples for labeling, denoted as "KLR+Active." The active learning strategy is based on a simple maximum entropy criteria given in the pervious section. The UKLR scheme is implemented based on the algorithm in Figure 6. For active learning evaluation, we choose a batch of 10 most informative unlabeled examples for labeling in each trial of evaluations. Table 3 summarizes the experimental results of average test set accuarcy performances on four UCI datasets. From the experimental results, we can observe that the active learning classification schems outperform the randomly sampled classification schemes in most cases. This shows the suggested simple active learning strategy is effectiveness. Further, among all compared schemes, the suggsted UKLR solution significantly outperforms other classification approaches in most cases. These results show that the unified scheme is effective and promising to integrate traditional learning methods together in a unified solution . 6.4 Discussions Although the experimental results have shown that our scheme is promising, some open issues in our current solution need be further explored in future work. One problem to investigate more effective active learning methods in selecting the most informative examples for labeling. One solution to this issue is to employ the batch mode active learning methods that can be more efficient for large-scale classification tasks [11, 23, 24]. Moreover, we will study more effective kernel learning algorithms without the assumption of spectral kernels. Further, we may examine the theoretical analysis of generalization performance of our method [27]. Finally, we may combine some kernel machine speedup techniques to deploy our scheme efficiently for large-scale applications [26]. CONCLUSION This paper presented a novel general framework of learning the Unified Kernel Machines (UKM) for classification. Different from traditional classification schemes, our UKM framework integrates supervised learning, semi-supervised learning, unsupervised kernel design and active learning in a unified solution, making it more effective for classification tasks. For the proposed framework, we focus our attention on tackling a core problem of learning semi-supervised kernels from labeled and unlabled data. We proposed a Spectral 194 Research Track Paper Table 2: Classification performance of different kernels using KLR classifiers on four datasets. The mean accuracies and standard errors are shown in the table. 3 standard kernels and 5 semi-supervised kernels are compared. Each cell in the table has two rows. The upper row shows the test set accuracy with standard error; the lower row gives the average time used in learning the semi-supervised kernels ("Order" and "Imp-Order" kernels are sovled by SeDuMi/YALMIP package; "SKL" kernels are solved directly by the Matlab quadprog function. Train Standard Kernels Semi-Supervised Kernels Size Linear Quadratic RBF Order Imp-Order SKL(Linear) SKL(Quad) SKL(RBF) Heart 10 67.19 1.94 71.90 1.23 70.04 1.61 63.60 1.94 63.60 1.94 70.58 1.63 72.33 1.60 73.37 1.50 ( 0.67 ) ( 0.81 ) ( 0.07 ) ( 0.06 ) ( 0.06 ) 20 67.40 1.87 70.36 1.51 72.64 1.37 65.88 1.69 65.88 1.69 76.26 1.29 75.36 1.30 76.30 1.33 ( 0.71 ) ( 0.81 ) ( 0.06 ) ( 0.06 ) ( 0.06 ) 30 75.42 0.88 70.71 0.83 74.40 0.70 71.73 1.14 71.73 1.14 78.42 0.59 78.65 0.52 79.23 0.58 ( 0.95 ) ( 0.97 ) ( 0.06 ) ( 0.06 ) ( 0.06 ) 40 78.24 0.89 71.28 1.10 78.48 0.77 75.48 0.69 75.48 0.69 80.61 0.45 80.26 0.45 80.98 0.51 ( 1.35 ) ( 1.34 ) ( 0.07 ) ( 0.07 ) ( 0.07 ) Ionosphere 10 73.71 1.27 71.30 1.70 73.56 1.91 71.86 2.79 71.86 2.79 75.53 1.75 71.22 1.82 83.36 1.31 ( 0.90 ) ( 0.87 ) ( 0.05 ) ( 0.05 ) ( 0.05 ) 20 75.62 1.24 76.00 1.58 81.71 1.74 83.04 2.10 83.04 2.10 78.78 1.60 80.30 1.77 88.55 1.32 ( 0.87 ) ( 0.79 ) ( 0.05 ) ( 0.06 ) ( 0.05 ) 30 76.59 0.82 79.10 1.46 86.21 0.84 87.20 1.16 87.20 1.16 82.18 0.56 83.08 1.36 90.39 0.84 ( 0.93 ) ( 0.97 ) ( 0.05 ) ( 0.05 ) ( 0.05 ) 40 77.97 0.79 82.93 1.33 89.39 0.65 90.56 0.64 90.56 0.64 83.26 0.53 87.03 1.02 92.14 0.46 ( 1.34 ) ( 1.38 ) ( 0.05 ) ( 0.04 ) ( 0.04 ) Sonar 10 63.01 1.47 62.85 1.53 60.76 1.80 59.67 0.89 59.67 0.89 64.27 1.91 64.37 1.64 65.30 1.78 ( 0.63 ) ( 0.63 ) ( 0.08 ) ( 0.07 ) ( 0.07 ) 20 68.09 1.11 69.55 1.22 67.63 1.15 64.68 1.57 64.68 1.57 70.61 1.14 69.79 1.30 71.76 1.07 ( 0.68 ) ( 0.82 ) ( 0.07 ) ( 0.07 ) ( 0.08 ) 30 66.40 1.06 69.80 0.93 68.23 1.48 66.54 0.79 66.54 0.79 70.20 1.48 68.48 1.59 71.69 0.87 ( 0.88 ) ( 1.02 ) ( 0.07 ) ( 0.07 ) ( 0.07 ) 40 64.94 0.74 71.37 0.52 71.61 0.89 69.82 0.82 69.82 0.82 72.35 1.06 71.28 0.96 72.89 0.68 ( 1.14 ) ( 1.20 ) ( 0.07 ) ( 0.08 ) ( 0.07 ) Wine 10 82.26 2.18 85.89 1.73 87.80 1.63 86.99 1.98 86.99 1.45 83.63 2.62 83.21 2.36 90.54 1.08 ( 1.02 ) ( 0.86 ) ( 0.09 ) ( 0.09 ) ( 0.09 ) 20 86.39 1.39 86.96 1.30 93.77 0.99 92.31 1.39 92.31 1.39 89.53 2.32 92.56 0.56 94.94 0.50 ( 0.92 ) ( 0.91 ) ( 0.09 ) ( 0.09 ) ( 0.09 ) 30 92.50 0.76 87.43 0.63 94.63 0.50 92.97 0.54 92.97 0.54 93.99 1.09 94.29 0.53 96.25 0.30 ( 1.28 ) ( 1.27 ) ( 0.09 ) ( 0.10 ) ( 0.09 ) 40 94.96 0.65 88.80 0.93 96.38 0.35 95.62 0.37 95.62 0.37 95.80 0.47 95.36 0.46 96.81 0.28 ( 1.41 ) ( 1.39 ) ( 0.08 ) ( 0.08 ) ( 0.10 ) Kernel Learning (SKL) algorithm, which is more effective and efficient for learning kernels from labeled and unlabeled data. Under the framework, we developed a paradigm of unified kernel machine based on Kernel Logistic Regression, i.e., Unified Kernel Logistic Regression (UKLR). Empirical results demonstrated that our proposed solution is more effective than the traditional classification approaches. ACKNOWLEDGMENTS The work described in this paper was fully supported by two grants, one from the Shun Hing Institute of Advanced Engineering, and the other from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. CUHK4205/04E). REFERENCES [1] M. Belkin and I. M. andd P. Niyogi. Regularization and semi-supervised learning on large graphs. In COLT, 2004. [2] M. Belkin and P. Niyogi. Semi-supervised learning on riemannian manifolds. Machine Learning, 2004. [3] E. Chang, S. C. Hoi, X. Wang, W.-Y. Ma, and M. Lyu. A unified machine learning framework for large-scale personalized information management. In The 5th Emerging Information Technology Conference, NTU Taipei, 2005. [4] E. Chang and M. Lyu. Unified learning paradigm for web-scale mining. In Snowbird Machine Learning Workshop, 2006. [5] O. Chapelle, A. Zien, and B. Scholkopf. Semi-supervised learning. MIT Press, 2006. [6] F. R. K. Chung. Spectral Graph Theory. American Mathematical Soceity, 1997. [7] D. A. Cohn, Z. Ghahramani, and M. I. Jordan. Active learning with statistical models. In NIPS, volume 7, pages 705712, 1995. [8] N. Cristianini, J. Shawe-Taylor, and A. Elisseeff. On kernel-target alignment. JMLR, 2002. 195 Research Track Paper Table 3: Classification performance of different classification schemes on four UCI datasets. The mean accuracies and standard errors are shown in the table. "KLR" represents the initial classifier with the initial train size; other three methods are trained with additional 10 random/active examples. Train Linear Kernel RBF Kernel Size KLR KLR+Rand KLR+Active UKLR KLR KLR+Rand KLR+Active UKLR Heart 10 67.19 1.94 68.22 2.16 69.22 1.71 77.24 0.74 70.04 1.61 72.24 1.23 75.36 0.60 78.44 0.88 20 67.40 1.87 73.79 1.29 73.77 1.27 79.27 1.00 72.64 1.37 75.10 0.74 76.23 0.81 79.88 0.90 30 75.42 0.88 77.70 0.92 78.65 0.62 81.13 0.42 74.40 0.70 76.43 0.68 76.61 0.61 81.48 0.41 40 78.24 0.89 79.30 0.75 80.18 0.79 82.55 0.28 78.48 0.77 78.50 0.53 79.95 0.62 82.66 0.36 Ionosphere 10 73.71 1.27 74.89 0.95 75.91 0.96 77.31 1.23 73.56 1.91 82.57 1.78 82.76 1.37 90.48 0.83 20 75.62 1.24 77.09 0.67 77.51 0.66 81.42 1.10 81.71 1.74 85.95 1.30 88.22 0.78 91.28 0.94 30 76.59 0.82 78.41 0.79 77.91 0.77 84.49 0.37 86.21 0.84 89.04 0.66 90.32 0.56 92.35 0.59 40 77.97 0.79 79.05 0.49 80.30 0.79 84.49 0.40 89.39 0.65 90.55 0.59 91.83 0.49 93.89 0.28 Sonar 10 61.19 1.56 63.72 1.65 65.51 1.55 66.12 1.94 57.40 1.48 60.19 1.32 59.49 1.46 67.13 1.58 20 67.31 1.07 68.85 0.84 69.38 1.05 71.60 0.91 62.93 1.36 64.72 1.24 64.52 1.07 72.30 0.98 30 66.10 1.08 67.59 1.14 69.79 0.86 71.40 0.80 63.03 1.32 63.72 1.51 66.67 1.53 72.26 0.98 40 66.34 0.82 68.16 0.81 70.19 0.90 73.04 0.69 66.70 1.25 68.70 1.19 67.56 0.90 73.16 0.88 Wine 10 82.26 2.18 87.31 1.01 89.05 1.07 87.31 1.03 87.80 1.63 92.75 1.27 94.49 0.54 94.87 0.49 20 86.39 1.39 93.99 0.40 93.82 0.71 94.43 0.54 93.77 0.99 95.57 0.38 97.13 0.18 96.76 0.26 30 92.50 0.76 95.25 0.47 96.96 0.40 96.12 0.47 94.63 0.50 96.27 0.35 97.17 0.38 97.21 0.26 40 94.96 0.65 96.21 0.63 97.54 0.37 97.70 0.34 96.38 0.35 96.33 0.45 97.97 0.23 98.12 0.21 [9] S. Fine, R. Gilad-Bachrach, and E. Shamir. Query by committee, linear separation and random walks. Theor. Comput. Sci., 284(1):2551, 2002. [10] Y. Freund, H. S. Seung, E. Shamir, and N. Tishby. Selective sampling using the query by committee algorithm. Mach. Learn., 28(2-3):133168, 1997. [11] S. C. Hoi, R. Jin, and M. R. Lyu. Large-scale text categorization by batch mode active learning. In WWW2006, Edinburg, 2006. [12] S. B. C. M. J.A.K. Suykens, G. Horvath and J. Vandewalle. Advances in Learning Theory: Methods, Models and Applications. NATO Science Series: Computer & Systems Sciences, 2003. [13] R. Kondor and J. Lafferty. Diffusion kernels on graphs and other discrete structures. 2002. [14] G. Lanckriet, N. Cristianini, P. Bartlett, L. E. Ghaoui, and M. Jordan. Learning the kernel matrix with semi-definite programming. JMLR, 5:2772, 2004. [15] G. Lanckriet, L. Ghaoui, C. Bhattacharyya, and M. Jordan. Minimax probability machine. In Advances in Neural Infonation Processing Systems 14, 2002. [16] R. Liere and P. Tadepalli. Active learning with committees for text categorization. In Proceedings 14th Conference of the American Association for Artificial Intelligence (AAAI), pages 591596, MIT Press, 1997. [17] R. Meir and G. Ratsch. An introduction to boosting and leveraging. In In Advanced Lectures on Machine Learning (LNAI2600), 2003. [18] A. Ng, M. Jordan, and Y. Weiss. On spectral clustering: Analysis and an algorithm. In In Advances in Neural Information Processing Systems 14, 2001. [19] N. Roy and A. McCallum. Toward optimal active learning through sampling estimation of error reduction. In 18th ICML, pages 441448, 2001. [20] B. Scholkopf, A. Smola, and K.-R. Muller. Nonlinear component analysis as a kernel eigenvalue problem. Neural Computation, 10:12991319, 1998. [21] A. Smola and R. Kondor. Kernels and regularization on graphs. In Intl. Conf. on Learning Theory, 2003. [22] M. Szummer and T. Jaakkola. Partially labeled classification with markov random walks. In Advances in Neural Information Processing Systems, 2001. [23] S. Tong and E. Chang. Support vector machine active learning for image retrieval. In Proc ACM Multimedia Conference, pages 107118, New York, 2001. [24] S. Tong and D. Koller. Support vector machine active learning with applications to text classification. In Proc. 17th ICML, pages 9991006, 2000. [25] V. N. Vapnik. Statistical Learning Theory. John Wiley & Sons, 1998. [26] G. Wu, Z. Zhang, and E. Y. Chang. Kronecker factorization for speeding up kernel machines. In SIAM Int. Conference on Data Mining (SDM), 2005. [27] T. Zhang and R. K. Ando. Analysis of spectral kernel design based semi-supervised learning. In NIPS, 2005. [28] D. Zhou, O. Bousquet, T. Lal, J. Weston, and B. Schlkopf. Learning with local and global consistency. In NIPS'16, 2005. [29] J. Zhu and T. Hastie. Kernel logistic regression and the import vector machine. In NIPS 14, pages 10811088, 2001. [30] X. Zhu. Semi-supervised learning literature survey. Technical report, Computer Sciences TR 1530, University of Wisconsin - Madison, 2005. [31] X. Zhu, Z. Ghahramani, and J. Lafferty. Semi-supervised learning using gaussian fields and harmonic functions. In Proc. ICML'2003, 2003. [32] X. Zhu, J. Kandola, Z. Ghahramani, and J. Lafferty. Nonparametric transforms of graph kernels for semi-supervised learning. In NIPS2005, 2005. 196 Research Track Paper

Active Learning;data mining;classification;unified kernel machine(UKM);Kernel Machines;spectral kernel learning (SKL);Kernel Logistic Regressions;Supervised Learning;supervised learning;Semi-Supervised Learning;active learning;Classification;Unsuper-vised Kernel Design;framework;Spectral Kernel Learning;semi-supervised kernel learning

127

Leo: A System for Cost Effective 3D Shaded Graphics

A physically compact, low cost, high performance 3D graphics accelerator is presented. It supports shaded rendering of triangles and antialiased lines into a double-buffered 24-bit true color frame buffer with a 24-bit Z-buffer. Nearly the only chips used besides standard memory parts are 11 ASICs (of four types). Special geometry data reformatting hardware on one ASIC greatly speeds and simplifies the data input pipeline. Floating-point performance is enhanced by another ASIC: a custom graphics microprocessor, with specialized graphics instructions and features. Screen primitive rasterization is carried out in parallel by five drawing ASICs, employing a new partitioning of the back-end rendering task. For typical rendering cases, the only system performance bottleneck is that intrinsically imposed by VRAM.

INTRODUCTION To expand the role of 3D graphics in the mainstream computer industry , cost effective, physically small, usable performance 3D shaded graphics architectures must be developed. For such systems, new features and sheer performance at any price can no longer be the driving force behind the architecture; instead, the focus must be on affordable desktop systems. The historical approach to achieving low cost in 3D graphics systems has been to compromise both performance and image quality. But now, falling memory component prices are bringing nearly ideal frame buffers into the price range of the volume market: double buffered 24-bit color with a 24-bit Z-buffer. The challenge is to drive these memory chips at their maximum rate with a minimum of supporting rendering chips, keeping the total system cost and physical size to an absolute minimum. To achieve this, graphics architectures must be repartitioned to reduce chip count and internal bus sizes, while still supporting existing 2D and 3D functionality. This paper describes a new 3D graphics system, Leo, designed to these philosophies. For typical cases, Leo's only performance limit is that intrinsically imposed by VRAM. This was achieved by a combination of new architectural techniques and advances in VLSI technology. The result is a system without performance or image quality compromises, at an affordable cost and small physical size. The Leo board set is about the size of one and a half paperback novels ; the complete workstation is slightly larger than two copies of Foley and Van Dam [7]. Leo supports both the traditional requirements of the 2D X window system and the needs of 3D rendering: shaded triangles, antialiased vectors, etc. ARCHITECTURAL ALTERNATIVES A generic pipeline for 3D shaded graphics is shown in Figure 1. ([7] Chapter 18 is a good overview of 3D graphics hardware pipeline issues .) This pipeline is truly generic, as at the top level nearly every commercial 3D graphics accelerator fits this abstraction. Where individual systems differ is in the partitioning of this rendering pipeline , especially in how they employ parallelism. Two major areas have been subject to separate optimization: the floating-point intensive initial stages of processing up to, and many times including, primitive set-up; and the drawing-intensive operation of generating pixels within a primitive and Z-buffering them into the frame buffer. For low end accelerators, only portions of the pixel drawing stages of the pipeline are in hardware; the floating-point intensive parts of the pipe are processed by the host in software. As general purpose processors increase in floating-point power, such systems are starting to support interesting rendering rates, while minimizing cost [8]. But, beyond some limit, support of higher performance requires dedicated hardware for the entire pipeline. There are several choices available for partitioning the floating-point intensive stages. Historically, older systems performed these tasks in a serial fashion [2]. In time though, breaking the pipe into more pieces for more parallelism (and thus performance) meant that each section was devoting more and more of its time to I/O overhead rather than to real work. Also, computational variance meant that many portions of the pipe would commonly be idle while others were overloaded. This led to the data parallel designs of most recent 3D graphics architectures [12]. Leo: A System for Cost Effective 3D Shaded Graphics Michael F Deering, Scott R Nelson Sun Microsystems Computer Corporation Here the concept is that multiple parallel computation units can each process the entire floating-point intensive task, working in parallel on different parts of the scene to be rendered. This allows each pipe to be given a large task to chew on, minimizing handshake overhead. But now there is a different load balancing problem. If one pipe has an extra large task, the other parallel pipes may go idle waiting for their slowest peer, if the common requirement of in-order execution of tasks is to be maintained. Minor load imbalances can be averaged out by adding FIFO buffers to the inputs and outputs of the parallel pipes. Limiting the maximum size of task given to any one pipe also limits the maximum imbalance, at the expense of further fragmenting the tasks and inducing additional overhead. But the most severe performance bottleneck lies in the pixel drawing back-end. The most fundamental constraint on 3D computer graphics architecture over the last ten years has been the memory chips that comprise the frame buffer. Several research systems have attempted to avoid this bottleneck by various techniques [10][4][8], but all commercial workstation systems use conventional Z-buffer rendering algorithms into standard VRAMs or DRAMs. How this RAM is organized is an important defining feature of any high performance rendering system. LEO OVERVIEW Figure 2 is a diagram of the Leo system. This figure is not just a block diagram; it is also a chip level diagram, as every chip in the system is shown in this diagram. All input data and window system interactions enter through the LeoCommand chip. Geometry data is reformatted in this chip before being distributed to the array of LeoFloat chips below. The LeoFloat chips are microcoded specialized DSP-like processors that tackle the floating-point intensive stages of the rendering pipeline. The LeoDraw chips handle all screen space pixel rendering and are directly connected to the frame buffer RAM chips. LeoCross handles the back-end color look-up tables, double buffering, and video timing, passing the final digital pixel values to the RAMDAC. Leo Command Leo Float Leo Draw Leo Cross RAM DAC Clock Generator Boot PROM SRAM SRAM SRAM SRAM VRAM VRAM VRAM VRAM VRAM VRAM VRAM VRAM VRAM DRAM DRAM Leo Draw VRAM VRAM VRAM VRAM VRAM VRAM VRAM VRAM VRAM DRAM DRAM Leo Draw VRAM VRAM VRAM VRAM VRAM VRAM VRAM VRAM VRAM DRAM DRAM Leo Draw VRAM VRAM VRAM VRAM VRAM VRAM VRAM VRAM VRAM DRAM DRAM Leo Draw VRAM VRAM VRAM VRAM VRAM VRAM VRAM VRAM VRAM DRAM DRAM Leo Float SRAM SRAM SRAM SRAM Leo Float SRAM SRAM SRAM SRAM Leo Float SRAM SRAM SRAM SRAM CF Bus CD Bus CX Bus (Subset of CD Bus) Figure 2: The Leo Block Diagram. Every chip in the system is represented in this diagram. Video Output SBus Data Input Transformation Clip Test Face Determination Lighting Clip (if needed) Perspective Divide Screen Space Conversion Set Up for Incremental Render Edge-Walk Span-Interpolate Z-Buffered Blend VRAM Frame Buffer Double Buffered MUX Output Lookup Table Digital to Analog Conversion Figure 1: Generic 3D Graphics Pipeline Drawing Floating-point Intensive Functions Intensive Functions CD Bus 102 The development of the Leo architecture started with the constraints imposed by contemporary VRAM technology. As will be derived in the LeoDraw section below, these constraints led to the partitioning of the VRAM controlling LeoDraw chips, and set a maximum back-end rendering rate. This rate in turn set the performance goal for LeoFloat, as well as the data input bandwidth and processing rate for LeoCommand. After the initial partitioning of the rendering pipeline into these chips, each chip was subjected to additional optimization. Throughput bottlenecks in input geometry format conversion, floating-point processing, and pixel rendering were identified and overcome by adding reinforcing hardware to the appropriate chips. Leo's floating-point intensive section uses data parallel partitioning . LeoCommand helps minimize load balancing problems by breaking down rendering tasks to the smallest isolated primitives: individual triangles, vectors, dots, portions of pixel rasters, rendering attributes, etc., at the cost of precluding optimizations for shared data in triangle strips and polylines. This was considered acceptable due to the very low average strip length empirically observed in real applications. The overhead of splitting geometric data into isolated primitives is minimized by the use of dedicated hardware for this task. Another benefit of converting all rendering operations to isolated primitives is that down-stream processing of primitives is considerably simplified by only needing to focus on the isolated case. INPUT PROCESSING LEO COMMAND Feeding the pipe Leo supports input of geometry data both as programmed I/O and through DMA. The host CPU can directly store up to 32 data words in an internal LeoCommand buffer without expensive read back testing of input status every few words. This is useful on hosts that do not support DMA, or when the host must perform format conversions beyond those supported in hardware. In DMA mode, LeoCommand employs efficient block transfer protocols on the system bus to transfer data from system memory to its input buffer, allowing much higher bandwidth than simple programmed I/O. Virtual memory pointers to application's geometry arrays are passed directly to LeoCommand, which converts them to physical memory addresses without operating system intervention (except when a page is marked as currently non-resident). This frees the host CPU to perform other computations during the data transfer. Thus the DMA can be efficient even for pure immediate-mode applications, where the geometry is being created on the fly. Problem: Tower of Babel of input formats One of the problems modern display systems face is the explosion of different input formats for similar drawing functions that need to be supported. Providing optimized microcode for each format rapidly becomes unwieldy. The host CPU could be used to pretrans-late the primitive formats, but at high speeds this conversion operation can itself become a system bottleneck. Because DMA completely bypasses the host CPU, LeoCommand includes a programmable format conversion unit in the geometry data pipeline. This reformatter is considerably less complex than a general purpose CPU, but can handle the most commonly used input formats, and at very high speeds. The geometry reformatting subsystem allows several orthogonal operations to be applied to input data. This geometric input data is abstracted as a stream of vertex packets. Each vertex packet may contain any combination of vertex position, vertex normal, vertex color, facet normal, facet color, texture map coordinates, pick IDs, headers, and other information. One conversion supports arbitrary re-ordering of data within a vertex, allowing a standardized element order after reformatting. Another operation supports the conversion of multiple numeric formats to 32-bit IEEE floating-point. The source data can be 8-bit or 16-bit fixed-point, or 32-bit or 64-bit IEEE floating-point. Additional miscellaneous reformatting allows the stripping of headers and other fields, the addition of an internally generated sequential pick ID, and insertion of constants. The final reformatting stage re-packages vertex packets into complete isolated geometry primitives (points, lines, triangles). Chaining bits in vertex headers delineate which vertices form primitives. Like some other systems, Leo supports a generalized form of triangle strip (see Figure 3), where vertex header bits within a strip specify how the incoming vertex should be combined with previous vertices to form the next triangle. A stack of the last three vertices used to form a triangle is kept. The three vertices are labeled oldest, middle , and newest. An incoming vertex of type replace _oldest causes the oldest vertex to be replaced by the middle, the middle to be replaced by the newest, and the incoming vertex becomes the newest. This corresponds to a PHIGS PLUS triangle strip (sometimes called a "zig-zag" strip). The replacement type replace _middle leaves the oldest vertex unchanged, replaces the middle vertex by the newest, and the incoming vertex becomes the newest. This corresponds to a triangle star. The replacement type restart marks the oldest and middle vertices as invalid, and the incoming vertex becomes the newest. Generalized triangle strips must always start with this code. A triangle will be output only when a replacement operation results in three valid vertices. Restart corresponds to a "move" operation in polylines, and allows multiple unconnected variable-length triangle strips to be described by a single data structure passed in by the user, 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 22 23 24 25 26 27 28 29 30 31 32 33 21 1 Restart 2 RO 3 RO 4 RO 5 RO 6 RO 7 Restart 8 RO 9 RO 10 RM 11 RM 12 RM 13 RM 14 RM 15 Restart 16 RO 17 RO 18 Restart 19 RO 20 RO 21 RO 22 Restart 23 RO 24 RO 25 RO 26 RO 27 RO 28 RO 29 RM 30 RM 31 RM 32 RM 33 RO Triangle Strip Triangle Star Independent Triangle Independent Quad Mixed Strip Figure 3: A Generalized Triangle Strip Vertex Codes RO = Replace Oldest RM = Replace Middle 103 reducing the overhead. The generalized triangle strip's ability to effectively change from "strip" to "star" mode in the middle of a strip allows more complex geometry to be represented compactly, and requires less input data bandwidth. The restart capability allows several pieces of disconnected geometry to be passed in one DMA operation . Figure 3 shows a single generalized triangle strip, and the associated replacement codes. LeoCommand also supports header-less strips of triangle vertices either as pure strips, pure stars, or pure independent triangles. LeoCommand hardware automatically converts generalized triangle strips into isolated triangles. Triangles are normalized such that the front face is always defined by a clockwise vertex order after transformation. To support this, a header bit in each restart defines the initial face order of each sub-strip, and the vertex order is reversed after every replace _oldest. LeoCommand passes each com-pleted triangle to the next available LeoFloat chip, as indicated by the input FIFO status that each LeoFloat sends back to LeoCommand . The order in which triangles have been sent to each LeoFloat is scoreboarded by LeoCommand, so that processed triangles are let out of the LeoFloat array in the same order as they entered . Non-sequential rendering order is also supported, but the automatic rendering task distribution hardware works so well that the performance difference is less than 3%. A similar, but less complex vertex repackaging is supported for polylines and multi-polylines via a move/draw bit in the vertex packet header. To save IC pins and PC board complexity, the internal Leo data busses connecting LeoCommand, LeoFloat, and LeoDraw are 16 bits in size. When colors, normals, and texture map coefficients are being transmitted on the CF-bus between LeoCommand and the LeoFloats , these components are (optionally) compressed from 32-bit IEEE floating-point into 16-bit fixed point fractions by LeoCommand , and then automatically reconverted back to 32-bit IEEE floating-point values by LeoFloat. This quantization does not effect quality. Color components will eventually end up as 8-bit values in the frame buffer. For normals, 16-bit (signed) accuracy represents a resolution of approximately plus or minus an inch at one mile. This optimization reduces the required data transfer bandwidth by 25%. FLOATING-POINT PROCESSING LEO FLOAT After canonical format conversion, the next stages of processing triangles in a display pipeline are: transformation, clip test, face determination , lighting, clipping (if required), screen space conversion, and set-up. These operations are complex enough to require the use of a general purpose processor. Use of commercially available DSP (Digital Signal Processing) chips for this work has two major drawbacks. First, most such processors require a considerable number of surrounding glue chips, especially when they are deployed as multi-processors. These glue chips can easily quadruple the board area dedicated to the DSP chip, as well as adversely affecting power, heat, cost, and reliability. Second, few of these chips have been optimized for 3D graphics. A better solution might be to augment the DSP with a special ASIC that would replace all of these glue chips. Given the expense of developing an ASIC, we decided to merge that ASIC with a custom DSP core optimized for graphics. The resulting chip was LeoFloat. LeoFloat combines a 32-bit mi-crocodable floating-point core with concurrent input and output packet communication subsystems (see Figure 4 . ), similar to the approach of [3]. The only support chips required are four SRAM chips for external microcode store. A number of specialized graphics instructions and features make LeoFloat different from existing DSP processors. Each individual feature only makes a modest incremen-tal contribution to performance, and indeed many have appeared in other designs. What is novel about LeoFloat is the combination of features, whose cumulative effect leads to impressive overall system performance. The following sections describe some of the more important special graphics instructions and features. Double buffered asynchronous I/O register files. All input and output commands are packaged up by separate I/O packet hardware. Variable length packets of up to 32 32-bit words are automatically written into (or out of) on-chip double-buffered register files (the I and O registers). These are mapped directly into microcode register space. Special instructions allow complete packets to be requested, relinquished, or queued for transmission in one instruction cycle. Enough internal registers. Most commercial DSP chips support a very small number of internal fast registers, certainly much smaller than the data needed by the inner loops of most 3D pipeline algorithms . They attempt to make up for this with on-chip SRAM or data caches, but typically SRAMs are not multi-ported and the caches not user-schedulable. We cheated with LeoFloat. We first wrote the code for the largest important inner loop (triangles), counted how many registers were needed (288), and built that many into the chip. Parallel internal function units . The floating-point core functions (32-bit IEEE format) include multiply, ALU, reciprocal, and integer operations, all of which can often be executed in parallel. It is particularly important that the floating-point reciprocal operation not tie up the multiply and add units, so that perspective or slope calculations can proceed in parallel with the rest of geometric processing . Less frequently used reciprocal square root hardware is shared with the integer function unit. Put all non-critical algorithms on the host. We avoided the necessity of building a high level language compiler (and support instructions ) for LeoFloat by moving any code not worth hand coding in microcode to the host processor. The result is a small, clean kernel of graphics routines in microcode. (A fairly powerful macro-assembler with a `C'-like syntax was built to support the hand coding.) Software pipeline scheduling. One of the most complex parts of modern CPUs to design and debug is their scoreboard section, which schedules the execution of instructions across multiple steps in time and function units, presenting the programmer with the Figure 4: LeoFloat arithmetic function units, registers and data paths. I0 -I31 I0'-I31' * * + + + IALU 1/X FALU FMULT Input from off-chip Off-chip output P0 -P31 P32-P63 P64-P91 R0 -R31 R32-R63 O0 -O31 O0'-O31' 104 illusion that individual instructions are executed in one shot. LeoFloat avoided all this hardware by using more direct control fields, like horizontal microprogrammable machines, and leaving it to the assembler (and occasionally the programmer) to skew one logical instruction across several physical instructions. Special clip condition codes & clip branch. For clip testing we employ a modified Sutherland-Hodgman algorithm, which first computes a vector of clip condition bits. LeoFloat has a clip test instruction that computes these bits two at a time, shifting them into a special clip-bits register. After the bits have been computed, special branch instructions decode these bits into the appropriate case: clip rejected, clip accepted, single edge clip (six cases), or needs general clipping. There are separate branch instructions for triangles and vectors. (A similar approach was taken in [9].) The branch instructions allow multiple other conditions to be checked at the same time, including backfacing and model clipping. Register Y sort instruction. The first step of the algorithm we used for setting up triangles for scan conversion sorts the three triangle vertices in ascending Y order. On a conventional processor this requires either moving a lot of data, always referring to vertex data through indirect pointers, or replicating the set-up code for all six possible permutations of triangle vertex order. LeoFloat has a special instruction that takes the results of the last three comparisons and reorders part of the R register file to place vertices in sorted order. Miscellaneous. LeoFloat contains many performance features tra-ditionally found on DSP chips, including an internal subroutine stack, block load/store SRAM, and integer functions. Also there is a "kitchen sink" instruction that initiates multiple housekeeping functions in one instruction, such as "transmit current output packet (if not clip pending), request new input packet, extract op-code and dispatch to next task." Code results: equivalent to 150 megaflop DSP. Each 25 MHz LeoFloat processes the benchmark isolated triangle (including clip-test and set-up) in 379 clocks. (With a few exceptions, microcode instructions issue at a rate of one per clock tick.) The same graphics algorithm was tightly coded on several RISC processors and DSP chips (SPARC, i860, C30, etc.), and typically took on the order of 1100 clocks. Thus the 379 LeoFloat instruction at 25 MHz do the equivalent work of a traditional DSP chip running at 75 MHz (even though there are only 54 megaflops of hardware). Of course these numbers only hold for triangles and vectors, but that's most of what LeoFloat does. Four LeoFloats assure that floating-point processing is not the bottleneck for 100-pixel isolated, lighted triangles. SCREEN SPACE RENDERING: LEO DRAW VRAM limits Commercial VRAM chips represent a fundamental constraint on the possible pixel rendering performance of Leo's class of graphics accelerator. The goal of the Leo architecture was to ensure to the greatest extent possible that this was the only performance limit for typical rendering operations. The fundamental memory transaction for Z-buffered rendering algorithms is a conditional read-modify-write cycle. Given an XY address and a computed RGBZ value, the old Z value at the XY address is first read, and then if the computed Z is in front of the old Z, the computed RGBZ value is written into the memory. Such transactions can be mapped to allowable VRAM control signals in many different ways: reads and writes may be batched, Z may be read out through the video port, etc. VRAM chips constrain system rendering performance in two ways. First, they impose a minimum cycle time per RAM bank for the Z-buffered read-modify-write cycle. Figure 5 is a plot of this cycle time (when in "page" mode) and its changes over a half-decade period. VRAMs also constrain the ways in which a frame buffer can be partitioned into independently addressable banks. Throughout the five year period in Figure 5, three generations of VRAM technology have been organized as 256K by 4, 8, and 16-bit memories. For contemporary display resolutions of 1280 1024, the chips comprising a minimum frame buffer can be organized into no more than five separately-addressed interleave banks. Combining this information , a theoretical maximum rendering speed for a primitive can be computed. The second line in Figure 5 is the corresponding performance for rendering 100-pixel Z-buffered triangles, including the overhead for entering page mode, content refresh, and video shift register transfers (video refresh). Higher rendering rates are only possible if additional redundant memory chips are added, allowing for higher interleaving factors, at the price of increased system cost. Even supporting five parallel interleaves has a cost: at least 305 memory interface pins (five banks of (24 RGB + 24 Z + 13 address/ control)) are required, more pins than it is currently possible to dedicate to a memory interface on one chip. Some systems have used external buffer chips, but on a minimum cost and board area system , this costs almost as much as additional custom chips. Thus, on the Leo system we opted for five separate VRAM control chips (LeoDraws). Triangle scan conversion Traditional shaded triangle scan conversion has typically been via a linear pipeline of edge-walking followed by scan interpolation [12]. There have been several approaches to achieving higher throughput in rasterization. [2] employed a single edge-walker, but parallel scan interpolation. [4][10] employed massively parallel rasterizers. [6] and other recent machines use moderately parallel rasterizers, with additional logic to merge the pixel rasterization streams back together. In the Leo design we chose to broadcast the identical triangle specification to five parallel rendering chips, each tasked with rendering only those pixels visible in the local interleave. Each chip performs its own complete edge-walk and span interpolation of the triangle, biased by the chip's local interleave. By paying careful attention to proper mathematical sampling theory for rasterized pixels, the five 90 91 92 93 94 180 ns 160 ns 140 ns 260K 220K 200K 240K 200 ns 100 pixel triangle theoretical maximum render rate VRAM minimum Z-buffer RGB read/modify/write cycle time (on page) (off page = 1.5x) Figure 5: VRAM cycle time and theoretical maximum triangle rendering rate (for five-way interleaved frame buffers). VRAM Cycle T ime T riangle Rendering Rate 1 Meg VRAM 2 Meg VRAM 4 Meg 105 chips can act in concert to produce the correct combined rasterized image. Mathematically, each chip thinks it is rasterizing the triangle into an image memory with valid pixel centers only every five original pixels horizontally, with each chip starting off biased one more pixel to the right. To obtain the speed benefits of parallel chips, most high performance graphics systems have split the edge-walk and span-interpolate functions into separate chips. But an examination of the relative amounts of data flow between rendering pipeline stages shows that the overall peak data transfer bandwidth demand occurs between the edge-walk and span-interpolate sections, induced by long thin triangles, which commonly occur in tessellated geometry. To minimize pin counts and PC board bus complexity, Leo decided to replicate the edge-walking function into each of the five span-interpolation chips. One potential drawback of this approach is that the edge-walking section of each LeoDraw chip will have to advance to the next scan line up to five times more often than a single rasterization chip would. Thus LeoDraw's edge-walking circuit was designed to operate in one single pixel cycle time (160 ns. read-modify-write VRAM cycle), so it would never hold back scan conversion. Other usual pipelining techniques were used, such as loading in and buffering the next triangle to be drawn in parallel with rasterizing the current triangle. Window clipping, blending, and other pixel post processing are handled in later pipelined stages. Line scan conversion As with triangles, the mathematics of the line rasterization algorithms were set up to allow distributed rendering of aliased and antialiased lines and dots, with each LeoDraw chip handling the 1/5 of the frame buffer pixels that it owns. While the Leo system uses the X11 semantics of Bresenham lines for window system operations, these produce unacceptable motion artifacts in 3D wireframe rendering. Therefore, when rendering 3D lines, Leo employs a high-accuracy DDA algorithm, using 32 bits internally for sufficient subpixel precision. At present there is no agreement in the industry on the definition of a high quality antialiased line. We choose to use the image quality of vector strokers of years ago as our quality standard, and we tested different algorithms with end users, many of whom were still using cal-ligraphic displays. We found users desired algorithms that displayed no roping, angle sensitivities, short vector artifacts, or end-point artifacts . We submitted the resulting antialiased line quality test patterns as a GPC [11] test image. In achieving the desired image quality level , we determined several properties that a successful line antialias-ing algorithm must have. First, the lines must have at least three pixels of width across the minor axis. Two-pixel wide antialiased lines exhibit serious roping artifacts. Four-pixel wide lines offer no visible improvement except for lines near 45 degrees. Second, proper endpoint ramps spread over at least two pixels are necessary both for seamless line segment joins as well as for isolated line-ends. Third, proper care must be taken when sampling lines of subpixel length to maintain proper final intensity. Fourth, intensity or filter adjustments based on the slope are necessary to avoid artifacts when rotating wireframe images. To implement all this, we found that we needed at least four bits of subpixel positional accuracy after cumulative interpolation error is factored in. That is why we used 32 bits for XY coordinate accuracy: 12 for pixel location, 4 for subpixel location, and 16 for DDA interpolation error. (The actual error limit is imposed by the original, user-supplied 32-bit IEEE floating-point data.) Because of the horizontal interleaving and preferred scan direction, the X-major and Y-major aliased and antialiased line rasterization algorithms are not symmetric, so separate optimized algorithms were employed for each. Antialiased dots Empirical testing showed that only three bits of subpixel precision are necessary for accurate rendering of antialiased dots. For ASIC implementation, this was most easily accomplished using a brute-force table lookup of one of 64 precomputed 3 3 pixel dot images. These images are stored in on-chip ROM, and were generated using a circular symmetric Gaussian filter. Triangle, line, and dot hardware Implementation of the triangle and antialiased vector rasterization algorithms require substantial hardware resources. Triangles need single pixel cycle edge-walking hardware in parallel with RGBZ span interpolation hardware. To obtain the desired quality of antialiased vectors, our algorithms require hardware to apply multiple waveform shaping functions to every generated pixel. As a result, the total VLSI area needed for antialiased vectors is nearly as large as for triangles. To keep the chip die size reasonable, we reformu-lated both the triangle and antialiased vector algorithms to combine and reuse the same function units. The only difference is how the separate sequencers set up the rasterization pipeline. Per-pixel depth cue Depth cueing has long been a heavily-used staple of wireframe applications , but in most modern rendering systems it is an extra time expense feature, performed on endpoints back in the floating-point section. We felt that we were architecting Leo not for benchmarks, but for users, and many wireframe users want to have depth cueing on all the time. Therefore, we built a parallel hardware depth cue function unit into each LeoDraw. Each triangle, vector, or dot rendered by Leo can be optionally depth cued at absolutely no cost in performance. Another benefit of per-pixel depth cueing is full compliance with the PHIGS PLUS depth cueing specification. For Leo, per-pixel depth cueing hardware also simplifies the LeoFloat microcode , by freeing the LeoFloats from ever having to deal with it. Picking support Interactive graphics requires not only the rapid display of geometric data, but also interaction with that data: the ability to pick a particular part or primitive within a part. Any pixels drawn within the bounds of a 3D pick aperture result in a pick hit, causing the current pick IDs to be automatically DMAed back to host memory. Window system support Many otherwise sophisticated 3D display systems become somewhat befuddled when having to deal simultaneously with 3D rendering applications and a 2D window system. Modern window systems on interactive workstations require frequent context switching of the rendering pipeline state. Some 3D architectures have tried to minimize the overhead associated with context switching by supporting multiple 3D contexts in hardware. Leo goes one step further , maintaining two completely separate pipelines in hardware: one for traditional 2D window operations; the other for full 3D rendering . Because the majority of context switch requests are for 2D window system operations, the need for more complex 3D pipeline context switching is significantly reduced. The 2D context is much lighter weight and correspondingly easier to context switch. The two separate graphics pipelines operate completely in parallel, allowing simultaneous access by two independent CPUs on a multi-processor host. 2D functionality abstracts the frame buffer as a 1-bit, 8-bit, or 24-bit pixel array. Operations include random pixel access, optimized character cell writes, block clear, block copy, and the usual menagerie of 106 boolean operations, write masks, etc. Vertical block moves are special cased, as they are typically used in vertical scrolling of text windows, and can be processed faster than the general block move because the pixel data does not have to move across LeoDraw chip interleaves. Rendering into non-rectangular shaped windows is supported by special clip hardware, resulting in no loss in performance . A special block clear function allows designated windows (and their Z-buffers) to be initialized to any given constant in under 200 microseconds. Without this last feature, 30 Hz or faster animation of non-trivial objects would have been impossible. 7 VIDEO OUTPUT: LEO CROSS Leo's standard video output format is 1280 1024 at 76 Hz refresh rate, but it also supports other resolutions, including 1152 900, interlaced 640 480 RS-170 (NTSC), interlaced 768 576 PAL timing, and 960 680 113 Hz field sequential stereo. LeoCross contains several color look-up tables, supporting multiple pseudo color maps without color map flashing. The look-up table also supports two different true color abstractions: 24-bit linear color (needed by rendering applications), and REC-709 non-linear color (required by many imaging applications). Virtual reality support Stereo output is becoming increasingly important for use in Virtual Reality applications. Leo's design goals included support for the Virtual Holographic Workstation system configuration described in [5]. Leo's stereo resolution was chosen to support square pixels, so that lines and antialiased lines are displayed properly in stereo, and standard window system applications can co-exist with stereo. Stereo can be enabled on a per-window basis (when in stereo mode windows are effectively quad-buffered). Hooks were included in LeoCross to support display technologies other than CRT's, that may be needed for head-mounted virtual reality displays. 8 NURBS AND TEXTURE MAP SUPPORT/b> One of the advantages to using programmable elements within a graphics accelerator is that additional complex functionality, such as NURBS and texture mapping, can be accelerated. Texture mapping is supported through special LeoFloat microcode and features of LeoCommand. LeoFloat microcode also includes algorithms to accelerate dynamic tessellation of trimmed NURBS surfaces. The dynamic tessellation technique involves reducing trimmed NURBS surfaces into properly sized triangles according to a display/pixel space approximation criteria [1]; i.e. the fineness of tessellation is view dependent. In the past, dynamic tessellation tended to be mainly useful as a compression technique, to avoid storing all the flattened triangles from a NURBS surface in memory. Dynamic tessellation was not viewed as a performance enhancer, for while it might generate only a third as many triangles as a static tessellation, the triangles were generated at least an order of magnitude or more slower than brute force triangle rendering. In addition it had other problems, such as not handling general trimming. For many cases, Leo's dynamic tesselator can generate and render triangles only a small integer multiple slower than prestored triangle rendering, which for some views, can result in faster overall object rendering. 9 RESULTS Leo is physically a-two board sandwich, measuring 5.7 6.7 0.6 inches, that fits in a standard 2S SBus slot. Figure 6 is a photo of the two boards, separated, showing all the custom ASICs. Figure 7 is a photo of the complete Leo workstation, next to two of our units of scale and the board set. Leo can render 210K 100-pixel isolated, lighted, Gouraud shaded, Z-buffered, depth cued triangles per second, with one infinite diffuse and one ambient light source enabled. At 100 pixels, Leo is still VRAM rendering speed limited; smaller triangles render faster. Isolated 10-pixel antialiased, constant color, Z-buffered, depth cued lines (which are actually 12 pixels long due to endpoint ramps, and three pixels wide) render at a 422K per second rate. Corresponding aliased lines render at 730K. Aliased and antialiased constant color, Z-buffered, depth cued dots are clocked at 1100K. 24-bit image rasters can be loaded onto the screen at a 10M pixel per second rate. Screen scrolls, block moves, and raster character draws all also have competitive performance. Figure 8 is a sample of shaded triangle rendering. 10 SIMULATION A system as complex as Leo cannot be debugged after the fact. All the new rendering mathematics were extensively simulated before being committed to hardware design. As each chip was defined, high, medium, and low level simulators of its function were written and continuously used to verify functionality and performance. Complete images of simulated rendering were generated throughout the course of the project, from within weeks of its start. As a result, the window system and complex 3D rendering were up and running on a complete board set within a week of receiving the first set of chips. 11 CONCLUSIONS By paying careful attention to the forces that drive both performance and cost, a physically compact complete 3D shaded graphics accelerator was created. The focus was not on new rendering features , but on cost reduction and performance enhancement of the most useful core of 3D graphics primitives. New parallel algorithms were developed to allow accurate screen space rendering of primitives. Judicious use of hardware to perform some key traditional software functions (such as format conversion and primitive vertex reassembly) greatly simplified the microcode task. A specialized floating-point core optimized for the primary task of processing lines and triangles also supports more general graphics processing , such as rasters and NURBS. The final system performance is limited by the only chips not custom designed for Leo: the standard RAM chips. ACKNOWLEDGEMENTS The authors would like to thank the entire Leo team for their efforts in producing the system, and Mike Lavelle for help with the paper. REFERENCES 1. Abi-Ezzi, Salim, and L. Shirman. Tessellation of Curved Surfaces under Highly Varying Transformations. Proc. Euro-graphics '91 (Vienna, Austria, September 1991), 385-397. 2. Akeley, Kurt and T. Jermoluk. High-Performance Polygon Rendering, Proceedings of SIGGRAPH '88 (Atlanta, GA, Aug 1-5, 1988). In Computer Graphics 22, 4 (July 1988), 239-246. 3. Anido, M., D. Allerton and E. Zaluska. MIGS - A Multiprocessor Image Generation System using RISC-like Micropro-cessors . Proceedings of CGI '89 (Leeds, UK, June 1989), Springer Verlag 1990. 4. Deering, Michael, S. Winner, B. Schediwy, C. Duffy and N. Hunt. The Triangle Processor and Normal Vector Shader: A VLSI system for High Performance Graphics. Proceedings of SIGGRAPH '88 (Atlanta, GA, Aug 1-5, 1988). In Computer Graphics 22, 4 (July 1988), 21-30. 107 5. Deering, Michael. High Resolution Virtual Reality. Proceedings of SIGGRAPH '92 (Chicago, IL, July 26-31, 1992). In Computer Graphics 26, 2 (July 1992), 195-202. 6. Dunnett, Graham, M. White, P. Lister and R. Grimsdale. The Image Chip for High Performance 3D Rendering. IEEE Computer Graphics and Applications 12, 6 (November 1992), 41-52. 7. Foley, James, A. van Dam, S. Feiner and J Hughes. Computer Graphics: Principles and Practice, 2nd ed., Addison-Wesley , 1990. 8. Kelley, Michael, S. Winner, K. Gould. A Scalable Hardware Render Accelerator using a Modified Scanline Algorithm. Proceedings of SIGGRAPH '92 (Chicago, IL, July 26-31, 1992). In Computer Graphics 26, 2 (July 1992), 241-248. 9. Kirk, David, and D. Voorhies. The Rendering Architecture of the DN10000VS. Proceedings of SIGGRAPH '90 (Dallas, TX, August 6-10, 1990). In Computer Graphics 24, 4 (August 1990), 299-307. 10. Molnar, Steven, J. Eyles, J. Poulton. PixelFlow: High-Speed Rendering Using Image Composition. Proceedings of SIGGRAPH '92 (Chicago, IL, July 26-31, 1992). In Computer Graphics 26, 2 (July 1992), 231-240. 11. Nelson, Scott. GPC Line Quality Benchmark Test. GPC Test Suite, NCGA GPC committee 1991. 12. Torborg, John. A Parallel Processor Architecture for Graphics Arithmetic Operations. Proceedings of SIGGRAPH '87 (Anaheim, CA, July 27-31, 1987). In Computer Graphics 21, 4 (July 1987), 197-204. Figure 8: Traffic Jam to Point Reyes. A scene containing 2,322,000 triangles, rendered by Leo Hardware. Sto-chastically super-sampled 8 times. Models courtesy of Viewpoint Animation Engineering. Figure 7: The complete SPARCstation ZX workstation, next to two of our units of scale and the Leo board set. Figure 6: The two boards, unfolded. 108

input processing;3D graphics hardware;parallel algorithms;video output;general graphics processing;parallel graphics algorithms;small physical size;geometry data;3D shaded graphics;rendering;screen space rendering;antialiased lines;floating-point microprocessors;low cost;floating point processing;gouraud shading

128

Location based Indexing Scheme for DAYS

Data dissemination through wireless channels for broadcasting information to consumers is becoming quite common. Many dissemination schemes have been proposed but most of them push data to wireless channels for general consumption. Push based broadcast [1] is essentially asymmetric, i.e., the volume of data being higher from the server to the users than from the users back to the server. Push based scheme requires some indexing which indicates when the data will be broadcast and its position in the broadcast. Access latency and tuning time are the two main parameters which may be used to evaluate an indexing scheme. Two of the important indexing schemes proposed earlier were tree based and the exponential indexing schemes. None of these schemes were able to address the requirements of location dependent data (LDD) which is highly desirable feature of data dissemination. In this paper, we discuss the broadcast of LDD in our project DAta in Your Space (DAYS), and propose a scheme for indexing LDD. We argue that this scheme, when applied to LDD, significantly improves performance in terms of tuning time over the above mentioned schemes. We prove our argument with the help of simulation results.

INTRODUCTION Wireless data dissemination is an economical and efficient way to make desired data available to a large number of mobile or static users. The mode of data transfer is essentially asymmetric, that is, the capacity of the transfer of data (downstream communication) from the server to the client (mobile user) is significantly larger than the client or mobile user to the server (upstream communication). The effectiveness of a data dissemination system is judged by its ability to provide user the required data at anywhere and at anytime. One of the best ways to accomplish this is through the dissemination of highly personalized Location Based Services (LBS) which allows users to access personalized location dependent data. An example would be someone using their mobile device to search for a vegetarian restaurant. The LBS application would interact with other location technology components or use the mobile user's input to determine the user's location and download the information about the restaurants in proximity to the user by tuning into the wireless channel which is disseminating LDD. We see a limited deployment of LBS by some service providers. But there are every indications that with time some of the complex technical problems such as uniform location framework, calculating and tracking locations in all types of places, positioning in various environments, innovative location applications, etc., will be resolved and LBS will become a common facility and will help to improve market productivity and customer comfort. In our project called DAYS, we use wireless data broadcast mechanism to push LDD to users and mobile users monitor and tune the channel to find and download the required data. A simple broadcast, however, is likely to cause significant performance degradation in the energy constrained mobile devices and a common solution to this problem is the use of efficient air indexing. The indexing approach stores control information which tells the user about the data location in the broadcast and how and when he could access it. A mobile user, thus, has some free time to go into the doze mode which conserves valuable power. It also allows the user to personalize his own mobile device by selectively tuning to the information of his choice. Access efficiency and energy conservation are the two issues which are significant for data broadcast systems. Access efficiency refers to the latency experienced when a request is initiated till the response is received. Energy conservation [7, 10] refers to the efficient use of the limited energy of the mobile device in accessing broadcast data. Two parameters that affect these are the tuning time and the access latency. Tuning time refers to the time during which the mobile unit (MU) remains in active state to tune the channel and download its required data. It can also be defined as the number of buckets tuned by the mobile device in active state to get its required data. Access latency may be defined as the time elapsed since a request has been issued till the response has been received. 1 This research was supported by a grant from NSF IIS-0209170. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MobiDE'05, June 12, 2005, Baltimore, Maryland, USA. Copyright 2005 ACM 1-59593-088-4/05/0006...$5.00. 17 Several indexing schemes have been proposed in the past and the prominent among them are the tree based and the exponential indexing schemes [17]. The main disadvantages of the tree based schemes are that they are based on centralized tree structures. To start a search, the MU has to wait until it reaches the root of the next broadcast tree. This significantly affects the tuning time of the mobile unit. The exponential schemes facilitate index replication by sharing links in different search trees. For broadcasts with large number of pages, the exponential scheme has been shown to perform similarly as the tree based schemes in terms of access latency. Also, the average length of broadcast increases due to the index replication and this may cause significant increase in the access latency. None of the above indexing schemes is equally effective in broadcasting location dependent data. In addition to providing low latency, they lack properties which are used to address LDD issues. We propose an indexing scheme in DAYS which takes care of some these problems. We show with simulation results that our scheme outperforms some of the earlier indexing schemes for broadcasting LDD in terms of tuning time. The rest of the paper is presented as follows. In section 2, we discuss previous work related to indexing of broadcast data. Section 3 describes our DAYS architecture. Location dependent data, its generation and subsequent broadcast is presented in section 4. Section 5 discusses our indexing scheme in detail. Simulation of our scheme and its performance evaluation is presented in section 6. Section 7 concludes the paper and mentions future related work. PREVIOUS WORK Several disk-based indexing techniques have been used for air indexing. Imielinski et al. [5, 6] applied the B+ index tree, where the leaf nodes store the arrival times of the data items. The distributed indexing method was proposed to efficiently replicate and distribute the index tree in a broadcast. Specifically, the index tree is divided into a replicated part and a non replicated part. Each broadcast consists of the replicated part and the non-replicated part that indexes the data items immediately following it. As such, each node in the non-replicated part appears only once in a broadcast and, hence, reduces the replication cost and access latency while achieving a good tuning time. Chen et al. [2] and Shivakumar et al. [8] considered unbalanced tree structures to optimize energy consumption for non-uniform data access. These structures minimize the average index search cost by reducing the number of index searches for hot data at the expense of spending more on cold data. Tan and Yu discussed data and index organization under skewed broadcast Hashing and signature methods have also been suggested for wireless broadcast that supports equality queries [9]. A flexible indexing method was proposed in [5]. The flexible index first sorts the data items in ascending (or descending) order of the search key values and then divides them into p segments. The first bucket in each data segment contains a control index, which is a binary index mapping a given key value to the segment containing that key, and a local index, which is an m-entry index mapping a given key value to the buckets within the current segment. By tuning the parameters of p and m, mobile clients can achieve either a good tuning time or good access latency. Another indexing technique proposed is the exponential indexing scheme [17]. In this scheme, a parameterized index, called the exponential index is used to optimize the access latency or the tuning time. It facilitates index replication by linking different search trees. All of the above mentioned schemes have been applied to data which are non related to each other. These non related data may be clustered or non clustered. However, none of them has specifically addressed the requirements of LDD. Location dependent data are data which are associated with a location. Presently there are several applications that deal with LDD [13, 16]. Almost all of them depict LDD with the help of hierarchical structures [3, 4]. This is based on the containment property of location dependent data. The Containment property helps determining relative position of an object by defining or identifying locations that contains those objects. The subordinate locations are hierarchically related to each other. Thus, Containment property limits the range of availability or operation of a service. We use this containment property in our indexing scheme to index LDD. DAYS ARCHITECTURE DAYS has been conceptualized to disseminate topical and non-topical data to users in a local broadcast space and to accept queries from individual users globally. Topical data, for example, weather information, traffic information, stock information, etc., constantly changes over time. Non topical data such as hotel, restaurant, real estate prices, etc., do not change so often. Thus, we envision the presence of two types of data distribution: In the first case, server pushes data to local users through wireless channels. The other case deals with the server sending results of user queries through downlink wireless channels. Technically, we see the presence of two types of queues in the pull based data access. One is a heavily loaded queue containing globally uploaded queries. The other is a comparatively lightly loaded queue consisting of locally uploaded queries. The DAYS architecture [12] as shown in figure 1 consists of a Data Server, Broadcast Scheduler, DAYS Coordinator, Network of LEO satellites for global data delivery and a Local broadcast space. Data is pushed into the local broadcast space so that users may tune into the wireless channels to access the data. The local broadcast space consists of a broadcast tower, mobile units and a network of data staging machines called the surrogates. Data staging in surrogates has been earlier investigated as a successful technique [12, 15] to cache users' related data. We believe that data staging can be used to drastically reduce the latency time for both the local broadcast data as well as global responses. Query request in the surrogates may subsequently be used to generate the popularity patterns which ultimately decide the broadcast schedule [12]. 18 Popularity Feedback from Surrogates for Broadcast Scheduler Local Broadcast Space Broadcast Tower Surrogate MU MU MU MU Data Server Broadcast scheduler DAYS Coordinator Local downlink channel Global downlink channel Pull request queue Global request queue Local request queue Location based index Starbucks Plaza Kansas City Figure 1. DAYS Architecture Figure 2. Location Structure of Starbucks, Plaza LOCATION DEPENDENT DATA (LDD) We argue that incorporating location information in wireless data broadcast can significantly decrease the access latency. This property becomes highly useful for mobile unit which has limited storage and processing capability. There are a variety of applications to obtain information about traffic, restaurant and hotel booking, fast food, gas stations, post office, grocery stores, etc. If these applications are coupled with location information, then the search will be fast and highly cost effective. An important property of the locations is Containment which helps to determine the relative location of an object with respect to its parent that contains the object. Thus, Containment limits the range of availability of a data. We use this property in our indexing scheme. The database contains the broadcast contents which are converted into LDD [14] by associating them with respective locations so that it can be broadcasted in a clustered manner. The clustering of LDD helps the user to locate information efficiently and supports containment property. We present an example to justify our proposition. Example: Suppose a user issues query "Starbucks Coffee in Plaza please." to access information about the Plaza branch of Starbucks Coffee in Kansas City. In the case of location independent set up the system will list all Starbucks coffee shops in Kansas City area. It is obvious that such responses will increase access latency and are not desirable. These can be managed efficiently if the server has location dependent data, i.e., a mapping between a Starbucks coffee shop data and its physical location. Also, for a query including range of locations of Starbucks, a single query requesting locations for the entire region of Kansas City, as shown in Figure 2, will suffice. This will save enormous amount of bandwidth by decreasing the number of messages and at the same time will be helpful in preventing the scalability bottleneck in highly populated area. 4.1 Mapping Function for LDD The example justifies the need for a mapping function to process location dependent queries. This will be especially important for pull based queries across the globe for which the reply could be composed for different parts of the world. The mapping function is necessary to construct the broadcast schedule. We define Global Property Set (GPS) [11], Information Content (IC) set, and Location Hierarchy (LH) set where IC GPS and LH GPS to develop a mapping function. LH = {l 1 , l 2 , l 3 ...,l k } where l i represent locations in the location tree and IC = {ic 1 , ic 2 , ic 3 ,...,ic n } where ic i represent information type. For example, if we have traffic, weather, and stock information are in broadcast then IC = {ic traffic , ic weather , and ic stock }. The mapping scheme must be able to identify and select an IC member and a LH node for (a) correct association, (b) granularity match, (c) and termination condition. For example, weather IC could be associated with a country or a state or a city or a town of LH. The granularity match between the weather and a LH node is as per user requirement. Thus, with a coarse granularity weather information is associated with a country to get country's weather and with town in a finer granularity. If a town is the finest granularity, then it defines the terminal condition for association between IC and LH for weather. This means that a user cannot get weather information about subdivision of a town. In reality weather of a subdivision does not make any sense. We develop a simple heuristic mapping approach scheme based on user requirement. Let IC = {m 1 , m 2 ,m 3 .,..., m k }, where m i represent its element and let LH = {n 1 , n 2 , n 3, ..., n l }, where n i represents LH's member. We define GPS for IC (GPSIC) GPS and for LH (GPSLH) GPS as GPSIC = {P 1 , P 2 ,..., P n }, where P 1 , P 2 , P 3 ,..., P n are properties of its members and GPSLH = {Q 1 , Q 2 ,..., Q m } where Q 1 , Q 2 ,..., Q m are properties of its members. The properties of a particular member of IC are a subset of GPSIC. It is generally true that (property set (m i IC) property set (m j IC)) , however, there may be cases where the intersection is not null. For example, stock IC and movie IC rating do not have any property in common. We assume that any two or more members of IC have at least one common geographical property (i.e. location) because DAYS broadcasts information about those categories, which are closely tied with a location. For example, stock of a company is related to a country, weather is related to a city or state, etc. We define the property subset of m i IC as PSm i m i IC and PSm i = {P 1 , P 2 , ..., P r } where r n. P r {P r PSm i P r GPS IC } which implies that i, PSm i GPS IC . The geographical properties of this set are indicative of whether m i IC can be mapped to only a single granularity level (i.e. a single location) in LH or a multiple granularity levels (i.e. more than one nodes in 19 the hierarchy) in LH. How many and which granularity levels should a m i map to, depends upon the level at which the service provider wants to provide information about the m i in question. Similarly we define a property subset of LH members as PSn j n j LH which can be written as PSn j ={Q 1 , Q 2 , Q 3 , ..., Q s } where s m. In addition, Q s {Q s PSn j Q s GPS LH } which implies that j, PSn j GPSLH. The process of mapping from IC to LH is then identifying for some m x IC one or more n y LH such that PSm x PSn v . This means that when m x maps to n y and all children of n y if m x can map to multiple granularity levels or m x maps only to n y if m x can map to a single granularity level. We assume that new members can join and old member can leave IC or LH any time. The deletion of members from the IC space is simple but addition of members to the IC space is more restrictive. If we want to add a new member to the IC space, then we first define a property set for the new member: PSm new_m ={P 1 , P 2 , P 3 , ..., P t } and add it to the IC only if the condition: P w {P w PSp new_m P w GPS IC } is satisfied. This scheme has an additional benefit of allowing the information service providers to have a control over what kind of information they wish to provide to the users. We present the following example to illustrate the mapping concept. IC = {Traffic, Stock, Restaurant, Weather, Important history dates, Road conditions} LH = {Country, State, City, Zip-code, Major-roads} GPS IC = {Surface-mobility, Roads, High, Low, Italian-food, StateName, Temp, CityName, Seat-availability, Zip, Traffic-jams, Stock-price, CountryName, MajorRoadName, Wars, Discoveries, World} GPS LH = {Country, CountrySize, StateName, CityName, Zip, MajorRoadName} Ps(IC Stock ) = {Stock-price, CountryName, High, Low} Ps(IC Traffic ) = {Surface-mobility, Roads, High, Low, Traffic-jams, CityName} Ps(IC Important dates in history ) = {World, Wars, Discoveries} Ps(IC Road conditions ) = {Precipitation, StateName, CityName} Ps(IC Restaurant ) = {Italian-food, Zip code} Ps(IC Weather ) = {StateName, CityName, Precipitation, Temperature} PS(LH Country ) = {CountryName, CountrySize} PS(LH State = {StateName, State size}, PS(LH City ) ={CityName, City size} PS(LH Zipcode ) = {ZipCodeNum } PS(LH Major roads ) = {MajorRoadName} Now, only PS(IC Stock ) PS Country . In addition, PS(IC Stock ) indicated that Stock can map to only a single location Country. When we consider the member Traffic of IC space, only PS(IC Traffic ) PS city . As PS(IC Traffic ) indicates that Traffic can map to only a single location, it maps only to City and none of its children. Now unlike Stock, mapping of Traffic with Major roads, which is a child of City, is meaningful. However service providers have right to control the granularity levels at which they want to provide information about a member of IC space. PS(IC Road conditions ) PS State and PS(IC Road conditions ) PS City . So Road conditions maps to State as well as City. As PS(IC Road conditions ) indicates that Road conditions can map to multiple granularity levels, Road conditions will also map to Zip Code and Major roads, which are the children of State and City. Similarly, Restaurant maps only to Zip code, and Weather maps to State, City and their children, Major Roads and Zip Code. LOCATION BASED INDEXING SCHEME This section discusses our location based indexing scheme (LBIS). The scheme is designed to conform to the LDD broadcast in our project DAYS. As discussed earlier, we use the containment property of LDD in the indexing scheme. This significantly limits the search of our required data to a particular portion of broadcast. Thus, we argue that the scheme provides bounded tuning time. We describe the architecture of our indexing scheme. Our scheme contains separate data buckets and index buckets. The index buckets are of two types. The first type is called the Major index. The Major index provides information about the types of data broadcasted. For example, if we intend to broadcast information like Entertainment, Weather, Traffic etc., then the major index points to either these major types of information and/or their main subtypes of information, the number of main subtypes varying from one information to another. This strictly limits number of accesses to a Major index. The Major index never points to the original data. It points to the sub indexes called the Minor index. The minor indexes are the indexes which actually points to the original data. We called these minor index pointers as Location Pointers as they points to the data which are associated with a location. Thus, our search for a data includes accessing of a major index and some minor indexes, the number of minor index varying depending on the type of information. Thus, our indexing scheme takes into account the hierarchical nature of the LDD, the Containment property, and requires our broadcast schedule to be clustered based on data type and location. The structure of the location hierarchy requires the use of different types of index at different levels. The structure and positions of index strictly depend on the location hierarchy as described in our mapping scheme earlier. We illustrate the implementation of our scheme with an example. The rules for framing the index are mentioned subsequently. 20 A1 Entertainment Resturant Movie A2 A3 A4 R1 R2 R3 R4 R5 R6 R7 R8 Weather KC SL JC SF Entertainment R1 R2 R3 R4 R5 R6 R7 R8 KC SL JC SF (A, R, NEXT = 8) 3, R5 4, R7 Type (S, L) ER W E EM (1, 4) (5, 4) (1, 4), (9, 4) (9, 4) Type (S, L) W E EM ER (1, 4) (5, 8) (5, 4) (9, 4) Type (S, L) E EM ER W (1, 8) (1, 4) (5, 4) (9, 4) A1 A2 A3 A4 Movie Resturant Weather 1 2 3 4 5 6 7 8 9 10 11 12 Major index Major index Major index Minor index Major index Minor index Figure 3. Location Mapped Information for Broadcast Figure 4. Data coupled with Location based Index Example: Let us suppose that our broadcast content contains IC entertainment and IC weather which is represented as shown in Fig. 3. Ai represents Areas of City and Ri represents roads in a certain area. The leaves of Weather structure represent four cities. The index structure is given in Fig. 4 which shows the position of major and minor index and data in the broadcast schedule. We propose the following rules for the creation of the air indexed broadcast schedule: The major index and the minor index are created. The major index contains the position and range of different types of data items (Weather and Entertainment, Figure 3) and their categories. The sub categories of Entertainment, Movie and Restaurant, are also in the index. Thus, the major index contains Entertainment (E), Entertainment-Movie (EM), Entertainment-Restaurant (ER), and Weather (W). The tuple (S, L) represents the starting position (S) of the data item and L represents the range of the item in terms of number of data buckets. The minor index contains the variables A, R and a pointer Next. In our example (Figure 3), road R represents the first node of area A. The minor index is used to point to actual data buckets present at the lowest levels of the hierarchy. In contrast, the major index points to a broader range of locations and so it contains information about main and sub categories of data. Index information is not incorporated in the data buckets. Index buckets are separate containing only the control information. The number of major index buckets m=#(IC), IC = {ic 1 , ic 2 , ic 3 ,...,ic n } where ic i represent information type and # represents the cardinality of the Information Content set IC. In this example, IC= {ic Movie , ic Weather , ic Restaurant } and so #(IC) =3. Hence, the number of major index buckets is 3. Mechanism to resolve the query is present in the java based coordinator in MU. For example, if a query Q is presented as Q (Entertainment, Movie, Road_1), then the resultant search will be for the EM information in the major index. We say, Q EM. Our proposed index works as follows: Let us suppose that an MU issues a query which is represented by Java Coordinator present in the MU as "Restaurant information on Road 7". This is resolved by the coordinator as Q ER. This means one has to search for ER unit of index in the major index. Let us suppose that the MU logs into the channel at R2. The first index it receives is a minor index after R2. In this index, value of Next variable = 4, which means that the next major index is present after bucket 4. The MU may go into doze mode. It becomes active after bucket 4 and receives the major index. It searches for ER information which is the first entry in this index. It is now certain that the MU will get the position of the data bucket in the adjoining minor index. The second unit in the minor index depicts the position of the required data R7. It tells that the data bucket is the first bucket in Area 4. The MU goes into doze mode again and becomes active after bucket 6. It gets the required data in the next bucket. We present the algorithm for searching the location based Index. Algorithm 1 Location based Index Search in DAYS 1. Scan broadcast for the next index bucket, found=false 2. While (not found) do 3. if bucket is Major Index then 4. Find the Type & Tuple (S, L) 5. if S is greater than 1, go into doze mode for S seconds 6. end if 7. Wake up at the S th bucket and observe the Minor Index 8. end if 9. if bucket is Minor Index then 10. if Type Requested not equal to Type found and (A,R) Request not equal to (A,R) found then 11. Go into doze mode till NEXT & repeat from step 3 12. end if 13. else find entry in Minor Index which points to data 14. Compute time of arrival T of data bucket 15. Go into doze mode till T 16. Wake up at T and access data, found = true 17. end else 18. end if 19. end While 21 PERFORMANCE EVALUATION Conservation of energy is the main concern when we try to access data from wireless broadcast. An efficient scheme should allow the mobile device to access its required data by staying active for a minimum amount of time. This would save considerable amount of energy. Since items are distributed based on types and are mapped to suitable locations, we argue that our broadcast deals with clustered data types. The mobile unit has to access a larger major index and a relatively much smaller minor index to get information about the time of arrival of data. This is in contrast to the exponential scheme where the indexes are of equal sizes. The example discussed and Algorithm 1 reveals that to access any data, we need to access the major index only once followed by one or more accesses to the minor index. The number of minor index access depends on the number of internal locations. As the number of internal locations vary for item to item (for example, Weather is generally associated with a City whereas traffic is granulated up to major and minor roads of a city), we argue that the structure of the location mapped information may be visualized as a forest which is a collection of general trees, the number of general trees depending on the types of information broadcasted and depth of a tree depending on the granularity of the location information associated with the information. For our experiments, we assume the forest as a collection of balanced M-ary trees. We further assume the M-ary trees to be full by assuming the presence of dummy nodes in different levels of a tree. Thus, if the number of data items is d and the height of the tree is m, then n= (m*d-1)/(m-1) where n is the number of vertices in the tree and i= (d-1)/(m-1) where i is the number of internal vertices. Tuning time for a data item involves 1 unit of time required to access the major index plus time required to access the data items present in the leaves of the tree. Thus, tuning time with d data items is t = log m d+1 We can say that tuning time is bounded by O(log m d). We compare our scheme with the distributed indexing and exponential scheme. We assume a flat broadcast and number of pages varying from 5000 to 25000. The various simulation parameters are shown in Table 1. Figure 5-8 shows the relative tuning times of three indexing algorithms, ie, the LBIS, exponential scheme and the distributed tree scheme. Figure 5 shows the result for number of internal location nodes = 3. We can see that LBIS significantly outperforms both the other schemes. The tuning time in LBIS ranges from approx 6.8 to 8. This large tuning time is due to the fact that after reaching the lowest minor index, the MU may have to access few buckets sequentially to get the required data bucket. We can see that the tuning time tends to become stable as the length of broadcast increases. In figure 6 we consider m= 4. Here we can see that the exponential and the distributed perform almost similarly, though the former seems to perform slightly better as the broadcast length increases. A very interesting pattern is visible in figure 7. For smaller broadcast size, the LBIS seems to have larger tuning time than the other two schemes. But as the length of broadcast increases, it is clearly visible the LBIS outperforms the other two schemes. The Distributed tree indexing shows similar behavior like the LBIS. The tuning time in LBIS remains low because the algorithm allows the MU to skip some intermediate Minor Indexes. This allows the MU to move into lower levels directly after coming into active mode, thus saving valuable energy. This action is not possible in the distributed tree indexing and hence we can observe that its tuning time is more than the LBIS scheme, although it performs better than the exponential scheme. Figure 8, in contrast, shows us that the tuning time in LBIS, though less than the other two schemes, tends to increase sharply as the broadcast length becomes greater than the 15000 pages. This may be attributed both due to increase in time required to scan the intermediate Minor Indexes and the length of the broadcast. But we can observe that the slope of the LBIS curve is significantly less than the other two curves. Table 1 Simulation Parameters P Definition Values N Number of data Items 5000 - 25000 m Number of internal location nodes 3, 4, 5, 6 B Capacity of bucket without index (for exponential index) 10,64,128,256 i Index base for exponential index 2,4,6,8 k Index size for distributed tree 8 bytes The simulation results establish some facts about our location based indexing scheme. The scheme performs better than the other two schemes in terms of tuning time in most of the cases. As the length of the broadcast increases, after a certain point, though the tuning time increases as a result of factors which we have described before, the scheme always performs better than the other two schemes. Due to the prescribed limit of the number of pages in the paper, we are unable to show more results. But these omitted results show similar trend as the results depicted in figure 5-8. CONCLUSION AND FUTURE WORK In this paper we have presented a scheme for mapping of wireless broadcast data with their locations. We have presented an example to show how the hierarchical structure of the location tree maps with the data to create LDD. We have presented a scheme called LBIS to index this LDD. We have used the containment property of LDD in the scheme that limits the search to a narrow range of data in the broadcast, thus saving valuable energy in the device. The mapping of data with locations and the indexing scheme will be used in our DAYS project to create the push based architecture. The LBIS has been compared with two other prominent indexing schemes, i.e., the distributed tree indexing scheme and the exponential indexing scheme. We showed in our simulations that the LBIS scheme has the lowest tuning time for broadcasts having large number of pages, thus saving valuable battery power in the MU. 22 In the future work we try to incorporate pull based architecture in our DAYS project. Data from the server is available for access by the global users. This may be done by putting a request to the source server. The query in this case is a global query. It is transferred from the user's source server to the destination server through the use of LEO satellites. We intend to use our LDD scheme and data staging architecture in the pull based architecture. We will show that the LDD scheme together with the data staging architecture significantly improves the latency for global as well as local query. REFERENCES [1] Acharya, S., Alonso, R. Franklin, M and Zdonik S. Broadcast disk: Data management for asymmetric communications environments. In Proceedings of ACM SIGMOD Conference on Management of Data, pages 199210, San Jose, CA, May 1995. [2] Chen, M.S.,Wu, K.L. and Yu, P. S. Optimizing index allocation for sequential data broadcasting in wireless mobile computing. IEEE Transactions on Knowledge and Data Engineering (TKDE), 15(1):161173, January/February 2003. Figure 5. Broadcast Size (# buckets) Dist tree Expo LBIS Figure 6. Broadcast Size (# buckets) Dist tree Expo LBIS Figure 7. Broadcast Size (# buckets) Dist tree Expo LBIS Figure 8. Broadcast Size (# buckets) Dist tree Expo LBIS Av era g e tunin g t i me Av era g e t u n i n g t i me Av era g e tunin g t i me Av era g e t u n i n g t i me 23 [3] Hu, Q. L., Lee, D. L. and Lee, W.C. Performance evaluation of a wireless hierarchical data dissemination system. In Proceedings of the 5 th Annual ACM International Conference on Mobile Computing and Networking (MobiCom'99), pages 163173, Seattle, WA, August 1999. [4] Hu, Q. L. Lee, W.C. and Lee, D. L. Power conservative multi-attribute queries on data broadcast. In Proceedings of the 16th International Conference on Data Engineering (ICDE'00), pages 157166, San Diego, CA, February 2000. [5] Imielinski, T., Viswanathan, S. and Badrinath. B. R. Power efficient filtering of data on air. In Proceedings of the 4th International Conference on Extending Database Technology (EDBT'94), pages 245258, Cambridge, UK, March 1994. [6] Imielinski, T., Viswanathan, S. and Badrinath. B. R. Data on air Organization and access. IEEE Transactions on Knowledge and Data Engineering (TKDE), 9(3):353372, May/June 1997. [7] Shih, E., Bahl, P. and Sinclair, M. J. Wake on wireless: An event driven energy saving strategy for battery operated devices. In Proceedings of the 8th Annual ACM International Conference on Mobile Computing and Networking (MobiCom'02), pages 160171, Atlanta, GA, September 2002. [8] Shivakumar N. and Venkatasubramanian, S. Energy-efficient indexing for information dissemination in wireless systems. ACM/Baltzer Journal of Mobile Networks and Applications (MONET), 1(4):433446, December 1996. [9] Tan K. L. and Yu, J. X. Energy efficient filtering of non uniform broadcast. In Proceedings of the 16th International Conference on Distributed Computing Systems (ICDCS'96), pages 520527, Hong Kong, May 1996. [10] Viredaz, M. A., Brakmo, L. S. and Hamburgen, W. R. Energy management on handheld devices. ACM Queue, 1(7):4452, October 2003. [11] Garg, N. Kumar, V., & Dunham, M.H. "Information Mapping and Indexing in DAYS", 6th International Workshop on Mobility in Databases and Distributed Systems, in conjunction with the 14th International Conference on Database and Expert Systems Applications September 1-5, Prague, Czech Republic, 2003. [12] Acharya D., Kumar, V., & Dunham, M.H. InfoSpace: Hybrid and Adaptive Public Data Dissemination System for Ubiquitous Computing". Accepted for publication in the special issue of Pervasive Computing. Wiley Journal for Wireless Communications and Mobile Computing, 2004. [13] Acharya D., Kumar, V., & Prabhu, N. Discovering and using Web Services in M-Commerce, Proceedings for 5th VLDB Workshop on Technologies for E-Services, Toronto, Canada,2004. [14] Acharya D., Kumar, V. Indexing Location Dependent Data in broadcast environment. Accepted for publication, JDIM special issue on Distributed Data Management, 2005. [15] Flinn, J., Sinnamohideen, S., & Satyanarayan, M. Data Staging on Untrusted Surrogates, Intel Research, Pittsburg, Unpublished Report, 2003. [16] Seydim, A.Y., Dunham, M.H. & Kumar, V. Location dependent query processing, Proceedings of the 2nd ACM international workshop on Data engineering for wireless and mobile access, p.47-53, Santa Barbara, California, USA, 2001. [17] Xu, J., Lee, W.C., Tang., X. Exponential Index: A Parameterized Distributed Indexing Scheme for Data on Air. In Proceedings of the 2nd ACM/USENIX International Conference on Mobile Systems, Applications, and Services (MobiSys'04), Boston, MA, June 2004. 24

containment;indexing scheme;access efficiency;indexing;Wireless data broadcast;mapping function;location based services;wireless;energy conservation;location dependent data;broadcast;push based architecture;data dissemination;data staging

129

Lossless Online Bayesian Bagging

Bagging frequently improves the predictive performance of a model. An online version has recently been introduced, which attempts to gain the benefits of an online algorithm while approximating regular bagging. However, regular online bagging is an approximation to its batch counterpart and so is not lossless with respect to the bagging operation. By operating under the Bayesian paradigm, we introduce an online Bayesian version of bagging which is exactly equivalent to the batch Bayesian version, and thus when combined with a lossless learning algorithm gives a completely lossless online bagging algorithm. We also note that the Bayesian formulation resolves a theoretical problem with bagging, produces less variability in its estimates, and can improve predictive performance for smaller data sets.

Introduction In a typical prediction problem, there is a trade-off between bias and variance, in that after a certain amount of fitting, any increase in the precision of the fit will cause an increase in the prediction variance on future observations. Similarly, any reduction in the prediction variance causes an increase in the expected bias for future predictions. Breiman (1996) introduced bagging as a method of reducing the prediction variance without affecting the prediction bias. As a result, predictive performance can be significantly improved. Bagging, short for "Bootstrap AGGregatING", is an ensemble learning method. Instead of making predictions from a single model fit on the observed data, bootstrap samples are taken of the data, the model is fit on each sample, and the predictions are averaged over all of the fitted models to get the bagged prediction. Breiman (1996) explains that bagging works well for unstable modeling procedures, i.e. those for which the conclusions are sensitive to small changes in the data. He also gives a theoretical explanation of how bagging works, demonstrating the reduction in mean-squared prediction error for unstable c 2004 Herbert K. H. Lee and Merlise A. Clyde. Lee and Clyde procedures. Breiman (1994) demonstrated that tree models, among others, are empirically unstable. Online bagging (Oza and Russell, 2001) was developed to implement bagging sequentially , as the observations appear, rather than in batch once all of the observations have arrived. It uses an asymptotic approximation to mimic the results of regular batch bagging, and as such it is not a lossless algorithm. Online algorithms have many uses in modern computing. By updating sequentially, the update for a new observation is relatively quick compared to re-fitting the entire database, making real-time calculations more feasible. Such algorithms are also advantageous for extremely large data sets where reading through the data just once is time-consuming, so batch algorithms which would require multiple passes through the data would be infeasible. In this paper, we consider a Bayesian version of bagging (Clyde and Lee, 2001) based on the Bayesian bootstrap (Rubin, 1981). This overcomes a technical difficulty with the usual bootstrap in bagging. It also leads to a theoretical reduction in variance over the bootstrap for certain classes of estimators, and a significant reduction in observed variance and error rates for smaller data sets. We present an online version which, when combined with a lossless online model-fitting algorithm, continues to be lossless with respect to the bagging operation, in contrast to ordinary online bagging. The Bayesian approach requires the learning base algorithm to accept weighted samples. In the next section we review the basics of the bagging algorithm, of online bagging, and of Bayesian bagging. Next we introduce our online Bayesian bagging algorithm. We then demonstrate its efficacy with classification trees on a variety of examples. Bagging In ordinary (batch) bagging, bootstrap re-sampling is used to reduce the variability of an unstable estimator. A particular model or algorithm, such as a classification tree, is specified for learning from a set of data and producing predictions. For a particular data set X m , denote the vector of predictions (at the observed sites or at new locations) by G(X m ). Denote the observed data by X = (x 1 , . . . , x n ). A bootstrap sample of the data is a sample with replacement, so that X m = (x m 1 , . . . , x m n ), where m i {1, . . . , n} with repetitions allowed. X m can also be thought of as a re-weighted version of X, where the weights, (m) i are drawn from the set {0, 1 n , 2 n , . . . , 1}, i.e., n (m) i is the number of times that x i appears in the mth bootstrap sample. We denote the weighted sample as (X, (m) ). For each bootstrap sample, the model produces predictions G(X m ) = G(X m ) 1 , . . . , G(X m ) P where P is the number of prediction sites. M total bootstrap samples are used. The bagged predictor for the jth element is then 1 M M m =1 G(X m ) j = 1 M M m =1 G(X, (m) ) j , or in the case of classification, the jth element is predicted to belong to the most frequently predicted category by G(X 1 ) j , . . . , G(X M ) j . A version of pseudocode for implementing bagging is 1. For m {1, . . . , M }, 144 Lossless Online Bayesian Bagging (a) Draw a bootstrap sample, X m , from X. (b) Find predicted values G(X m ). 2. The bagging predictor is 1 M M m =1 G (X m ). Equivalently, the bootstrap sample can be converted to a weighted sample (X, (m) ) where the weights (m) i are found by taking the number of times x i appears in the bootstrap sample and dividing by n. Thus the weights will be drawn from {0, 1 n , 2 n , . . . , 1} and will sum to 1. The bagging predictor using the weighted formulation is 1 M M m =1 G (X m , (m) ) for regression, or the plurality vote for classification. 2.1 Online Bagging Online bagging (Oza and Russell, 2001) was recently introduced as a sequential approximation to batch bagging. In batch bagging, the entire data set is collected, and then bootstrap samples are taken from the whole database. An online algorithm must process observations as they arrive, and thus each observation must be resampled a random number of times when it arrives. The algorithm proposed by Oza and Russell resamples each observation according to a Poisson random variable with mean 1, i.e., P (K m = k) = exp(-1)/k!, where K m is the number of resamples in "bootstrap sample" m, K m {0, 1, . . .}. Thus as each observation arrives, it is added K m times to X m , and then G(X m ) is updated, and this is done for m {1, . . . , M }. Pseudocode for online bagging is For i {1, . . . , n}, 1. For m {1, . . . , M }, (a) Draw a weight K m from a Poisson(1) random variable and add K m copies of x i to X m . (b) Find predicted values G(X m ). 2. The current bagging predictor is 1 M M m =1 G (X m ). Ideally, step 1(b) is accomplished with a lossless online update that incorporates the K m new points without refitting the entire model. We note that n may not be known ahead of time, but the bagging predictor is a valid approximation at each step. Online bagging is not guaranteed to produce the same results as batch bagging. In particular, it is easy to see that after n points have been observed, there is no guarantee that X m will contain exactly n points, as the Poisson weights are not constrained to add up to n like a regular bootstrap sample. While it has been shown (Oza and Russell, 2001) that these samples converge asymptotically to the appropriate bootstrap samples, there may be some discrepancy in practice. Thus while it can be combined with a lossless online learning algorithm (such as for a classification tree), the bagging part of the online ensemble procedure is not lossless. 145 Lee and Clyde 2.2 Bayesian Bagging Ordinary bagging is based on the ordinary bootstrap, which can be thought of as replacing the original weights of 1 n on each point with weights from the set {0, 1 n , 2 n , . . . , 1}, with the total of all weights summing to 1. A variation is to replace the ordinary bootstrap with the Bayesian bootstrap (Rubin, 1981). The Bayesian approach treats the vector of weights as unknown parameters and derives a posterior distribution for , and hence G(X, ). The non-informative prior n i =1 -1 i , when combined with the multinomial likelihood, leads to a Dirichlet n (1, . . . , 1) distribution for the posterior distribution of . The full posterior distribution of G(X, ) can be estimated by Monte Carlo methods: generate (m) from a Dirichlet n (1, . . . , 1) distribution and then calculate G(X, (m) ) for each sample. The average of G(X, (m) ) over the M samples corresponds to the Monte Carlo estimate of the posterior mean of G(X, ) and can be viewed as a Bayesian analog of bagging (Clyde and Lee, 2001). In practice, we may only be interested in a point estimate, rather than the full posterior distribution. In this case, the Bayesian bootstrap can be seen as a continuous version of the regular bootstrap. Thus Bayesian bagging can be achieved by generating M Bayesian bootstrap samples, and taking the average or majority vote of the G(X, (m) ). This is identical to regular bagging except that the weights are continuous-valued on (0, 1), instead of being restricted to the discrete set {0, 1 n , 2 n , . . . , 1}. In both cases, the weights must sum to 1. In both cases, the expected value of a particular weight is 1 n for all weights, and the expected correlation between weights is the same (Rubin, 1981). Thus Bayesian bagging will generally have the same expected point estimates as ordinary bagging. The variability of the estimate is slightly smaller under Bayesian bagging, as the variability of the weights is n n +1 times that of ordinary bagging. As the sample size grows large, this factor becomes arbitrarily close to one, but we do note that it is strictly less than one, so the Bayesian approach does give a further reduction in variance compared to the standard approach. In practice, for smaller data sets, we often find a significant reduction in variance, possibly because the use of continuous-valued weights leads to fewer extreme cases than discrete-valued weights. Pseudocode for Bayesian bagging is 1. For m {1, . . . , M }, (a) Draw random weights (m) from a Dirichlet n (1, . . . , 1) to produce the Bayesian bootstrap sample (X, (m) ). (b) Find predicted values G(X, (m) ). 2. The bagging predictor is 1 M M m =1 G (X, (m) ). Use of the Bayesian bootstrap does have a major theoretical advantage, in that for some problems, bagging with the regular bootstrap is actually estimating an undefined quantity. To take a simple example, suppose one is bagging the fitted predictions for a point y from a least-squares regression problem. Technically, the full bagging estimate is 1 M 0 m ^ y m where m ranges over all possible bootstrap samples, M 0 is the total number of possible bootstrap samples, and ^ y m is the predicted value from the model fit using the mth bootstrap sample. The issue is that one of the possible bootstrap samples contains the 146 Lossless Online Bayesian Bagging first data point replicated n times, and no other data points. For this bootstrap sample, the regression model is undefined (since at least two different points are required), and so ^ y and thus the bagging estimator are undefined. In practice, only a small sample of the possible bootstrap samples is used, so the probability of drawing a bootstrap sample with an undefined prediction is very small. Yet it is disturbing that in some problems, the bagging estimator is technically not well-defined. In contrast, the use of the Bayesian bootstrap completely avoids this problem. Since the weights are continuous-valued, the probability that any weight is exactly equal to zero is zero. Thus with probability one, all weights are strictly positive, and the Bayesian bagging estimator will be well-defined (assuming the ordinary estimator on the original data is well-defined). We note that the Bayesian approach will only work with models that have learning algorithms that handle weighted samples. Most standard models either have readily available such algorithms, or their algorithms are easily modified to accept weights, so this restriction is not much of an issue in practice. Online Bayesian Bagging Regular online bagging cannot be exactly equivalent to the batch version because the Poisson counts cannot be guaranteed to sum to the number of actual observations. Gamma random variables can be thought of as continuous analogs of Poisson counts, which motivates our derivation of Bayesian online bagging. The key is to recall a fact from basic probability -- a set of independent gamma random variables divided by its sum has a Dirichlet distribution, i.e., If w i ( i , 1), then w 1 w i , w 2 w i , . . . , w k w i Dirichlet n ( 1 , 2 , . . . , k ) . (See for example, Hogg and Craig, 1995, pp. 187188.) This relationship is a common method for generating random draws from a Dirichlet distribution, and so is also used in the implementation of batch Bayesian bagging in practice. Thus in the online version of Bayesian bagging, as each observation arrives, it has a realization of a Gamma(1) random variable associated with it for each bootstrap sample, and the model is updated after each new weighted observation. If the implementation of the model requires weights that sum to one, then within each (Bayesian) bootstrap sample, all weights can be re-normalized with the new sum of gammas before the model is updated. At any point in time, the current predictions are those aggregated across all bootstrap samples, just as with batch bagging. If the model is fit with an ordinary lossless online algorithm, as exists for classification trees (Utgoff et al., 1997), then the entire online Bayesian bagging procedure is completely lossless relative to batch Bayesian bagging. Furthermore, since batch Bayesian bagging gives the same mean results as ordinary batch bagging, online Bayesian bagging also has the same expected results as ordinary batch bagging. Pseudocode for online Bayesian bagging is For i {1, . . . , n}, 1. For m {1, . . . , M }, 147 Lee and Clyde (a) Draw a weight (m) i from a Gamma(1, 1) random variable, associate weight with x i , and add x i to X. (b) Find predicted values G(X, (m) ) (renormalizing weights if necessary). 2. The current bagging predictor is 1 M M m =1 G (X, (m) ). In step 1(b), the weights may need to be renormalized (by dividing by the sum of all current weights) if the implementation requires weights that sum to one. We note that for many models, such as classification trees, this renormalization is not a major issue; for a tree, each split only depends on the relative weights of the observations at that node, so nodes not involving the new observation will have the same ratio of weights before and after renormalization and the rest of the tree structure will be unaffected; in practice, in most implementations of trees (including that used in this paper), renormalization is not necessary. We discuss the possibility of renormalization in order to be consistent with the original presentation of the bootstrap and Bayesian bootstrap, and we note that ordinary online bagging implicitly deals with this issue equivalently. The computational requirements of Bayesian versus ordinary online bagging are comparable . The procedures are quite similar, with the main difference being that the fitting algorithm must handle non-integer weights for the Bayesian version. For models such as trees, there is no significant additional computational burden for using non-integer weights. Examples We demonstrate the effectiveness of online Bayesian bagging using classification trees. Our implementation uses the lossless online tree learning algorithms (ITI) of Utgoff et al. (1997) (available at http://www.cs.umass.edu/lrn/iti/). We compared Bayesian bagging to a single tree, ordinary batch bagging, and ordinary online bagging, all three of which were done using the minimum description length criterion (MDL), as implemented in the ITI code, to determine the optimal size for each tree. To implement Bayesian bagging, the code was modified to account for weighted observations. We use a generalized MDL to determine the optimal tree size at each stage, replacing all counts of observations with the sum of the weights of the observations at that node or leaf with the same response category. Replacing the total count directly with the sum of the weights is justified by looking at the multinomial likelihood when written as an exponential family in canonical form; the weights enter through the dispersion parameter and it is easily seen that the unweighted counts are replaced by the sums of the weights of the observations that go into each count. To be more specific, a decision tree typically operates with a multinomial likelihood, leaves j classes k p n jk jk , where p jk is the true probability that an observation in leaf j will be in class k, and n jk is the count of data points in leaf j in class k. This is easily re-written as the product over all observations, n i =1 p i where if observation i is in leaf j and a member of class k then p i = p jk . For simplicity, we consider the case k = 2 as the generalization to larger k is straightforward. Now consider a single point, y, which takes values 0 or 1 depending on which class is it a member of. Transforming to the canonical parameterization, let = p 1-p , 148 Lossless Online Bayesian Bagging where p is the true probability that y = 1. Writing the likelihood in exponential family form gives exp y + log 1 1+exp{} a where a is the dispersion parameter, which would be equal to 1 for a standard data set, but would be the reciprocal of the weight for that observation in a weighted data set. Thus the likelihood for an observation y with weight w is exp y + log 1 1+exp{} (1/w) = p wy (1 - p) w (1-y) and so returning to the full multinomial, the original counts are simply replaced by the weighted counts. As MDL is a penalized likelihood criterion, we thus use the weighted likelihood and replace each count with a sum of weights. We note that for ordinary online bagging, using a single Poisson weight K with our generalized MDL is exactly equivalent to including K copies of the data point in the data set and using regular MDL. Table 1 shows the data sets we used for classification problems, the number of classes in each data set, and the sizes of their respective training and test partitions. Table 2 displays the results of our comparison study. All of the data sets, except the final one, are available online at http://www.ics.uci.edu/mlearn/MLRepository.html, the UCI Machine Learning Repository. The last data set is described in Lee (2001). We compare the results of training a single classification tree, ordinary batch bagging, online bagging, and Bayesian online bagging (or equivalently Bayesian batch). For each of the bagging techniques, 100 bootstrap samples were used. For each data set, we repeated 1000 times the following procedure : randomly choose a training/test partition; fit a single tree, a batch bagged tree, an online bagged tree, and a Bayesian bagged tree; compute the misclassification error rate for each fit. Table 2 reports the average error rate for each method on each data set, as well as the estimated standard error of this error rate. Size of Size of Number of Training Test Data Set Classes Data Set Data Set Breast cancer (WI) 2 299 400 Contraceptive 3 800 673 Credit (German) 2 200 800 Credit (Japanese) 2 290 400 Dermatology 6 166 200 Glass 7 164 50 House votes 2 185 250 Ionosphere 2 200 151 Iris 3 90 60 Liver 3 145 200 Pima diabetes 2 200 332 SPECT 2 80 187 Wine 3 78 100 Mushrooms 2 1000 7124 Spam 2 2000 2601 Credit (American) 2 4000 4508 Table 1: Sizes of the example data sets 149 Lee and Clyde Bayesian Single Batch Online Online/Batch Data Set Tree Bagging Bagging Bagging Breast cancer (WI) 0.055 (.020) 0.045 (.010) 0.045 (.010) 0.041 (.009) Contraceptive 0.522 (.019) 0.499 (.017) 0.497 (.017) 0.490 (.016) Credit (German) 0.318 (.022) 0.295 (.017) 0.294 (.017) 0.285 (.015) Credit (Japanese) 0.155 (.017) 0.148 (.014) 0.147 (.014) 0.145 (.014) Dermatology 0.099 (.033) 0.049 (.017) 0.053 (.021) 0.047 (.019) Glass 0.383 (.081) 0.357 (.072) 0.361 (.074) 0.373 (.075) House votes 0.052 (.011) 0.049 (.011) 0.049 (.011) 0.046 (.010) Ionosphere 0.119 (.026) 0.094 (.022) 0.099 (.022) 0.096 (.021) Iris 0.062 (.029) 0.057 (.026) 0.060 (.025) 0.058 (.025) Liver 0.366 (.036) 0.333 (.032) 0.336 (.034) 0.317 (.033) Pima diabetes 0.265 (.027) 0.250 (.020) 0.247 (.021) 0.232 (.017) SPECT 0.205 (.029) 0.200 (.030) 0.202 (.031) 0.190 (.027) Wine 0.134 (.042) 0.094 (.037) 0.101 (.037) 0.085 (.034) Mushrooms 0.004 (.003) 0.003 (.002) 0.003 (.002) 0.003 (.002) Spam 0.099 (.008) 0.075 (.005) 0.077 (.005) 0.077 (.005) Credit (American) 0.350 (.007) 0.306 (.005) 0.306 (.005) 0.305 (.006) Table 2: Comparison of average classification error rates (with standard error) We note that in all cases, both online bagging techniques produce results similar to ordinary batch bagging, and all bagging methods significantly improve upon the use of a single tree. However, for smaller data sets (all but the last three), online/batch Bayesian bagging typically both improves prediction performance and decreases prediction variability. Discussion Bagging is a useful ensemble learning tool, particularly when models sensitive to small changes in the data are used. It is sometimes desirable to be able to use the data in an online fashion. By operating in the Bayesian paradigm, we can introduce an online algorithm that will exactly match its batch Bayesian counterpart. Unlike previous versions of online bagging, the Bayesian approach produces a completely lossless bagging algorithm. It can also lead to increased accuracy and decreased prediction variance for smaller data sets. Acknowledgments This research was partially supported by NSF grants DMS 0233710, 9873275, and 9733013. The authors would like to thank two anonymous referees for their helpful suggestions. 150 Lossless Online Bayesian Bagging References L. Breiman. Heuristics of instability in model selection. Technical report, University of California at Berkeley, 1994. L. Breiman. Bagging predictors. Machine Learning, 26(2):123140, 1996. M. A. Clyde and H. K. H. Lee. Bagging and the Bayesian bootstrap. In T. Richardson and T. Jaakkola, editors, Artificial Intelligence and Statistics 2001, pages 169174, 2001. R. V. Hogg and A. T. Craig. Introduction to Mathematical Statistics. Prentice-Hall, Upper Saddle River, NJ, 5th edition, 1995. H. K. H. Lee. Model selection for neural network classification. Journal of Classification, 18:227243, 2001. N. C. Oza and S. Russell. Online bagging and boosting. In T. Richardson and T. Jaakkola, editors, Artificial Intelligence and Statistics 2001, pages 105112, 2001. D. B. Rubin. The Bayesian bootstrap. Annals of Statistics, 9:130134, 1981. P. E. Utgoff, N. C. Berkman, and J. A. Clouse. Decision tree induction based on efficient tree restructuring. Machine Learning, 29:544, 1997. 151

classification;Dirichlet Distribution;online bagging;bootstrap;Classification Tree;Bayesian Bootstrap;mean-squared prediction error;Bayesian bagging;bagging;lossless learning algorithm

13

A Machine Learning Based Approach for Table Detection on The Web

Table is a commonly used presentation scheme, especially for describing relational information. However, table understanding remains an open problem. In this paper, we consider the problem of table detection in web documents. Its potential applications include web mining, knowledge management , and web content summarization and delivery to narrow-bandwidth devices. We describe a machine learning based approach to classify each given table entity as either genuine or non-genuine . Various features re ecting the layout as well as content characteristics of tables are studied. In order to facilitate the training and evaluation of our table classi er, we designed a novel web document table ground truthing protocol and used it to build a large table ground truth database. The database consists of 1,393 HTML les collected from hundreds of di erent web sites and contains 11,477 leaf &lt;TABLE&gt; elements, out of which 1,740 are genuine tables. Experiments were conducted using the cross validation method and an F-measure of 95 : 89% was achieved.

INTRODUCTION The increasing ubiquity of the Internet has brought about a constantly increasing amount of online publications. As a compact and e cient way to present relational information , tables are used frequently in web documents. Since tables are inherently concise as well as information rich, the automatic understanding of tables has many applications including knowledge management, information retrieval, web Copyright is held by the author/owner(s). WWW2002 , May 711, 2002, Honolulu, Hawaii, USA. ACM 1-58113-449-5/02/0005. mining, summarization, and content delivery to mobile devices . The processes of table understanding in web documents include table detection, functional and structural analysis and nally table interpretation 6]. In this paper, we concentrate on the problem of table detection. The web provides users with great possibilities to use their own style of communication and expressions. In particular, people use the &lt;TABLE&gt; tag not only for relational information display but also to create any type of multiple-column layout to facilitate easy viewing, thus the presence of the &lt;TABLE&gt; tag does not necessarily indicate the presence of a relational table. In this paper, we de ne genuine tables to be document entities where a two dimensional grid is semantically signi cant in conveying the logical relations among the cells 10]. Conversely, Non-genuine tables are document entities where &lt;TABLE&gt; tags are used as a mechanism for grouping contents into clusters for easy viewing only. Figure 1 gives a few examples of genuine and non-genuine tables. While genuine tables in web documents could also be created without the use of &lt;TABLE&gt; tags at all, we do not consider such cases in this article as they seem very rare from our experience . Thus, in this study, Table detection refers to the technique which classi es a document entity enclosed by the &lt;TABLE&gt;&lt;/TABLE&gt; tags as genuine or non-genuine tables. Several researchers have reported their work on web table detection 2, 10, 6, 14]. In 2], Chen et al. used heuristic rules and cell similarities to identify tables. They tested their table detection algorithm on 918 tables from airline information web pages and achieved an F-measure of 86 : 50%. Penn et al. proposed a set of rules for identifying genuinely tabular information and news links in HTML documents 10]. They tested their algorithm on 75 web site front-pages and achieved an F-measure of 88 : 05%. Yoshida et al. proposed a method to integrate WWW tables according to the category of objects presented in each table 14]. Their data set contains 35,232 table tags gathered from the web. They estimated their algorithm parameters using all of table data and then evaluated algorithm accuracy on 175 of the tables. The average F-measure reported in their paper is 82 : 65%. These previous methods all relied on heuristic rules and were only tested on a database that is either very small 10], or highly domain speci c 2]. Hurst mentioned that a Naive Bayes classi er algorithm produced adequate results but no detailed algorithm and experimental information was provided 6]. We propose a new machine learning based approach for 242 Figure 1: Examples of genuine and non-genuine tables. table detection from generic web documents. In particular , we introduce a set of novel features which re ect the layout as well as content characteristics of tables. These features are used in classi ers trained on thousands of examples . To facilitate the training and evaluation of the table classi ers, we designed a novel web document table ground truthing protocol and used it to build a large table ground truth database. The database consists of 1,393 HTML les collected from hundreds of di erent web sites and contains 11,477 leaf &lt;TABLE&gt; elements, out of which 1,740 are genuine tables. Experiments on this database using the cross validation method demonstrate signi cant performance improvements over previous methods. The rest of the paper is organized as follows. We describe our feature set in Section 2, followed by a brief discussion of the classi ers we experimented with in Section 3. In Section 4, we present a novel table ground truthing protocol and explain how we built our database. Experimental results are then reported in Section 5 and we conclude with future directions in Section 6. FEATURES FOR WEB TABLE DETECTION Feature selection is a crucial step in any machine learning based methods. In our case, we need to nd a combination of features that together provide signi cant separation between genuine and non-genuine tables while at the same time constrain the total number of features to avoid the curse of dimensionality. Past research has clearly indicated that layout and content are two important aspects in table understanding 6]. Our features were designed to capture both of these aspects. In particular, we developed 16 features which can be categorized into three groups: seven layout features, eight content type features and one word group feature. In the rst two groups, we attempt to capture the global composition of tables as well as the consistency within the whole table and across rows and columns. The last feature looks at words used in tables and is derived directly from the vector space model commonly used in Information Retrieval. Before feature extraction, each HTML document is rst parsed into a document hierarchy tree using Java Swing XML parser with W3C HTML 3.2 DTD 10]. A &lt;TABLE&gt; node is said to be a leaf table if and only if there are no &lt;TABLE&gt; nodes among its children 10]. Our experience indicates that almost all genuine tables are leaf tables. Thus in this study only leaf tables are considered candidates for genuine tables and are passed on to the feature extraction stage. In the following we describe each feature in detail. 2.1 Layout Features In HTML documents, although tags like &lt;TR&gt; and &lt;TD&gt; (or &lt;TH&gt; ) may be assumed to delimit table rows and table cells, they are not always reliable indicators of the number of rows and columns in a table. Variations can be caused by spanning cells created using &lt;ROWSPAN&gt; and &lt;COLSPAN&gt; tags. Other tags such as &lt;BR&gt; could be used to move content into the next row. Therefore to extract layout features reliably one can not simply count the number of &lt;TR&gt; 's and &lt;TD&gt; 's. For this purpose, we maintain a matrix to record all 243 the cell spanning information and serve as a pseudo rendering of the table. Layout features based on row or column numbers are then computed from this matrix. Given a table T , assuming its numbers of rows and columns are rn and cn respectively, we compute the following layout features: Average number of columns, computed as the average number of cells per row: c = 1 rn rn X i =1 c i where c i is the number of cells in row i , i = 1 ::: rn Standard deviation of number of columns: dC = v u u t 1 rn rn X i =1 ( c i ; c ) ( c i ; c ) Average number of rows, computed as the average number of cells per column: r = 1 rn cn X i =1 r i where r i is the number of cells in column i , i = 1 ::: cn Standard deviation of number of rows: dR = v u u t 1 cn cn X i =1 ( r i ; r ) ( r i ; r ) : Since the majority of tables in web documents contain characters, we compute three more layout features based on cell length in terms of number of characters: Average overall cell length: cl = 1 en P en i =1 cl i , where en is the total number of cells in a given table and cl i is the length of cell i , i = 1 ::: en Standard deviation of cell length: dCL = v u u t 1 en en X i =1 ( cl i ; cl ) ( cl i ; cl ) Average Cumulative length consistency , CLC . The last feature is designed to measure the cell length consistency along either row or column directions. It is inspired by the fact that most genuine tables demonstrate certain consistency either along the row or the column direction, but usually not both, while non-genuine tables often show no consistency in either direction. First, the average cumulative within-row length consistency, CLC r , is computed as follows. Let the set of cell lengths of the cells from row i be R i , i = 1 ::: r (considering only non-spanning cells): 1. Compute the mean cell length, m i , for row R i . 2. Compute cumulative length consistency within each R i : CLC i = X cl 2R i LC cl : Here LC cl is de ned as: LC cl = 0 : 5 ; D , where D = min fj cl ; m i j =m i 1 : 0 g . Intuitively, LC cl measures the degree of consistency between cl and the mean cell length, with ; 0 : 5 indicating extreme inconsistency and 0 : 5 indicating extreme consistency. When most cells within R i are consistent, the cumulative measure CLC i is positive, indicating a more or less consistent row. 3. Take the average across all rows: CLC r = 1 r r X i =1 CLC i : After the within-row length consistency CLC r is computed , the within-column length consistency CLC c is computed in a similar manner. Finally, the overall cumulative length consistency is computed as CLC = max( CLC r CLC c ). 2.2 Content Type Features Web documents are inherently multi-media and has more types of content than any traditional documents. For example , the content within a &lt;TABLE&gt; element could include hyperlinks, images, forms, alphabetical or numerical strings, etc. Because of the relational information it needs to convey, a genuine table is more likely to contain alpha or numerical strings than, say, images. The content type feature was designed to re ect such characteristics. We de ne the set of content types T = f Image Form Hyperlink Alphabetical Digit Empty Others g . Our content type features include: The histogram of content type for a given table. This contributes 7 features to the feature set Average content type consistency , CTC . The last feature is similar to the cell length consistency feature . First, within-row content type consistency CTC r is computed as follows. Let the set of cell type of the cells from row i as T i , i = 1 ::: r (again, considering only non-spanning cells): 1. Find the dominant type, DT i , for T i . 2. Compute the cumulative type consistency with each row R i , i = 1 ::: r : CTC i = X ct 2R i D where D = 1 if ct is equal to DT i and D = ; 1, otherwise . 3. Take the average across all rows: CTC r = 1 r r X i =1 CTC i The within-column type consistency is then computed in a similar manner. Finally, the overall cumulative type consistency is computed as: CTC = max( CTC r CTC c ). 244 2.3 Word Group Feature If we treat each table as a \mini-document&quot; by itself, table classi cation can be viewed as a document categorization problem with two broad categories: genuine tables and non-genuine tables. We designed the word group feature to incorporate word content for table classi cation based on techniques developed in information retrieval 7, 13]. After morphing 11] and removing the infrequent words, we obtain the set of words found in the training data, W . We then construct weight vectors representing genuine and non-genuine tables and compare that against the frequency vector from each new incoming table. Let Z represent the non-negative integer set. The following functions are de ned on set W . df G : W ! Z , where df G ( w i ) is the number of genuine tables which include word w i , i = 1 ::: jW j tf G : W ! Z , where tf G ( w i ) is the number of times word w i , i = 1 ::: jW j , appears in genuine tables df N : W ! Z , where df N ( w i ) is the number of non-genuine tables which include word w i , i = 1 ::: jW j tf N : W ! Z , where tf N ( w i ) is the number of times word w i , i = 1 ::: jW j , appears in non-genuine tables. tf T : W ! Z , where tf T ( w i ) is the number of times word w i , w i 2 W appears in a new test table. To simplify the notations, in the following discussion, we will use df Gi , tf Gi , df Ni and tf Ni to represent df G ( w i ), tf G ( w i ), df N ( w i ) and tf N ( w i ), respectively. Let N G , N N be the number of genuine tables and non-genuine tables in the training collection, respectively and let C = max( N G N N ). Without loss of generality, we assume N G 6 = 0 and N N 6 = 0. For each word w i in W , i = 1 ::: jW j , two weights, p G i and p N i are computed: p G i = 8 &lt; : tf Gi log ( df G i N G N N df N i + 1) when df Ni 6 = 0 tf Gi log ( df G i N G C + 1) when df Ni = 0 p N i = 8 &lt; : tf Ni log ( df N i N N N G df G i + 1) when df Gi 6 = 0 tf Ni log ( df N i N N C + 1) when df Gi = 0 As can be seen from the formulas, the de nitions of these weights were derived from the traditional tf idf measures used in informational retrieval, with some adjustments made for the particular problem at hand. Given a new incoming table, let us denote the set including all the words in it as W n . Since W is constructed using thousands of tables, the words that are present in both W and W n are only a small subset of W . Based on the vector space model, we de ne the similarity between weight vectors representing genuine and non-genuine tables and the frequency vector representing the incoming table as the corresponding dot products. Since we only need to consider the words that are present in both W and W n , we rst compute the e ective word set : W e = W \ W n . Let the words in W e be represented as w m k , where m k k = 1 ::: jW e j , are indexes to the words from set W = f w 1 w 2 ::: w jW j g . we de ne the following vectors: Weight vector representing the genuine table group: ! G S = p G m 1 U p G m 2 U p G m jW e j U ! where U is the cosine normalization term: U = v u u t jW e j X k =1 p G m k p G m k : Weight vector representing the non-genuine table group: ! N S = p N m 1 V p N m 2 V p N m jW e j V ! where V is the cosine normalization term: V = v u u t jW e j X k =1 p N m k p N m k : Frequency vector representing the new incoming table: ! I T = tf Tm 1 tf Tm 2 tf Tm jW e j : Finally, the word group feature is de ned as the ratio of the two dot products: wg = 8 &gt; &gt; &gt; &lt; &gt; &gt; &gt; : ! I T ! G S ! I T ! N S when ! I T ! N S 6 = 0 1 when ! I T ! G S = 0 and ! I T ! N S = 0 10 when ! I T ! G S 6 = 0 and ! I T ! N S = 0 CLASSIFICATION SCHEMES Various classi cation schemes have been widely used in document categorization as well as web information retrieval 13, 8]. For the table detection task, the decision tree classi-er is particularly attractive as our features are highly non-homogeneous . We also experimented with Support Vector Machines (SVM), a relatively new learning approach which has achieved one of the best performances in text categorization 13]. 3.1 Decision Tree Decision tree learning is one of the most widely used and practical methods for inductive inference. It is a method for approximating discrete-valued functions that is robust to noisy data. Decision trees classify an instance by sorting it down the tree from the root to some leaf node, which provides the classi cation of the instance. Each node in a discrete-valued decision tree speci es a test of some attribute of the instance, and each branch descending from that node corresponds to one of the possible values for this attribute. Continuous-valued decision attributes can be incorporated by dynami-cally de ning new discrete-valued attributes that partition the continuous attribute value into a discrete set of intervals 9].Animplementationof thecontinuous-valueddecision tree described in 4] was used for our experiments. The decision tree is constructed using a training set of feature vectors with true class labels. At each node, a discriminant threshold 245 is chosen such that it minimizes an impurity value. The learned discriminant function splits the training subset into two subsets and generates two child nodes. The process is repeated at each newly generated child node until a stopping condition is satis ed, and the node is declared as a terminal node based on a majority vote. The maximum impurity reduction, the maximum depth of the tree, and minimum number of samples are used as stopping conditions. 3.2 SVM Support Vector Machines (SVM) are based on the Structural Risk Management principle from computational learning theory 12]. The idea of structural risk minimization is to nd a hypothesis h for which the lowest true error is guaranteed. The true error of h is the probability that h will make an error on an unseen and randomly selected test example. The SVM method is de ned over a vector space where the goal is to nd a decision surface that best separates the data points in two classes. More precisely, the decision surface by SVM for linearly separable space is a hyperplane which can be written as ~w ~x ; b = 0 where ~ x is an arbitrary data point and the vector ~w and the constant b are learned from training data. Let D = ( y i ~x i ) denote the training set, and y i 2 f +1 ; 1 g be the classi cation for ~x i , the SVM problem is to nd ~w and b that satis es the following constraints: ~w ~x i ; b +1 for y i = +1 ~w ~x i ; b ; 1 for y i = ; 1 while minimizing the vector 2-norm of ~w . The SVM problem in linearly separable cases can be e ciently solved using quadratic programming techniques, while the non-linearly separable cases can be solved by either introducing soft margin hyperplanes, or by mapping the original data vectors to a higher dimensional space where the data points become linearly separable 12, 3]. One reason why SVMs are very powerful is that they are very universal learners. In their basic form, SVMs learn linear threshold functions. Nevertheless, by a simple \plug-in&quot; of an appropriate kernel function, they can be used to learn polynomial classi ers, radial basis function (RBF) networks, three-layer sigmoid neural nets, etc. 3]. For our experiments, we used the SV M light system implemented by Thorsten Joachims. 1 DATA COLLECTION AND TRUTHING Since there are no publicly available web table ground truth database, researchers tested their algorithms in di erent data sets in the past 2, 10, 14]. However, their data sets either had limited manually annotated table data ( e.g. , 918 table tags in 2], 75 HTML pages in 10], 175 manually annotated table tags in 14]), or were collected from some speci c domains ( e.g. , a set of tables selected from airline information pages were used in 2]). To develop our machine learning based table detection algorithm, we needed to build a general web table ground truth database of signi cant size. 1 http://svmlight.joachims.org 4.1 Data Collection Instead of working within a speci c domain, our goal of data collection was to get tables of as many di erent varieties as possible from the web. To accomplish this, we composed a set of key words likely to indicate documents containing tables and used those key words to retrieve and download web pages using the Google search engine. Three directories on Google were searched: the business directory and news directory using key words: ftable, stock, bonds, figure, schedule, weather, score, service, results, valueg , and the science directory using key words ftable, results, valueg . A total of 2,851 web pages were down-loaded in this manner and we ground truthed 1,393 HTML pages out of these (chosen randomly among all the HTML pages). These 1,393 HTML pages from around 200 web sites comprise our database. 4.2 Ground Truthing There has been no previous report on how to systemati-cally generate web table ground truth data. To build a large web table ground truth database, a simple, exible and complete ground truth protocol is required. Figure 4.2(a) shows the diagram of our ground truthing procedure. We created a new Document Type De nition(DTD) which is a super-set of W3C HTML 3.2 DTD. We added three attributes for &lt;TABLE&gt; element, which are \tabid&quot;, \genuine table&quot; and \table title&quot;. The possible value of the second attribute is yes or no and the value of the rst and third attributes is a string. We used these three attributes to record the ground truth of each leaf &lt;TABLE&gt; node. The bene t of this design is that the ground truth data is inside HTML le format. We can use exactly the same parser to process the ground truth data. We developed a graphical user interface for web table ground truthing using the Java 1] language. Figure 4.2(b) is a snapshot of the interface. There are two windows. After reading an HTML le, the hierarchy of the HTML le is shown in the left window. When an item is selected in the hierarchy, the HTML source for the selected item is shown in the right window. There is a panel below the menu bar. The user can use the radio button to select either genuine table or non-genuine table. The text window is used to input table title. 4.3 Database Description Our nal table ground truth database consists of 1,393 HTML pages collected from around 200 web sites. There are a total of 14,609 &lt;TABLE&gt; nodes, including 11,477 leaf &lt;TABLE&gt; nodes. Out of the 11,477 leaf &lt;TABLE&gt; nodes, 1,740 are genuine tables and 9,737 are non-genuine tables. Not every genuine table has its title and only 1,308 genuine tables have table titles. We also found at least 253 HTML les have unmatched &lt;TABLE&gt; , &lt;/TABLE&gt; pairs or wrong hierarchy, which demonstrates the noisy nature of web documents EXPERIMENTS A hold-out method is used to evaluate our table classi-er . We randomly divided the data set into nine parts. Each classi er was trained on eight parts and then tested on the remaining one part. This procedure was repeated nine times, each time with a di erent choice for the test 246 Parser Adding attributes HTML with attributes and unique index to each table(ground truth) Validation HTML File Hierarchy (a) (b) Figure 2: (a) The diagram of ground truthing procedure (b) A snapshot of the ground truthing software. part. Then the combined nine part results are averaged to arrive at the overall performance measures 4]. For the layout and content type features, this procedure is straightforward. However it is more complicated for the word group feature training. To compute w g for training samples, we need to further divide the training set into two groups, a larger one (7 parts) for the computation of the weights p G i and p N i , i = 1 ::: jW j , and a smaller one (1 part) for the computation of the vectors ! G S , ! N S , and ! I T . This partition is again rotated to compute w g for each table in the training set. Table 1: Possible true- and detected-state combinations for two classes. True Class Assigned Class genuine table non-genuine table genuine table N gg N gn non-genuine table N ng N nn The output of each classi er is compared with the ground truth and a contingency table is computed to indicate the number of a particular class label that are identi ed as members of one of two classes. The rows of the contingency table represent the true classes and the columns represent the assigned classes. The cell at row r and column c is the number of tables whose true class is r while its assigned class is c . The possible true- and detected-state combination is shown in Table 1. Three performance measures Recall Rate(R) , Precision Rate(P) and F-measure(F) are computed as follows : R = N gg N gg + N gn P = N gg N gg + N ng F = R + P 2 : For comparison among di erent features and learning algorithms we report the performance measures when the best F-measure is achieved. First, the performance of various feature groups and their combinations were evaluated using the decision tree classi er. The results are given in Table 2. Table 2: Experimental results using various feature groups and the decision tree classi er. L T LT LTW R (%) 87.24 90.80 94.20 94.25 P (%) 88.15 95.70 97.27 97.50 F (%) 87.70 93.25 95.73 95.88 L: La y out only . T: Con ten t t yp e only . L T: La y out and con ten t t yp e. L TW: La y out, con ten t t yp e and w ord group. As seen from the table, content type features performed better than layout features as a single group, achieving an F-measure of 93 : 25%. However, when the two groups were combined the F-measure was improved substantially to 95 : 73%, recon rming the importance of combining layout and content features in table detection. The addition of the word group feature improved the F-measure slightly more to 95 : 88%. Table 3 compares the performances of di erent learning algorithms using the full feature set. The leaning algorithms tested include the decision tree classi er and the SVM al-247 gorithm with two di erent kernels { linear and radial basis function (RBF). Table 3: Experimental results using di erent learning algorithms. Tree SVM (linear) SVM (RBF) R (%) 94.25 93.91 95.98 P (%) 97.50 91.39 95.81 F (%) 95.88 92.65 95.89 As seen from the table, for this application the SVM with radial basis function kernel performed much better than the one with linear kernel. It achieved an F measure of 95 : 89%, comparable to the 95 : 88% achieved by the decision tree classi er. Figure 3 shows two examples of correctly classi ed tables, where Figure 3(a) is a genuine table and Figure 3(b) is a non-genuine table. Figure 4 shows a few examples where our algorithm failed. Figure 4(a) was misclassi ed as a non-genuine table, likely because its cell lengths are highly inconsistent and it has many hyperlinks which is unusual for genuine tables. The reason why Figure 4(b) was misclassi ed as non-genuine is more interesting. When we looked at its HTML source code, we found it contains only two &lt;TR&gt; tags. All text strings in one rectangular box are within one &lt;TD&gt; tag. Its author used &lt;p&gt; tags to put them in di erent rows. This points to the need for a more carefully designed pseudo-rendering process. Figure 4(c) shows a non-genuine table misclassi-ed as genuine. A close examination reveals that it indeed has good consistency along the row direction. In fact, one could even argue that this is indeed a genuine table, with implicit row headers of Title, Name, Company A liation and Phone Number . This example demonstrates one of the most di cult challenges in table understanding, namely the ambiguous nature of many table instances (see 5] for a more detailed analysis on that). Figure 4(d) was also misclassi-ed as a genuine table. This is a case where layout features and the kind of shallow content features we used are not enough | deeper semantic analysis would be needed in order to identify the lack of logical coherence which makes it a non-genuine table. For comparison, we tested the previously developed rule-based system 10] on the same database. The initial results (shown in Table 4 under \Original Rule Based&quot;) were very poor. After carefully studying the results from the initial experiment we realized that most of the errors were caused by a rule imposing a hard limit on cell lengths in genuine tables. After deleting that rule the rule-based system achieved much improved results (shown in Table 4 under \Modi ed Rule Based&quot;). However, the proposed machine learning based method still performs considerably better in comparison. This demonstrates that systems based on handcrafted rules tend to be brittle and do not generalize well. In this case, even after careful manual adjustment in a new database, it still does not work as well as an automatically trained classi er. (a) (b) Figure 3: Examples of correctly classi ed tables. (a): a genuine table (b): a non-genuine table. Table 4: Experimental results of a previously developed rule based system. Original Rule Based Modi ed Rule Based R (%) 48.16 95.80 P (%) 75.70 79.46 F (%) 61.93 87.63 248 (a) (b) (c) (d) Figure 4: Examples of misclassi ed tables. (a) and (b): Genuine tables misclassi ed as non-genuine (c) and (d): Non-genuine tables misclassi ed as genuine. A direct comparison to other previous results 2, 14] is not possible currently because of the lack of access to their system. However, our test database is clearly more general and far larger than the ones used in 2] and 14], while our precision and recall rates are both higher. CONCLUSION AND FUTURE WORK Table detection in web documents is an interesting and challenging problem with many applications. We present a machine learning based table detection algorithm for HTML documents. Layout features, content type features and word group features were used to construct a novel feature set. Decision tree and SVM classi ers were then implemented and tested in this feature space. We also designed a novel table ground truthing protocol and used it to construct a large web table ground truth database for training and testing. Experiments on this large database yielded very promising results. Our future work includes handling more di erent HTML styles in pseudo-rendering, detecting table titles of the rec-ognized genuine tables and developing a machine learning based table interpretation algorithm. We would also like to investigate ways to incorporate deeper language analysis for both table detection and interpretation. ACKNOWLEDGMENT We would like to thank Kathie Shipley for her help in collecting the web pages, and Amit Bagga for discussions on vector space models. REFERENCES 1] M. Campione, K. Walrath, and A. Huml. The java(tm) tutorial: A short course on the basics (the java(tm) series). 2] H.-H. Chen, S.-C. Tsai, and J.-H. Tsai. Mining tables from large scale html texts. In Proc. 18th International Conference on Computational Linguistics , Saabrucken, Germany, July 2000. 3] C. Cortes and V. Vapnik. Support-vector networks. Machine Learning , 20:273{296, August 1995. 4] R. Haralick and L. Shapiro. Computer and Robot Vision , volume 1. Addison Wesley, 1992. 5] J. Hu, R. Kashi, D. Lopresti, G. Nagy, and G. Wilfong. Why table ground-truthing is hard. In Proc. 6th International Conference on Document Analysis and Recognition (ICDAR01) , pages 129{133, Seattle, WA, USA, September 2001. 6] M. Hurst. Layout and language: Challenges for table understanding on the web. In Proc. 1st International Workshop on Web Document Analysis , pages 27{30, Seattle, WA, USA, September 2001. 7] T. Joachims. A probabilistic analysis of the rocchio algorithm with t df for text categorization. In Proc. 14th International Conference on Machine Learning , pages 143{151, Morgan Kaufmann, 1997. 8] A. McCallum, K. Nigam, J. Rennie, and K. Seymore. Automating the construction of internet portals with machine learning. In Information Retrieval Journal , volume 3, pages 127{163, Kluwer, 2000. 249 9] T. M. Mitchell. Machine Learning . McGraw-Hill, 1997. 10] G. Penn, J. Hu, H. Luo, and R. McDonald. Flexible web document analysis for delivery to narrow-bandwidth devices. In Proc. 6th International Conference on Document Analysis and Recognition (ICDAR01) , pages 1074{1078, Seattle, WA, USA, September 2001. 11] M. F. Porter. An algorithm for su x stripping. Program , 14(3):130{137, 1980. 12] V. N. Vapnik. The Nature of Statistical Learning Theory , volume 1. Springer, New York, 1995. 13] Y. Yang and X. Liu. A re-examination of text categorization methods. In Proc. SIGIR'99 , pages 42{49, Berkeley, California, USA, August 1999. 14] M. Yoshida, K. Torisawa, and J. Tsujii. A method to integrate tables of the world wide web. In Proc. 1st International Workshop on Web Document Analysis , pages 31{34, Seattle, WA, USA, September 2001. 250

Table detection;table ground truthing protocol;Layout Analysis;classifers;word group;presentation;Information Retrieval;Algorithms;Support Vector Machine;classifcation schemes;Machine Learning;Table Detection;Layout;machine learning based approach;content type;Decision tree;HTML document

130

Low Latency Photon Mapping Using Block Hashing

For hardware accelerated rendering, photon mapping is especially useful for simulating caustic lighting effects on non-Lambertian surfaces. However, an efficient hardware algorithm for the computation of the k nearest neighbours to a sample point is required. Existing algorithms are often based on recursive spatial subdivision techniques, such as kd-trees. However, hardware implementation of a tree-based algorithm would have a high latency, or would require a large cache to avoid this latency on average. We present a neighbourhood-preserving hashing algorithm that is low-latency and has sub-linear access time. This algorithm is more amenable to fine-scale parallelism than tree-based recursive spatial subdivision, and maps well onto coherent block-oriented pipelined memory access. These properties make the algorithm suitable for implementation using future programmable fragment shaders with only one stage of dependent texturing.

Introduction Photon mapping, as described by Jensen , is a technique for reconstructing the incoming light field at surfaces everywhere in a scene from sparse samples generated by forward light path tracing. In conjunction with path tracing, photon mapping can be used to accelerate the computation of both diffuse and specular global illumination . It is most effective for specular or glossy reflectance effects, such as caustics . The benefits of migrating photo-realistic rendering techniques towards a real-time, hardware-assisted implementation are obvious. Recent work has shown that it is possible to implement complex algorithms, such as ray-tracing, using the programmable features of general-purpose hardware accelerators and/or specialised hardware . We are interested in hardware support for photon mapping: specifically , the application of photon maps to the direct visualisation of caustics on non-Lambertian surfaces, since diffuse global illumination effects are probably best handled in a real-time renderer using alternative techniques such as irradiance . Central to photon mapping is the search for the set of photons nearest to the point being shaded. This is part of the interpolation step that joins light paths propagated from light sources with rays traced from the camera during rendering, and it is but one application of the well-studied k nearest neighbours (kNN) problem. Jensen uses the kd-tree , data structure to find these nearest photons. However, solving the kNN problem via kd-trees requires a search that traverses the tree. Even if the tree is stored as a heap, traversal still requires random-order memory access and memory to store a stack. More importantly, a search-path pruning algorithm, based on the data already examined, is required to avoid accessing all data in the tree. This introduces serial dependencies between one memory lookup and the next. Consequently, a hardware implementation of a kd-tree-based kNN solution would either have high latency, or would require a large cache to avoid such latency. In either case a custom hardware implementation would be required. These properties motivated us to look at alternatives to tree search. Since photon mapping is already making an approximation by using kNN interpolation, we conjectured that an approximate kNN (AkNN) solution should suffice so long as visual quality is maintained. In this paper we investigate a hashing-based AkNN solution in the context of high-c The Eurographics Association 2002. Ma and McCool / Low Latency Photon Mapping Using Block Hashing performance hardware-based (specifically, programmable shader-based) photon mapping. Our major contribution is an AkNN algorithm that has bounded query time, bounded memory usage, and high potential for fine-scale parallelism. Moreover, our algorithm results in coherent, non-redundant accesses to block-oriented memory. The results of one memory lookup do not affect subsequent memory lookups, so accesses can take place in parallel within a pipelined memory system. Our algorithm is based on array access, and is more compatible with current texture-mapping capabilities than tree-based algorithms. Furthermore, any photon mapping acceleration technique that continues to rely on a form of kNN (such as irradiance caching ) can still benefit from our technique. In Section 2 , we first review previous work on the kNN and the approximate k-nearest neighbour (AkNN) problems. Section 3 describes the context and assumptions of our research and illustrates the basic hashing technique used in our algorithm. Sections 4 and 5 describe the details of our algorithm. Section 6 presents numerical, visual quality, and performance results. Section 7 discusses the mapping of the algorithm onto a shader-based implementation. Finally, we conclude in Section 8 . Previous Work Jensen's book 25 covers essentially all relevant previous work leading up to photon mapping. Due to space limitations, we will refer the reader to that book and focus our literature review on previous approaches to the kNN and AkNN problems . Any non-trivial algorithm that claims to be able to solve the kNN problem faster than brute-force does so by reducing the number of candidates that have to be examined when computing the solution set. Algorithms fall into the following categories: recursive spatial subdivision, point location, neighbourhood graphs, and hashing. Amongst algorithms based on recursive spatial subdivision , the kd-tree 5 method is the approach commonly used to solve the kNN problem 14 . An advantage of the kd-tree is that if the tree is balanced it can be stored as a heap in a single array. While it has been shown that kd-trees have optimal expected-time complexity 6 , in the worst case finding the k nearest neighbours may require an exhaustive search of the entire data structure via recursive decent. This requires a stack the same size as the depth of the tree. During the recursion , a choice is made of which subtree to search next based on a test at each internal node. This introduces a dependency between one memory access and the next and makes it hard to map this algorithm into high-latency pipelined memory accesses. Much work has been done to find methods to optimise the kd-tree method of solving the kNN and AkNN problems. See Christensen 26 , Vanco et al. 44 , Havran 19 , and Sample et al. 39 . Many other recursive subdivision-based techniques have also been proposed for the kNN and AkNN problems, including kd-B-trees 36 , BBD-trees 4 , BAR-trees 9 , Principal-Axis Trees 33 , the R-tree family of data structures 27 , and ANN-trees 30 . Unfortunately, all schemes based on recursive search over a tree share the same memory dependency problem as the kd-tree. The second category of techniques are based on building and searching graphs that encode sample-adjacency information . The randomised neighbourhood graph approach 3 builds and searches an approximate local neighbourhood graph. Eppstein et al. 11 investigated the fundamental properties of a nearest neighbour graph. Jaromczyk and Toussaint surveyed data structures and techniques based on Relative Neighbourhood Graphs 23 . Graph-based techniques tend to have the same difficulties as tree-based approaches: searching a graph also involves stacks or queues, dependent memory accesses, and pointer-chasing unsuited to high-latency pipelined memory access. Voronoi diagrams can be used for optimal 1-nearest neighbour searches in 2D and 3D 10 . This and other point-location based techniques 17 for solving nearest neighbour problems do not need to calculate distances between the query point and the candidates, but do need another data structure (like a BSP tree) to test a query point for inclusion in a region. Hashing approaches to the kNN and AkNN problems have recently been proposed by Indyk et al. 20 , 21 and Gionis et al. 16 . These techniques have the useful property that multi-level dependent memory lookups are not required. The heart of these algorithms are simple hash functions that preserve spatial locality, such as the one proposed by Linial and Sasson 31 , and Gionis et al. 16 . We base our technique on the latter. The authors also recognise recent work by Wald et al. 45 on real-time global illumination techniques where a hashing-based photon mapping technique was apparently used (but not described in detail). Numerous surveys and books 1 , 2 , 15 , 42 , 43 , 17 provide further information on this family of problems and data structures developed to solve them. Context We have developed a novel technique called Block Hashing (BH) to solve the approximate kNN (AkNN) problem in the context of, but not limited to, photon mapping. Our algorithm uses hash functions to categorise photons by their positions. Then, a kNN query proceeds by deciding which hash bucket is matched to the query point and retrieving the photons contained inside the hash bucket for analysis . One attraction of the hashing approach is that evaluation of hash functions takes constant time. In addition, once we have the hash value, accessing data we want in the hash table takes only a single access. These advantages permit us to c The Eurographics Association 2002. 90 Ma and McCool / Low Latency Photon Mapping Using Block Hashing avoid operations that are serially dependent on one another, such as those required by kd-trees, and are major stepping stones towards a low-latency shader-based implementation. Our technique is designed under two assumptions on the behaviour of memory systems in (future) accelerators. First, we assume that memory is allocated in fixed-sized blocks . Second, we assume that access to memory is via burst transfer of blocks that are then cached. Under this assumption, if any part of a fixed-sized memory block is "touched", access to the rest of this block will be virtually zero-cost. This is typical even of software implementations on modern machines which rely heavily on caching and burst-mode transfers from SDRAM or RDRAM. In a hardware implementation with a greater disparity between processing power and memory speed, using fast block transfers and caching is even more important. Due to these benefits, in BH all memory used to store photon data is broken into fixed-sized blocks. 3.1. Locality-Sensitive Hashing Since our goal is to solve the kNN problem as efficiently as possible in a block-oriented cache-based context, our hashing technique requires hash functions that preserve spatial neighbourhoods. These hash functions take points that are close to each other in the domain space and hash them close to each other in hash space. By using such hash functions, photons within the same hash bucket as a query point can be assumed to be close to the query point in the original domain space. Consequently, these photons are good candidates for the kNN search. More than one such scheme is available; we chose to base our technique on the Locality-Sensitive Hashing (LSH) algorithm proposed by Gionis et al. 16 , but have added several refinements (which we describe in Section 4 ). The hash function in LSH groups one-dimensional real numbers in hash space by their spatial location. It does so by partitioning the domain space and assigning a unique hash value to each partition. Mathematically, let T = {t i | 0 i P} be a monotonically increasing sequence of P + 1 thresholds between 0 and 1. Assume t 0 = 0 and t P = 1, so there are P - 1 degrees of freedom in this sequence . Define a one-dimensional Locality-Sensitive Hash Function h T : [0, 1] {0 . . . P - 1} to be h T (t) = i, where t i t &lt; t i+1 . In other words, the hash value i can take on P different values, one for each "bucket" defined by the threshold pair [t i ,t i+1 ). An example is shown in Figure 1 . t 1 t 2 t 3 t 4 t 0 Figure 1: An example of h T . The circles and boxes represent values to be hashed, while the vertical lines are the thresholds in T . The boxes lie between t 1 and t 2 , thus they are given a hash value of 1. The function h T can be interpreted as a monotonic non-uniform quantisation of spatial position, and is characterised by P and the sequence T . It is important to note that h T gives each partition of the domain space delineated by T equal representation in hash space. Depending on the location of the thresholds, h T will contract some parts of the domain space and expand other parts. If we rely on only a single hash table to classify a data set, a query point will only hash to a single bucket within this table, and the bucket may represent only a subset of the true neighbourhood we sought. Therefore, multiple hash tables with different thresholds are necessary for the retrieval of a more complete neighbourhood around the query point (See Figure 2 .) h T1 h T3 h T2 t 1 t 2 t 3 t 4 t 0 t 4 t 0 t 4 t 0 t 1 t 1 t 2 t 2 t 3 t 3 Figure 2: An example of multiple hash functions classifying a dataset. The long vertical line represents the query value. The union of results multiple hash tables with different thresholds represents a more complete neighbourhood. To deal with n-dimensional points, each hash table will have one hash function per dimension. Each hash function generates one hash value per coordinate of the point (See Figure 3 .) The final hash value is calculated by n -1 i=0 h i P i , where h i are the hash values and P is the number of thresholds . If P were a power of two, then this amounts to concatenating the bits. The only reason we use the same number of thresholds for each dimension is simplicity. It is conceivable that a different number of thresholds could be used for each dimension to better adapt to the data. We defer the discussion of threshold generation and query procedures until Sections 4.2 and 4.4 , respectively. h x h y P Figure 3: Using two hash functions to handle a 2D point. Each hash function will be used to hash one coordinate. LSH is very similar to grid files 34 . However, the grid file was specifically designed to handle dynamic data. Here, we assume that the data is static during the rendering pass. Also, the grid file is more suitable for range searches than it is for solving the kNN problem. 3.2. Block-Oriented Memory Model It has been our philosophy that hardware implementations of algorithms should treat off-chip memory the same way software implementations treat disk: as a relatively slow, "out-of -core", high-latency, block-oriented storage device. This c The Eurographics Association 2002. 91 Ma and McCool / Low Latency Photon Mapping Using Block Hashing analogy implies that algorithms and data structures designed to optimise for disk access are potentially applicable to hardware design. It also drove us to employ fixed-sized blocks to store the data involved in the kNN search algorithm, which are photons in the context of this application. In our prototype software implementation of BH, each photon is stored in a structure similar to Jensen's "extended" photon representation 25 . As shown in Figure 4 , each component of the 3D photon location is represented by a 32-bit fixed-point number. The unit vectors representing incoming direction ^d and surface normal ^n are quantised to 16-bit values using Jensen's polar representation. Photon power is stored in four channels using sixteen-bit floating point numbers . This medium-precision signed representation permits other AkNN applications beyond that of photon mapping. Four colour channels are also included to better match the four-vectors supported in fragment shaders. For the photon mapping application specifically, our technique is compatible with the replacement of the four colour channels with a Ward RGBE colour representation 46 . Likewise, another implementation might use a different representation for the normal and incident direction unit vectors. | 32 | x y z ^ d ^ n c 1 c 2 c 3 c 4 | 16 | Figure 4: Representation of a photon record. The 32-bit values (x, y, z) denote the position of a photon and are used as keys. Two quantised 16-bit unit vectors ^d, ^n and four 16-bit floating point values are carried as data. All photon records are stored in fixed-sized memory blocks. BH uses a 64 32-bit-word block size, chosen to permit efficient burst-mode transfers over a wide bus to transaction-oriented DRAM. Using a 128-bit wide path to DDR SDRAM, for instance, transfer of this block would take eight cycles, not counting the overhead of command cycles to specify the operation and the address. Using next-generation QDR SDRAM this transfer would take only four cycles (or eight on a 64-bit bus, etc.) Our photon representation occupies six 32-bit words. Since photon records are not permitted to span block boundaries , ten photons will fit into a 64-word block with four words left over. Some of this extra space is used in our implementation to record how many photons are actually stored in each block. For some variants of the data structures we describe , this extra space could also be used for flags or pointers to other blocks. It might be possible or desirable in other implementations to support more or fewer colour channels, or channels with greater or lesser precision, in which case some of our numerical results would change. Block Hashing Block Hashing (BH) contains a preprocessing phase and a query phase. The preprocessing phase consists of three steps. After photons have been traced into the scene, the algorithm organises the photons into fixed-sized memory blocks, creates a set of hash tables, and inserts photon blocks into the hash tables. In the second phase, the hash tables will be queried for a set of candidate photons from which the k nearest photons will be selected for each point in space to be shaded by the renderer. 4.1. Organizing Photons into Blocks Due to the coherence benefits associated with block-oriented memory access, BH starts by grouping photons and storing them into fixed-sized memory blocks. However, these benefits are maximised when the photons within a group are close together spatially. We chose to use the Hilbert curve 13 to help group photons together. The advantage of the Hilbert curve encoding of position is that points mapped near each other on the Hilbert curve are guaranteed to be within a certain distance of each other in the original domain 22 . Points nearby in the original domain space have a high probability of being nearby on the curve, although there is a non-zero probability of them being far apart on the curve. If we sort photons by their Hilbert curve order before packing them into blocks, then the photons in each block will have a high probability of being spatially coherent. Each block then corresponds to an interval of the Hilbert curve, which in turn covers some compact region of the domain (see Figure 7 a). Each region of domain space represented by the blocks is independent, and regions do not overlap. BH sorts photons and inserts them into a B + -tree 8 using the Hilbert curve encoding of the position of each photon as the key. This method of spatially grouping points was first proposed by Faloutsos and Rong 12 for a different purpose. Since a B + -tree stores photon records only at leaves, with a compatible construction the leaf nodes of the B + -tree can serve as the photon blocks used in the later stages of BH. One advantage of using a B + -tree for sorting is that insertion cost is bounded: the tree is always balanced, and in the worst case we may have to split h nodes in the tree, when the height of the tree is h. Also, B + -trees are optimised for block-oriented storage, as we have assumed. The B + -tree used by BH has index and leaf nodes that are between half full to completely full. To minimise the final number of blocks required to store the photons, the leaf nodes can be compacted (see Figure 5 .) After the photons are sorted and compacted, the resulting photon blocks are ready to be used by BH, and the B + -tree index and any leaf nodes that are made empty by compaction are discarded. If the complete set of photons is known a priori, the compact B + -tree 37 for static data can be used instead. This data structure maintains full nodes and avoids the extra compaction step. c The Eurographics Association 2002. 92 Ma and McCool / Low Latency Photon Mapping Using Block Hashing (b) Index node Empty cell in leaf node Occupied cell in leaf node (a) Figure 5: Compaction of photon blocks. (a) B + -tree after inserting all photons. Many leaf nodes have empty cells. (b) All photon records are compacted in the leaf nodes. Regardless, observe that each photon block contains a spatially clustered set of photons disjoint from those contained in other blocks. This is the main result we are after ; any other data structures that can group photons into spatially-coherent groups, such as grid files 34 , can be used in place of the B + -tree and space-filling curve. 4.2. Creating the Hash Tables The hash tables used in BH are based on the LSH scheme described in Section 3.1 . BH generates L tables in total, serving as parallel and complementary indices to the photon data. Each table has three hash functions (since photons are classified by their 3D positions), and each hash function has P + 1 thresholds. BH employs an adaptive method that generates the thresholds based on the photon positions. For each dimension, a histogram of photon positions is built. Then, the histogram is integrated to obtain a cumulative distribution function (cdf ). Lastly, stratified samples are taken from the inverse of the cdf to obtain threshold locations. The resulting thresholds will be far apart where there are few photons, and close together where photons are numerous. Ultimately this method attempts to have a similar number of photons into each bucket. Hash tables are stored as a one-dimensional array structure , shown in Figure 6 . The hash key selects a bucket out of the P n available buckets in the hash table. Each bucket refers up to B blocks of photons, and has space for a validity flag per reference, and storage for a priority value. We defer the discussion on the choice of P, L and B until Section 5 . B V V V V Priority Figure 6: Hash table bucket layout 4.3. Inserting Photon Blocks In BH, references to entire photon blocks, rather than individual photons, are inserted into the hash tables. One reason for doing so is to reduce the memory required per bucket. Another, more important, reason is that when merging results from multiple hash tables (Section 3.1 ), BH needs to compare only block addresses instead of photons when weeding out duplicates as each block contains a unique set of photons. This means fewer comparisons have to be made and the individual photons are only accessed once per query, during post-processing of the merged candidate set to find the k nearest photons. Consequently, the transfer of each photon block through the memory system happens at most once per query. All photons accessed in a block are potentially useful contributions to the candidate set, since photons within a single block are spatially coherent. Due to our memory model assumptions, once we have looked at one photon in a block it should be relatively inexpensive to look at the rest. (a) (b) (c) Figure 7: Block hashing illustrated. (a) Each block corresponds to an interval of the Hilbert curve, which in turn covers some compact region of the domain. Consequently, each bucket (b) represents all photons (highlighted with squares) in each block with at least one photon hashed into it (c). Each bucket in a hash table corresponds to a rectangular region in a non-uniform grid as shown in Figure 7 b. Each block is inserted into the hash tables once for each photon within that block, using the position of these photons to create the keys. Each bucket of the hash table refers to not only the photons that have been hashed into that bucket, but also all the other photons that belong to the same blocks as the hashed photons (see Figure 7 c.) Since each photon block is inserted into each hash table multiple times, using different photons as keys, a block may be hashed into multiple buckets in the same hash table . Of course, a block should not be inserted into a bucket more than once. More importantly, our technique ensures that each block is inserted into at least one hash table. Orphaned blocks are very undesirable since the photons within will never be considered in the subsequent AkNN evaluation and will cause a constant error overhead. Hence, our technique does not navely drop a block that causes a bucket to overflow. However, there may be collisions that cause buckets to overflow, especially when a large bucket load factor is chosen to give a compact hash table size, or there exists a large variation in photon density (which, of course, is typical in this application). Our algorithm uses two techniques to address this problem. The first technique attempts to insert every block into every hash table, but in different orders on different hash tables, such that blocks that appear earlier in the ordering are not favoured for insertion in all tables. BH uses a technique similar to disk-striping 38 , illustrated by the c The Eurographics Association 2002. 93 Ma and McCool / Low Latency Photon Mapping Using Block Hashing pseudo code in Figure 8 . An example is given in the diagram in the same figure. for h from 0 to (number_of_hash_tables-1) for b from 0 to (number_of_blocks-1) idx = (h+b) modulo L insert block[b] into hashtable[idx] endfor endfor 0 1 2 0 1 2 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 3 4 5 3 4 5 3 4 5 Photon Block Hash Table Bucket 1st iteration 2nd iteration 3rd iteration Figure 8: Striping insertion strategy The second technique involves a strategy to deal with overflow in a bucket. For each photon block, BH keeps the count of buckets that the block has been hashed into so far. When a block causes overflow in a bucket, the block in the bucket that has the maximum count will be bumped if that count is larger than one, and larger than that of the incoming block. This way we ensure that all blocks are inserted into at least one bucket, given adequate hash table sizes, and no block is hashed into an excessive number of buckets. 4.4. Querying A query into the BH data structure proceeds by delegating the query to each of the L hash tables. These parallel accesses will yield as candidates all photon blocks represented by buckets that matched the query. The final approximate nearest neighbour set comes from scanning the unified candidate set for the nearest neighbours to the query point (see Figure 9 .) Note that unlike kNN algorithms based on hier-archical data structures, where candidates for the kNN set trickle in as the traversal progresses, in BH all candidates are available once the parallel queries are completed. Therefore, BH can use algorithms like selection 29 (instead of a priority queue) when selecting the k nearest photons. Each query will retrieve one bucket from each of the L hash tables. If the application can tolerate elevated inaccuracy in return for increased speed of query (for example, to pre-visualise a software rendering), it may be worthwhile to consider using only a subset of the L candidate sets. Block hashing is equipped with a user-specified accuracy setting: Let A IN be an integer multiplier. The block hashing algorithm will only consider Ak candidate photons in the final scan to determine the k nearest photons to a query. Obviously the smaller A is, the fewer photons will be processed in the final step; as such, query time is significantly reduced, but with an accuracy cost. Conversely, a higher A will lead to a more accurate result, but it will take more time. Experimental results that demonstrate the impact of the choice of A will be explored in Section 6 . Query point Matched point Data point (a) (b) (c) Figure 9: Merging the results from multiple hash tables. (a) the query point retrieves different candidates sets from different hash tables, (b) the union set of candidates after merging, and (c) the two closest neighbours selected. There needs to be a way to select the buckets from which the Ak candidate photons are obtained. Obviously, we want to devise a heuristic to pick the "best" candidates. Suppose every bucket in every hash table has a priority given by = |bucket_capacity - #entries - #overflows| where "#overflows" is the number of insertion attempts after the bucket became full. The priority can be pre-computed and stored in each bucket of each hash table during the insertion phase. The priority of a bucket is smallest when the bucket is full but never overflowed. Conversely, when the hash bucket is underutilised or overflow has occurred, will be larger. If a bucket is underutilised, it is probably too small spatially (relative to local sample density). If it has experienced overflow, it is likely too large spatially, and covers too many photon block regions. During each query, BH will sort the L buckets returned from the hash tables by their priority values, smallest values of first. Subsequently, buckets are considered in this order, one by one, until the required Ak photons are found. In this way the more "useful" buckets will be considered first. Choice of Parameter Values Block Hashing is a scheme that requires several parameters: B, the bucket capacity; L, the number of hash tables whose results are merged; and P, the number of thresholds per dimension . We would like to determine reasonable values for these parameters as functions of k, the number of nearest neighbours sought, and N, the number of photons involved. It is important to realize the implications of these parameters . The total number of 32-bit pointers to photon blocks is given by LP 3 B. Along with the number of thresholds 3LP, this gives the memory overhead required for BH. The upper bound for this value is 6N, the number of photons multiplied by the six 32-bit words each photon takes up in our implementation. If we allow B to be a fixed constant for now, c The Eurographics Association 2002. 94 Ma and McCool / Low Latency Photon Mapping Using Block Hashing the constraint LP 3 + 3LP N arises from the reasonable assumption that we do not want to have more references to blocks than there are photons, or more memory used in the index than in the data. Empirically, L = P = ln N has turned out to be a good choice. The value ln N remains sub-linear as N increases, and this value gives a satisfactory index memory overhead ratio: There are a total of B(ln N) 4 block references. Being four bytes each, the references require 4B(ln N) 4 bytes. With each hash table, there needs to be 3LP = 3(ln N) 2 thresholds . Represented by a 4-byte value each, the thresholds take another 12(ln N) 2 bytes. Next, assuming one photon block can hold ten photons, N photons requires N/10 blocks; each block requires 64 words, so the blocks require 25.6N bytes in total. The total memory required for N photons, each occupying 6 words, is 24N bytes. This gives an overhead ratio of (4B(ln N) 4 + 12(ln N) 2 + 25.6N - 24N)/24N. (1) The choice of B is also dependent on the value of k specified by the situation or the user. However, since it is usual in photon mapping that k is known ahead of time, B can be set accordingly. B should be set such that the total number of photons retrieved from the L buckets for each query will be larger than k. Mathematically speaking, each photon block in our algorithm has ten photons, hence 10LB k. In particular , 10LB &gt; Ak should also be satisfied. Since we choose L = ln N, rearranging the equation yields: B &gt; Ak/(10 ln N) For example, assuming A = 16, N = 2000000, k = 50, then B = 6. If we substitute B back into Equation 1 , we obtain the final overhead equation (4(Ak/10)(ln N) 3 + 12(ln N) 2 + 1.6N)/24N. (2) Figure 10 plots the number of photons versus the memory overhead. For the usual range of photon count in a photon mapping application, we see that the memory overhead, while relative large for small numbers of photons, becomes reasonable for larger numbers of photons, and has an asymptote of about 6%. Of course, if we assumed different block size (cache line size), these results would vary, but the analysis is the same. Memory Overhead Ratio (%) 5 15 20 25 30 0.50 0.75 1.00 1.25 1.50 1.75 A=16 A=8 A=4 10 0.25 2.00 Number of Photons Figure 10: Plot of photon count vs. memory overhead in-curred by BH, assuming k = 50. Results For BH to be a useful AkNN algorithm, it must have satisfactory algorithmic accuracy. Moreover, in the context of photon mapping, BH must also produce good visual accuracy . This section will demonstrate that BH satisfies both requirements , while providing a speed-up in terms of the time it takes to render an image, even in a software implementation . To measure algorithmic accuracy, our renderer was rigged to use both the kd-tree and BH based photon maps. For each kNN query the result sets were compared for the following metrics: False negatives: # photons incorrectly excluded from the kNN set. Maximum distance dilation: the ratio of bounding radii around the neighbours reported by the two algorithms. Average distance dilation: the ratio of the average distances between the query point and each of the nearest neighbours reported by the two algorithms. To gauge visual accuracy, we calculate a Caustic RMS Error metric, which compares the screen space radiance difference between the caustic radiance values obtained from kd-tree and BH. A timing-related Time Ratio metric is calculated as a ratio of the time taken for a query into the BH data structure versus that for the kd-tree data structure. Obviously, as A increases, the time required for photon mapping using BH approaches that for a kd-tree based photon mapping. Our first test scene, shown in Figure 11 , with numerical results in Figure 12 , consists of a highly specular ring placed on top of a plane with a Modified Phong 28 reflectance model. This scene tests the ability of BH to handle a caustic of varying density, and a caustic that has been cast onto a non-Lambertian surface. Figure 13 shows a second scene consisting of the Venus bust, with a highly specular ring placed around the neck of Venus. Figure 14 shows the numerical statistics of this test scene. The ring casts a cardioid caustic onto the (non-planar) chest of the Venus. This scene demonstrates a caustic on a highly tessellated curved surface. Global illumination is also used for both scenes, however the times given are only for the query time of the caustic maps. The general trend to notice is that for extremely low accuracy (A) settings the visual and algorithmic performance of BH is not very good. The candidate set is simply not large enough in these cases. However, as A increases, these performance indicators drop to acceptable levels very quickly, especially between values of A between 2 and 8. After A = 8 diminishing returns set in, and the increase in accuracy incurs a significant increase in the cost of the query time required . This numerical error comparison is parallelled by the c The Eurographics Association 2002. 95 Ma and McCool / Low Latency Photon Mapping Using Block Hashing (a) kd-tree (b) BH, A=4 (c) BH, A=16 (d) BH, A=8 Figure 11: "Ring" image 0 0.2 0.4 0.6 0.8 1 0 2 4 6 8 10 12 14 16 18 20 Accuracy Setting (A) False-Average #errors 1 1.1 1.2 1.3 1.4 1.5 1.6 0 2 4 6 8 10 12 14 16 18 20 Accuracy Setting (A) Max Radius Dilation Avg Radius Dilation Dilation Ratio 0 0.005 0.01 0.015 0.02 0 2 4 6 8 10 12 14 16 18 20 Accuracy Setting (A) RMS error Radiance RMS Error 0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 2 4 6 8 10 12 14 16 18 20 Accuracy Setting (A) Timing Ratio Time Ratio Figure 12: "Ring" numerical statistics visual comparison of the images: the image rendered with A = 8 and A = 16 are virtually indistinguishable. These results suggest that intermediate values of A, between 8 to 10, should be used as a good compromise between query speed and solution accuracy. It is apparent from the query time ratio plots that there exists a close-to-linear relationship between values of A and time required for a query into BH. This is consistent with the design of the A parameter; it corresponds directly to the number of photons accessed and processed for each query. Another important observation that can be made from the visual comparisons is that images with greater approximation value A look darker. This is because the density estimate is based on the inverse square of the radius of the sphere enclosing the k nearest neighbours. The approximate radius is always larger than the true radius. This is an inherent problem with any approximate solution to the kNN problem, and indeed even with the exact kNN density estimator: as k goes (a) kd-tree (b) BH, A=16 (c) BH, A=8 (d) BH, A=4 Figure 13: "Venus with Ring" images 0 0.5 1 1.5 2 2.5 3 3.5 4 0 2 4 6 8 10 12 14 16 18 20 Accuracy Setting (A) False-Average #errors 1 1.02 1.04 1.06 1.08 1.1 0 2 4 6 8 10 12 14 16 18 20 Accuracy Setting (A) Max Radius Dilation Avg Radius Dilation Dilation Ratio 0 0.01 0.02 0.03 0.04 0.05 0 2 4 6 8 10 12 14 16 18 20 Accuracy Setting (A) RMS error Radiance RMS Error 0 0.2 0.4 0.6 0.8 1 1.2 0 2 4 6 8 10 12 14 16 18 20 Accuracy Setting (A) Timing Ratio Time Ratio Figure 14: "Venus with Ring" numerical statistics to infinity, the kNN density estimator does converge on the true density, but always from below. Hardware Implementation There are two ways to approach a hardware implementation of an algorithm in a hardware accelerator: with a custom hardware implementation, or as an extension or exploitation of current hardware support. While there would be certain advantages in a full custom hardware implementation of BH, this would probably lead to a large chunk of hardware with low utilisation rates. Although there are many potential applications of AkNN search beyond photon mapping (we list several in the conclusions), it seems more reasonable to consider first if BH can be implemented using current hardware and programmable shader features, and if not, what the smallest incremental changes would be. We have concluded that BH, while not quite implementable on today's graphics hardware, should be implementable in the near future. We consider only the lookup phase here, since the preprocessing would indeed require some custom hardware support , but support which perhaps could be shared with other useful features. In the lookup phase, (1) we compute hash c The Eurographics Association 2002. 96 Ma and McCool / Low Latency Photon Mapping Using Block Hashing keys, (2) look up buckets in multiple hash tables, (3) merge and remove duplicates from the list of retrieved blocks and optionally sorting by priority, (4) retrieve the photon records stored in these blocks, and (5) process the photons. Steps (1) and (5) could be performed with current shader capabilities , although the ability to loop would be useful for the last step to avoid replicating the code to process each photon. Computing the hash function amounts to doing a number of comparisons, then adding up the zero-one results. This can be done in linear time with a relatively small number of instructions using the proposed DX9 fragment shader instruction set. If conditional assignment and array lookups into the register file are supported, this could be done in logarithmic time using binary search. Steps (2) and (4) amount to table lookups and can be implemented as nearest-neighbour texture-mapping operations with suitable encoding of the data. For instance, the hash tables might be supported with one texture map giving the priority and number of valid entries in the bucket, while another texture map or set of texture maps might give the block references, represented by texture coordinates pointing into another set of texture maps holding the photon records. Step (3) is difficult to do efficiently without true conditionals and conditional looping. Sorting is not the problem, as it could be done with conditional assignment. The problem is that removal of a duplicate block reduces the number of blocks in the candidate set. We would like in this case to avoid making redundant photon lookups and executing the instructions to process them. Without true conditionals, an inefficient work-around is to make these redundant texture accesses and process the redundant photons anyhow, but discard their contribution by multiplying them by zero. We have not yet attempted an implementation of BH on an actual accelerator. Without looping, current accelerators do not permit nearly enough instructions in shaders to process k photons for adequate density estimation. However, we feel that it might be feasible to implement our algorithm on a next-generation shader unit using the "multiply redundant photons by zero" approach, if looping (a constant number of times) were supported at the fragment level. We expect that the generation that follows DX9-class accelerators will probably have true conditional execution and looping, in which case implementation of BH will be both straightforward and efficient, and will not require any additional hardware or special shader instructions. It will also only require two stages of conditional texture lookup, and lookups in each stage can be performed in parallel. In comparison , it would be nearly impossible to implement a tree-based search algorithm on said hardware due to the lack of a stack and the large number of dependent lookups that would be required. With a sufficiently powerful shading unit, of course, we could implement any algorithm we wanted, but BH makes fewer demands than a tree-based algorithm does. Conclusion and Future Work We have presented an efficient, scalable, coherent and highly parallelisable AkNN scheme suitable for the high-performance implementation of photon mapping. The coherent memory access patterns of BH lead to im-proved performance even for a software implementation. However, in the near future we plan to implement the lookup phase of BH on an accelerator. Accelerator capabilities are not quite to the point where they can support this algorithm, but they are very close. What is missing is some form of true run-time conditional execution and looping, as well as greater capacity in terms of numbers of instructions. However , unlike tree-based algorithms, block hashing requires only bounded execution time and memory. An accelerator-based implementation would be most interesting if it is done in a way that permits other applications to make use of the fast AkNN capability it would provide. AkNN has many potential applications in graphics beyond photon maps. For rendering, it could also be used for sparse data interpolation (with many applications: visualisation of sparse volume data, BRDF and irradiance volume representation , and other sampled functions), sparse and multi-resolution textures, procedural texture generation (specifically , Worley's texture 47 functions), direct ray-tracing of point-based objects 40 , and gap-filling in forward projection point-based rendering 48 . AkNN could also potentially be used for non-rendering purposes: collision detection, surface reconstruction, and physical simulation (for interacting particle systems). Unlike the case with tree-based algorithms , we feel that it will be feasible to implement BH as a shader subroutine in the near future, which may make it a key component in many potential applications of advanced programmable graphics accelerators. For a more detailed description of the block hashing algorithm , please refer to the author's technical report 32 . Acknowledgements This research was funded by grants from the National Science and Engineering Research Council of Canada (NSERC), the Centre for Information Technology of Ontario (CITO), the Canadian Foundation for Innovation (CFI), the Ontario Innovation Trust (OIT), and the Bell University Labs initiative. References 1. P. K. Agarwal. Range Searching. In J. E. Goodman and J. O'Rourke, editors, Handbook of Discrete and Computational Geometry. CRC Press, July 1997. 2 2. P. K. Agarwal and J. Erickson. Geometric range searching and its relatives. Advances in Discrete and Computational Geometry, 23:156, 1999. 2 c The Eurographics Association 2002. 97 Ma and McCool / Low Latency Photon Mapping Using Block Hashing 3. S. Arya and D. M. Mount. Approximate Nearest Neighbor Queries in Fixed Dimensions. In Proc. ACM-SIAM SODA, 1993. 2 4. S. Arya, D. M. Mount, N. S. Netanyahu, R. Silverman, and A. Y. Wu. An Optimal Algorithm for Approximate Nearest Neighbor Searching. In Proc. ACM-SIAM SODA, pages 573582, 1994. 2 5. J. L. Bentley. Multidimensional binary search trees used for associative searching. Communications of the ACM, 18(9), September 1975. 1 , 2 6. J. L. Bentley, B. W. Weide, and A. C. Chow. Optimal Expected-Time Algorithms for Closest Point Problems. ACM TOMS, 6(4), December 1980. 1 , 2 7. Per H. Christensen. Faster Photon Map Global Illumination . Journal of Graphics Tools, 4(3):110, 1999. 2 8. D. Comer. The Ubiquitous B-Tree. ACM Computing Surveys, 11(2):121137, June 1979. 4 9. C. A. Duncan, M. T. Goodrich, and S. G. Kobourov. Balanced aspect ratio trees: Combining the advantages of k-d trees and octrees. In Proc. ACM-SIAM SODA, volume 10, pages 300309, 1999. 2 10. H. Edelsbrunner. Algorithms in Combinatorial Geometry . Springer-Verlag, 1987. 2 11. D. Eppstein, M. S. Paterson, and F. F. Yao. On nearest neighbor graphs. Discrete & Computational Geometry, 17(3):263282, April 1997. 2 12. C. Faloutsos and Y. Rong. DOT: A Spatial Access Method Using Fractals. In Proc. 7th Int. Conf. on Data Engineering,, pages 152159, Kobe, Japan, 1991. 4 13. C. Faloutsos and S. Roseman. Fractals for Secondary Key Retrieval. In Proc. 8th ACM PODS, pages 247 252, Philadelphia, PA, 1989. 4 14. J. H. Freidman, J. L. Bentley, and R. A. Finkel. An Algorithm for Finding Best Matches in Logarithmic Expected Time. ACM TOMS, 3(3):209226, 1977. 2 15. V. Gaede and O. Gnther. Multidimensional access methods. ACM Computing Surveys (CSUR), 30(2):170231, 1998. 2 16. A. Gionis, P. Indyk, and R. Motwani. Similarity Search in High Dimensions via Hashing. In Proc. VLDB, pages 518529, 1999. 2 , 3 17. J. E. Goodman and J. O'Rourke, editors. Handbook of Discrete and Computational Geometry. CRC Press, July 1997. ISBN: 0849385245. 2 18. G. Greger, P. Shirley, P. Hubbard, and D. Greenberg. The irradiance volume. IEEE CG&A, 18(2):3243, 1998. 1 19. V. Havran. Analysis of Cache Sensitive Representation for Binary Space Partitioning Trees. Informatica, 23(3):203210, May 2000. 2 20. P. Indyk and R. Motwani. Approximate Nearest Neighbors : Towards Removing the Curse of Dimensionality. In Proc. ACM STOC, pages 604613, 1998. 2 21. P. Indyk, R. Motwani, P. Raghavan, and S. Vem-pala . Locality-Preserving Hashing in Multidimensional Spaces. In Proc. ACM STOC, pages 618625, 1997. 2 22. H. V. Jagadish. Linear clustering of objects with multiple attributes. In Proc. Acm-sigmod, pages 332342, May 1990. 4 23. J. W. Jaromczyk and G. T. Toussaint. Relative Neighborhood Graphs and Their Relatives. Proc. IEEE, 80(9):15021517, September 1992. 2 24. H. W. Jensen. Rendering Caustics on Non-Lambertian Surfaces. Computer Graphics Forum, 16(1):5764, 1997. ISSN 0167-7055. 1 25. H. W. Jensen. Realistic Image Synthesis Using Photon Mapping. A.K. Peters, 2001. 1 , 2 , 4 26. H. W. Jensen, F. Suykens, and P. H. Christensen. A Practical Guide to Global Illumination using Photon Mapping. In SIGGRAPH 2001 Course Notes, number 38. ACM, August 2001. 2 27. J. K. P. Kuan and P. H. Lewis. Fast k Nearest Neighbour Search for R-tree Family. In Proc. of First Int. Conf. on Information, Communication and Signal Processing , pages 924928, Singapore, 1997. 2 28. E. P. Lafortune and Y. D. Willems. Using the Modified Phong BRDF for Physically Based Rendering. Technical Report CW197, Department of Computer Science, K.U.Leuven, November 1994. 7 29. C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to Algorithms. MIT Press, 2001. 6 30. K.-I. Lin and C. Yang. The ANN-Tree: An Index for Efficient Approximate Nearest-Neighbour Search. In Conf. on Database Systems for Advanced Applications, 2001. 2 31. N. Linial and O. Sasson. Non-Expansive Hashing. In Proc. acm stoc, pages 509518, 1996. 2 32. V. Ma. Low Latency Photon Mapping using Block Hashing. Technical Report CS-2002-15, School of Computer Science, University of Waterloo, 2002. 9 33. J. McNames. A Fast Nearest-Neighbor Algorithm Based on a Principal Axis Search Tree. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(9):964976, 2001. 2 34. J. Nievergelt, H. Hinterberger, and K. C. Sevcik. The Grid File: an adaptable, symmetric multikey file structure . ACM TODS, 9(1):3871, March 1984. 3 , 5 35. T. J. Purcell, I. Buck, W. R. Mark, and P. Hanrahan. Ray Tracing on Programmable Graphics Hardware. In to appear in Proc. SIGGRAPH, 2002. 1 36. J. T. Robinson. The K-D-B-tree: A Search Structure for Large Multidimensional Dynamic Indexes. In Proc. acm sigmod, pages 1018, 1981. 2 37. Arnold L. Rosenberg and Lawrence Snyder. Time- and space-optimality in b-trees. ACM TODS, 6(1):174193, 1981. 4 38. K. Salem and H. Garcia-Molina. Disk striping. In IEEE ICDE, pages 336342, 1986. 5 c The Eurographics Association 2002. 98 Ma and McCool / Low Latency Photon Mapping Using Block Hashing 39. N. Sample, M. Haines, M. Arnold, and T. Purcell. Optimizing Search Strategies in kd-Trees. May 2001. 2 40. G. Schaufler and H. W. Jensen. Ray Tracing Point Sampled Geometry. Rendering Techniques 2000, pages 319328, June 2000. 9 41. J. Schmittler, I. Wald, and P. Slusallek. SaarCOR - A Hardware Architecture for Ray Tracing. In to appear at EUROGRAPHICS Graphics Hardware, 2002. 1 42. M. Smid. Closest-Point Problems in Computational Geometry. In J. R. Sack and J. Urrutia, editors, Handbook on Computational Geometry. Elsevier Science, Amsterdam, North Holland, 2000. 2 43. P. Tsaparas. Nearest neighbor search in multidimen-sional spaces. Qualifying Depth Oral Report 319-02 , Dept. of Computer Science, University of Toronto, 1999. 2 44. M. Vanco, G. Brunnett, and T. Schreiber. A Hashing Strategy for Efficient k-Nearest Neighbors Computation . In Computer Graphics International, pages 120 128. IEEE, June 1999. 2 45. I. Wald, T. Kollig, C. Benthin, A. Keller, and P. Slusallek. Interactive global illumination. Technical report, Computer Graphics Group, Saarland University, 2002. to be published at EUROGRAPHICS Workshop on Rendering 2002. 2 46. G. Ward. Real Pixels. In James Arvo, editor, Graphics Gems II, pages 8083. Academic Press, 1991. 4 47. Steven Worley. A cellular texture basis function. In Proc. SIGGRAPH 1996, pages 291294. ACM Press, 1996. 9 48. M. Zwicker, H. Pfister, J. van Baar, and M. Gross. Surface Splatting. Proc. SIGGRAPH 2001, pages 371378, 2001. 9 c The Eurographics Association 2002. 99 Ma and McCool / Low Latency Photon Mapping Using Block Hashing (a) kd-tree (b) BH, A=16 (c) BH, A=8 (d) BH, A=4 Figure 13: "Ring" (a) kd-tree (b) BH, A=16 (c) BH, A=8 (d) BH, A=4 Figure 14: "Venus with Ring" c The Eurographics Association 2002. 158

photon mapping;block hashing (BH);hashing techniques;AkNN;kNN;accelerator

131

Lower Bounds & Competitive Algorithms for Online Scheduling of Unit-Size Tasks to Related Machines

In this paper we study the problem of assigning unit-size tasks to related machines when only limited online information is provided to each task. This is a general framework whose special cases are the classical multiple-choice games for the assignment of unit-size tasks to identical machines. The latter case was the subject of intensive research for the last decade. The problem is intriguing in the sense that the natural extensions of the greedy oblivious schedulers, which are known to achieve near-optimal performance in the case of identical machines, are proved to perform quite poorly in the case of the related machines. In this work we present a rather surprising lower bound stating that any oblivious scheduler that assigns an arbitrary number of tasks to n related machines would need log n polls of machine loads per task, in order to achieve a constant competitive ratio versus the optimum offline assignment of the same input sequence to these machines . On the other hand, we prove that the missing information for an oblivious scheduler to perform almost optimally , is the amount of tasks to be inserted into the system. In particular, we provide an oblivious scheduler that only uses O(loglog n) polls, along with the additional information of the size of the input sequence, in order to achieve a constant competitive ratio vs. the optimum offline assignment . The philosophy of this scheduler is based on an interesting exploitation of the slowfit concept ([1, 5, 3]; for a survey see [6, 9, 16]) for the assignment of the tasks to the related machines despite the restrictions on the provided online information, in combination with a layered induction argument for bounding the tails of the number of tasks passing from slower to faster machines. We finally use this oblivious scheduler as the core of an adaptive scheduler that does not demand the knowledge of the input sequence and yet achieves almost the same performance.

INTRODUCTION The problem of the Online Assignment of Tasks to Related Machines is defined as follows: there are n machines possibly of different speeds, that are determined by a speed vector c , and an input sequence of m independent tasks to be assigned to these machines. The tasks arrive sequentially, along with their associated weights (positive reals) and have to be assigned immediately and uniquely to the machines of the system. The size of the input sequence as well as the weights of the tasks are determined by an oblivious adversary (denoted here by ADV). Each task has to be assigned upon its arrival to one of the machines, using the following information: (possibly a portion of) the online information of current status of the system, the offline information of the machine speeds, and its own weight. The tasks are considered to be of infinite duration (permanent tasks) and no preemption is allowed. The cost of an online scheduler ALG for the assignment of an input sequence of tasks (denoted by ALG()) is the maximum load eventually appearing in the system. In case that a random-ized scheduler is taken into account, then the cost of the scheduler is the expectation of the corresponding random variable. The quality of an online scheduler is compared vs. the optimum offline assignment of the same input sequence to the n machines. We denote the optimum offline cost for by ADV(). That is, we consider the competitive ratio (or performance guarantee) to be the quality measure, (eg, see [6]): Definition 1.1. An online scheduler ALG is said to achieve a competitive ratio of parameters (a, ), if for any 124 input sequence the relation connecting its own cost ALG() and the optimum offline cost of ADV, are related by ALG() a ADV() + . ALG is strictly a-competitive if , ALG() a ADV(). In this work we study the consequences of providing only some portion of the online information to a scheduler. That is, we focus our interest on the case where each task is capable of checking the online status only by a (small wrt n) number d of polls from the n machines. In this case, the objective is to determine the trade-off between the number of polls that are available to each of the tasks and the performance guarantee of the online scheduler, or equivalently, to determine the minimum number of polls per task so that a strictly constant competitive ratio is achieved. Additionally, we consider the case of unit-size tasks that are assigned to related machines. Thus, each task t [m] has to be assigned to a machine host(t) [n] using the following information that is provided to it: the current loads of d suitably chosen machines (the kind of the "suitable" selection is one of the basic elements of a scheduler and will be called the polling strategy from now on) and an assignment strategy that determines the host of t among these d selected candidates on behalf of t. In what follows we shall consider homogeneous schedulers , ie, schedulers that apply exactly the same protocol on all the tasks that are inserted into the system. This choice is justified by the fact that no task is allowed to have access to knowledge concerning previous or forthcoming choices of other tasks in the system, except only for the current loads of those machines that have been chosen to be its candidate hosts. Additionally, we shall use the terms (capacitated) bins instead of (related) machines and (identical) balls instead of (unit-size) tasks interchangeably, due to the profound analogy of the problem under consideration with the corresponding Balls & Bins problem. 1.1 Polling Strategies The way a scheduler ALG lets each newly inserted task choose its d candidate hosts is called a polling strategy (PS). We call the strategies that poll candidate machines homogeneously for all the inserted tasks of the same size, homogeneous polling strategies (HPS). In the present work we consider the tasks to be indistinguishable: Each task upon its arrival knows only the loads of the machines that it polls, along with the speed (or equiv. capacity wrt bins) vector c of the system. This is why we focus our interest in schedulers belonging to HPS. Depending on the dependencies of the polls that are taken on behalf of a task, we classify the polling strategies as follows: Oblivious polling strategies(HOPS) In this case we consider that the polling strategy on behalf of a newly inserted task t consists of an independent (from other tasks) choice of a d-tuple from [n] d according to a fixed probability distribution f : [n] d [0, 1]. This probability distribution may only depend on the common offline information provided to each of the tasks. It should be clear that any kind of d independent polls (with or without replacement ) on behalf of each task, falls into this family of polling strategies. Thus the whole polling strategy is the sequence of m d-tuples chosen independently (using the same probability distribution f ) on behalf of the m tasks that are to be inserted into the system. Clearly for any polling strategy belonging to HOPS, the d random polls on behalf of each of the m tasks could have been fixed prior to the start of the assignments. Adaptive polling strategies (HAPS) In this case the i th poll on behalf of ball t [m] is allowed to exploit the information gained by the i - 1 previous polls on behalf of the same ball. That is, unlike the case of HOPS where the choice of d candidates of a task was oblivious to the current system state, now the polling strategy is allowed to direct the next poll to specific machines of the system according to the outcome of the polls up to this point. In this case all the polls on behalf of each task have to be taken at runtime, upon the arrival of the task. Remark: It is commented that this kind of polling strategies are not actually helpful in the case of identical machines, where HOPS schedulers achieve asymptotically optimal performance (see [18]). Nevertheless, we prove here that this is not the case for the related machines. It will be shown that oblivious strategies perform rather poorly in this setting , while HAPS schedulers achieve actually asymptotically optimal performance. 1.2 Assignment Strategies Having chosen the d-size set of candidate hosts for task t [m], the next thing is to assign this task to one of these machines given their current loads and possibly exploiting the knowledge on the way that they were selected. We call this procedure the assignment strategy. The significant question that arises here is the following: Given the polling strategy adopted and the knowledge that is acquired at runtime by the polled d-tuple on behalf of a task t [m], which would be the optimal assignment strategy for this task so that the eventual maximum load in the system be minimized? In the Unit Size Tasks-Identical Machines case, when each of the d polls is chosen iur (with replacement) from [n], Azar et al. ([2]) show that the best assignment strategy is the minimum load rule and requires O(log n) polls per task for a strictly constant competitive ratio. Consequently V ocking ([18]) has suggested the always go left strategy , which (in combination with a properly chosen oblivious polling strategy) only requires a total number of O(loglog n) in order to achieve a constant competitive ratio. In the same work it was also shown that one should not expect much by exploiting possible dependencies of the polls in the case of unit-size tasks that are placed into identical machines, since the load of the fullest machine is roughly the same as the one achieved in the case of non-uniform but independent polls using the always go left rule. Nevertheless, things are quite different in the Related Machines case: we show by our lower bound (section 3) that even if a scheduler ALG considers any oblivious polling strategy and the best possible assignment strategy, ALG has a strict competitive ratio of at least 2d n 4 d-2 , where d is the number of oblivious polls per task. This implies that in the case of the related machines there is still much space for the adaptive polling strategies until the lower bound of (loglog n) polls per task is matched. 125 1.3 Related Work In the case of assigning unit-size tasks to identical machines , there has been a lot of intensive research during the last decade. If each task is capable of viewing the whole status of the system upon its arrival (we call this the Full Information case), then Graham's greedy algorithm assures a competitive ratio that asymptotically tends to 2 1 n ([6]). Nevertheless, when the tasks are granted only a limited number of polls, things are much more complicated: In the case of unit-size tasks and a single poll per task, the result of Gonnet [10] has proved that for m n the maximum load is (1 + o(1)) ln n lnln n when the poll of each task is chosen iur from the n machines, whp. 1 In [15] an explicit form for the expected maximum load is given for all combinations of n and m. From this work it easily seen that for m n ln n, the maximum load is m n + ( m ln n/n), which implies that by means of competitive ratio, m n is actually the hardest instance. In the case of d 2 polls per task, a bunch of new techniques had to be applied for the analysis of such schedulers. The main tools used in the literature for this problem have been the layered induction, the witness tree argument and the method of fluid models (a comprehensive presentation of these techniques may be found in the very good survey of Mitzenmacher et al. [14]). In the seminal paper of Azar Broder Karlin and Upfal [2] it was proved that the proposed scheduler abku that chooses each of the polls of a task iur from [n] and then assigns the task to the candidate machine of minimum load, achieves a maximum load that is at most m n + lnln n ln d (1). This implies a strictly O lnln n ln d competitive ratio, or equivalently, at least O(ln n) polls per task would be necessary in order to achieve a strictly constant competitive ratio. In [18] the always go left algorithm was proposed, which assures a maximum load of at most m n + lnln n d ln 2 (1) and thus only needs an amount of O(loglog n) polls per task in order to achieve a strictly constant competitive ratio. In addition it was shown that this is the best possible that one may hope for in the case of assigning unit-size tasks to identical machines with only d (either oblivious or adaptive) polls per task. The Online Assignment of Tasks to Related Machines problem has been thoroughly studied during the past years for the Full Information case (eg, see chapter 12 in [6]). In particular, it has been shown that a strictly (small) constant competitive ratio can be achieved using the slowfit-like algorithms that are based on the idea of exploiting the least favourable machines (this idea first appeared in [17]). The case of Limited Information has attracted little attention up to this point: some recent works ([12, 13, 8]) study the case of each task having a single poll, for its assignment to one of the (possibly related) machines when the probability distributions of the tasks comprise a Nash Equilibrium . For example, in [8] it was shown that in the Related Machines case a coordination ratio (ie, the ratio between the worst possible expected maximum load among Nash Equilibria over the offline optimum) of O log n logloglog n . However, when all the task weights are equal then it was shown by Mavronicolas and Spirakis [13] that the coordination ratio 1 A probabilistic event A is said to hold with high probability (whp) if for some arbitrarily chosen constant &gt; 0, IP[A] 1 - n . is O log n loglog n . As for the case of d &gt; 1 in the Related Machines problem, up to the author's knowledge this is the first time that this problem is dealt with. 1.4 New results In this work we show that any HOPS scheduler requires at least O log n loglog n polls in order to achieve a strictly constant competitive ratio vs. an oblivious adversary. The key point in this lower bound argument is the construction of a system of d+1 groups of machines running at the same speed within each group, while the machine speeds (comparing machines of consecutive groups) fall by a fixed factor and on the other hand the cumulative capacities of the groups are raised by the same factor. Then it is intuitively clear that any HOPS scheduler cannot keep track of the current status within each of these d + 1 groups while having only d polls per new task, and thus it will have to pay the cost of misplacing balls in at least one of these groups. More specifically, we show the following lower bound: Theorem 1. d 1, the competitive ratio of any d-hops scheduler is at least 2d n 4 d-2 . Then we propose a new d-hops scheduler OBL which, if it is fortified with the additional knowledge of the total number of tasks to be inserted, then it achieves the following upper-bound: Theorem 2. Let lnln n - lnlnln n &gt; d 2 and suppose that the size of the input sequence is given as offline information.If OBL provides each task with (at most) 2d polls, then it has a strict competitive ratio that drops double-exponentially with d.In particular the cost of OBL is with high probability at most OBL (m) (1 + o(1))8 n d2 d+3 1 /(2 d+1 -1) + 1 ADV(m) It is commented that all the schedulers for the Identical Machines-Limited Information case up to now used minimum load as the profound assignment rule. On the other hand, OBL was inspired by the slowfit approaches for the Related Machines-Full Information problem and the fact that a greedy scheduler behaves badly in that case. Up to the author's knowledge, the idea of using the slow-est machine possible first appeared in [17]. Additionally, a layered induction argument is employed for bounding the amount of tasks that flow from the slower to the faster machines in the system. 2 This then allows the use of relatively simple arguments for bounding the maximum load of the tasks that end up in a small fraction of the system that consists of the fastest machines. Clearly this upper bound is near-optimal (up to a multiplicative constant), since it matches the (loglog n) lower bound of the Unit Size Tasks-Identical Machines problem ([18]) which is a subcase of our setting. Finally we propose a haps scheduler ( ADAPT ) that combines the previous hops scheduler with a classical guessing 2 Note that this does not imply preemption of tasks which is not allowed in our setting, but rather the event that a task hits slower machines that are already overloaded and thus has to assign itself to a faster machine. 126 argument for the cost of ADV and assures a cost roughly 5 times the cost of OBL : Theorem 3. For any input sequence of identical tasks that have to be assigned to n related machines using at most 2d + 1 polls per task, the cost of ADAPT is (whp), ADAPT () &lt; O n d2 d+3 1 /(2 d+1 -1) + 1 ADV() A SIMPLE LOWER BOUND ON HOMOGENEOUS SINGLE-POLL GAMES This section contains a simple argument for the claimed lower bound on online schedulers that devote a single poll per new task, ie, d = 1. Clearly by their nature these are HOPS schedulers, since there is no actual option for the assignment strategy. The proof for the lower bound of these schedulers is rather simple but it will and shed some light to the essence of the construction for the proof of the general lower bound that will follow in the next section. Let's assume that there exists a HOPS scheduler that only uses 1 poll per task and claims to be strict a-competitive against any oblivious adversary ADV. Initially ADV chooses an arbitrary real number r n which will be fixed in the end so as to maximize the lower bound on a. Let also the variables C total , C max denote the total capacity and the maximum possible polled capacity using one poll (ie, the maximum bin capacity in this case) in the system. Consequently ADV uses the following system of capacitated bins so that these values are preserved: C 1 = C max , C i = C max r , i = 2, . . . , C total - C max C max r + 1 ( n) Observe that the capacity of bin i 2 is r times smaller than C max , while on the other hand, the cumulative capacity of the last n - 1 machines is n-1 r times larger than the capacity of the largest bin in the system. Consider also the following abbreviations of probabilities and events that may occur upon the arrival of a new ball: E i "bin i is hit by a ball" P 1 IP[E 1 ], P 1 1 - P 1 Obviously due to the assumption of a-competitiveness, a C total C max = C max + ( n-1)C max r C max = 1 + n - 1 r since ALG could choose to assign all the incoming balls to the largest bin in the system. The question that arises is whether there exists a 1-poll scheduler that can do better than that. We consider the following input sequences: || = 1, w 1 = w: ALG() = IE[L max ()] P 1 w C max + P 1 w Cmax r ADV() = w C max / a-comp. / = a P 1 + r P 1 = r - P 1 (r - 1) P 1 r-a r-1 || = , t 1, w t = w: In this case the loads of all the bins will tend to their expected values, and thus || (1) = IE[ || (1)] = P 1 ||w C max ALG() = IE[L max ()] P 1 ||w C max ADV() = ||w C total / a-comp. / = a ||w C total P 1 ||w C max P 1 aC max C total = a n-1 r +1 Combining the two bounds on P 1 we get: aC max C total = a n-1 r + 1 P 1 r - a r - 1 a r - a n - 1 + r - a n - 1 r + 1 a r + n - 1 r r + n - 1 a r + n - 1 r + n-1 r = r 2 + n r - r r 2 + n - 1 which is maximized for r = n + 1 and assures a lower bound on a of n 2 . Remark: It is worth noting that the lower bound com-pletely depends on the number of bins in the system, and on the ratio r = C max C min and does not depend at all on the total capacity of the system, C total . THE LOWER BOUND ON MULTI-HOPS SHCEDULERS In this section we study the behaviour of homogeneous schedulers that adopt an oblivious polling strategy (ie, the polling strategy is from HOPS) and an arbitrary assignment strategy. We call these d-hops schedulers, since the choice of the d candidates on behalf of each ball is done independently for each ball, according to a common probability distribution f : [n] d [0, 1]. Recall that the choice of the candidate bins for each ball is oblivious to the current system state and thus could have been fixed prior to the beginning of the assignments. Theorem 1. d 1, the competitive ratio of any d-hops scheduler is at least 2d n 4 d-2 . Proof: Let f : [n] d [0, 1] be the adopted Oblivious polling strategy by an arbitrary d-hops scheduler, ALG. Assume also that ALG uses the best possible assignment strategy given this polling strategy, that is, each ball chooses its own candidate bins according to f and then it may assign itself to an arbitrarily chosen host among its candidates, depending on the current loads of the candidate bins. Assume also that ALG claims a (strict) competitive ratio a against oblivious adversaries. As parameters of the problem we consider again the quantities C total = n i=1 C i and C max : the total capacity of a system of n related machines and the maximum capacity that may be returned by a single poll. We shall describe an adversary ADV that initially chooses an arbitrary real number 1 r n and then considers the system of (d + 1 groups of) n capacitated bins that is described in Table 1. Observe that this construction preserves the following two invariants when considering two successive groups of bins F , F +1 : 1 d: 127 Group of Bins Number of Bins in group Capacity per Bin Cumulative Group Capacity F 1 1 C max C max F 2 r(r - 1) C max /r (r - 1)C max F 3 r 3 (r - 1) C max /r 2 r(r - 1)C max F 4 r 5 (r - 1) C max /r 3 r 2 (r - 1)C max . . . . . . . . . . . . F d r 2 d-3 (r - 1) C max /r d-1 r d-2 (r - 1)C max F d+1 n - 1 r r+1 (r 2 d-2 - 1) n - r 2 d-2 C max /r d ( n r d - r d-2 ) C max Table 1: The system of (d + 1 groups of ) capacitated bins considered by ADV for the proof of the Lower Bound on d-HOPS schedulers. (I1) when going from one group to its successor, the bin capacities decrease by a factor of r, and (I2) the cumulative capacity of the first + 1 groups is larger than the cumulative capacity of the first groups by a factor of r. We shall denote by C[F ] the cumulative capacity of any group of bins F [n]. Remark: The preservation of invariant (I2) when = d implies that C[F d+1 ] r d-1 (r-1)C max n r d - r d-2 C max r d-1 (r - 1)C max n r 2 d - r 2 d-1 + r 2 d-2 . We fortify ALG by allowing a perfect balance of the bins of a group F whenever at least one poll on behalf of a new ball goes to a bin of this group. This is actually in order to capture the notion of the "perfect assignment strategy given the polling strategy" claim stated above. Clearly this does not cause any problem since we are looking for a lower bound. Because ALG could lock its d choices to the first d groups of the system, it is obvious that its competitive ratio a is at most a C total C[ d =1 F ] = n r 2d-1 + 1. Consider now the d events E "F is hit by a ball" (1 d), while P IP[E ] (call it the hitting probability of group F ) is the probability of at least one bin from F being hit by a ball. We shall charge ALG according to the hitting probabilities that its polling strategy determines . Notice that these are fixed at the beginning of the assignments since the polling strategy of ALG is an oblivious strategy. Furthermore, the following conditional hitting probabilities are also determined uniquely by the polling strategy of ALG: i, j [n] : i &gt; j, P i|j IP[E i |E 1 E 2 E j ], Q i|j IP[E i |E 1 E 2 E j ]. Finally, let B () ( = 1, . . . , d) denote the maximum number of balls that may be hosted by bins of the set =1 F without violating the assumption of a-competitiveness of ALG, when the input sequence of tasks is chosen by ADV. The following lemma states an inherent property of any d-hops scheduler: Lemma 3.1. For any &gt; 1, unless ALG admits a competitive ratio a &gt; ( -1)(1-r -2 ) d 2 r , the following property holds: 1 d, P |-1 1 - 1 a r Proof: We prove this lemma by considering the following input sequences of balls of the same (arbitrarily chosen) size w: || = 1: In this case we know that ADV() = w C max . The cost (ie, the expectation of maximum load) of ALG is: ALG() P 1 w C max + (1 - P 1 ) Q 2 |1 rw C max + (1 - Q 2 |1 )Q 3 |2 r 2 w C max + + + (1 - Q 2 |1 ) (1 - Q d|d-1 ) r d w C max Due to the demand for a-competitiveness of ALG against ADV, this then implies a r 1 - P 1 P 1 1 a r . || = r 2-2 , = 2, , d: In this case ADV will use only the bins of =1 F and thus he will pay a cost of ADV() = r 2-2 w r -1 C max = r -1 w C max . As for the cost of ALG, we shall only charge it for the input subsequence of balls that definitely hit groups F 1 , . . . , F -1 (call it ^ ). Our purpose is from this sequence of tasks to determine P |-1 , ie, the conditional hitting probability of group F given that all the previous groups are hit by a ball. Clearly, IE[ |^|] = || -1 =1 P |-1 = r 2 -2 -1 =1 P |-1 (where for symmetry of representation we let P 1 |0 = P 1 ). Recall that B -1 () denotes the maximum number of balls that may be assigned to the bins of the first - 1 groups, given the claimed competitive ratio a by ALG and the input sequence . Then we have: wB -1 ( ) C[ -1 =1 F ] a ADV() B -1 () a r 2 -3 . Thus, there is a subsequence ~ of ^ that consists of those tasks which cannot be assigned to the bins of the first - 1 groups due to the a-competitiveness constraint. All these tasks have to exploit their remaining (at most) d - + 1 polls among the bins of [n] \ -1 =1 F . It is clear that ALG has no reason to spoil more than one poll per group due to the optimal assignment strategy that it adopts. Thus we can safely assume that there remain exactly d - + 1 polls for the remaining groups. Obviously IE[ |~|] IE[|^|] - B -1 () r 2 -2 -1 =1 P |-1 - ar 2 -3 r 2 -3 (r -1 - a), where for simplification of notation we use the bounding se-128 quence -1 = -2 1 -1 a r -1 = 1 -1 a r -1 and 0 = 1. This is true because P 1 1 a r 1 = 1 -1 a r , &gt; 1, while we assume inductively that -1 =1 P |-1 -1 . By showing that P |-1 1 -1 a r we shall also have assured that =1 P |-1 . We apply the Markov Inequality (on the complementary random variable ||-|~|) to find a lower bound on the size of ~ : &gt; 1, IP[|~| r 2 -3 (r - r + r -1 - a)] 1 - 1. Now it is clear that if ALG claims a competitive ratio a ( - 1)(1 - r -2 ) 2 d r ( - 1) 1 1 r 2-2 2 r , then at least one ball of will belong to ~ with probability at least 1 1 . Thus, either ALG has a &gt; ( -1)(1-r -2 ) 2 dr , or (by simply charging it only for this very specific ball) ALG() 1 - 1 [P |-1 + (1 - P |-1 )r] r -1 w C max which, combined with the demand for a-competitiveness and the cost of ADV for , implies that P |-1 1 - 1 a r . We finally try the following input sequence, in case that ALG still claims a competitive ratio a ( -1)(1-r -2 ) d 2 r: || = : For this input sequence it is clear that ADV() = ||w C total = ||w r d-1 - r d-2 + n r d C max . For ALG we again consider the subsequence ^ of balls that definitely hit the first d - 1 groups of the system. Clearly |^| = IE[|^|] d-1 || since we now consider an infinite sequence of incoming balls. As for the upper bound on the balls that the first d - 1 groups can host, this is again given by the demand for a-competitiveness: wB d-1 () C[ d-1 =1 F ] = wB d-1 () r d-2 C max a||w r d-1 - r d-2 + n r d C max B d-1 () r d-1 r d-1 - r d-2 + n r d a r || The subsequence ~ ^ that has to exploit a single poll among the bins of [n] \ d-1 =1 F has size at least |~| = IE[|~|] IE[|^|] - B d-1 () d-1 r d-1 r d-1 - r d-2 + n r d a r || 1 - (d - 1) - 1 a r r d-1 r d-1 - r d-2 + n r d a r || 1 - d - 1 - 1 a r || where for the last inequality we consider that n r 2 d-2 . Since we consider that a ( -1)(1-r -2 ) 2 dr , we can be sure that |~| 1 1 + 1 r 2 || and thus, the cost of ALG will be lower bounded by the expected load of the bins in F d due to the tasks of ~ : a ADV() ALG() P d|d-1 |~| w (r d-1 - r d-2 ) C max a r d-1 - r d-2 + n r d P d|d-1 1 ( -1)(1-r -2 ) 2 dr d-1 - r d-2 1 1 ( -1)(1-r -2 ) (d-1) 1 + n r 2d-2 ( r-1) 1 - 1 a r 1 - 1 - r -2 d - 1 which is not possible for any &gt; 1 and n r 2 d . Thus we conclude that ALG cannot avoid a competitive ratio a min - 1 (d - 1) r, n r 2 d-1 + 1 for any &gt; 1 and n r 2 d , which for = 2 and n = r 2 d gives the desired bound. DEALING WITH INPUT SEQUENCES OF KNOWN TOTAL SIZE In this section we prove that the missing information for an oblivious scheduler to perform efficiently is the size of the input sequence. More specifically, considering that the input size is provided as offline information to each of the newly inserted tasks, we construct an oblivious scheduler that exploits this information along with a slowfit assignment rule and a layered induction argument for the flow of balls from slower to faster bins, in order to achieve a strictly constant competitive ratio with only O(loglog n) polls per task. Assume that m unit-size balls are thrown into a system of n capacitated bins with capacities C max = C n C n-1 C 1 = C min . Assume also that each ball is allowed to poll up to 2d bins and then it has to assign itself to one of these candidates. We additionally assume that C max n 2 d+1 . As it will become clear later by the analysis, if this was not the case then it could only be in favour of the oblivious scheduler that we propose, because this would allow the absorption of the large additive constants in the performance guarantee of the scheduler. We consider (wlog) that the capacity vector c is nor-malized by n ||c|| 1 so that n i=1 C i = n. We also assume that the total size m the input sequence is given to every newly inserted ball. This implies that each ball can estimate the cost ADV(m) opt (ie, the optimum offline assignment of the m unit-size balls to the n capacitated bins), and thus it can know a priori the subset of bins that may have been used by ADV during the whole process . 3 Having this in mind, we can assume that every bin in the system is legitimate, that is, it might have been used by the optimum solution, otherwise we could have each ball ignore all the illegitimate bins in the system. Thus, opt max 1 C min , m C total = max 1 C min , m n . Finally, we assume that each of the legitimate bins of the system gets at 3 A bin i [n] may have been used by ADV if and only if 1/C i opt. 129 least one ball in the optimum offline schedule. This does not affect the performance of ADV, while it may only deteriorate the performance of an online scheduler. Nevertheless, it assures that m n opt m n + 1, meaning that the fractional load on the bins is actually a good estimation of opt. Let the load of bin i [n] at time t (that is, right after the assignment of the t th ball of the input sequence) be denoted as t (i) q t ( i) C i , where q t (i) is the number of balls assigned to bin i up to that time. The following definition refers to the notion of saturated bins in the system, ie, overloaded bins wrt the designed performance guarantee of an oblivious scheduler: Definition 4.1. A bin i [n] is called saturated upon the arrival of a new ball t m, if and only if it has t-1 (i) &gt; a opt (a), where (a) is called the designed performance guarantee of the oblivious scheduler. Let r [d], i r min{i [n] : i j=1 C j r =1 n 2 }. Then, we consider the following partition of the set of bins [n] into d + 1 groups of (roughly geometrically) decreasing cumulative capacities: F 1 {1, . . . , i 1 }, F r {i r-1 + 1, . . . , i r }, r = 2, . . . , d, F d+1 {i d + 1, . . . , n}. Although the cumulative capacity of group F r may vary from n 2 r -C max to n 2 r + C max , for ease of the following computations we assume that asymptotically r [d], C[F r ] n 2 r and C[F d+1 ] n 2 d . We denote by C min the capacity of the smallest bin in F , [d + 1]. We now consider the following ideal scheduler that uses an oblivious polling strategy and an assignment strategy based on the slowfit rule. This scheduler (we call it OBL ) initially discards all the illegitimate bins in the system, using the knowledge of m. Then first it normalizes the capacity vector of the remaining bins and afterwards it considers the grouping mentioned above and adopts the following pair of strategies: POLLING: 1 r d group F r gets exactly 1 poll, which is chosen among the bins of the group proportionally to the bin capacities. That is, r [d], i F r , IP[bin i F r is a candidate host of a ball] C i C[F r ] . The remaining d polls are assigned to the bins of group F d+1 , either to the d fastest bins, or according to the polling strategy of always go left, depending only on the parameters (c, d) of the problem instance. 4 ASSIGNMENT: Upon the arrival of a new ball t [m], the smallest polled bin from d =1 F (starting from F 1 , to F 2 , a.s.o.) that is unsaturated gets this ball (slowfit rule). In case that all the first d polls of a ball are already saturated , then this ball has to be assigned to a bin of F d+1 using its remaining d candidates. Within group F d+1 , either the minimum post load rule (ie, the bin of minimum load among the d choices from F d+1 is chosen, taking into account also the additional load of the new ball), or the slowfit rule 4 Observe that this is offline information and thus this decision can be made at the beginning of the assignment process, for all the balls of the input sequence. is applied, depending on the offline parameters (c , d) of the problem instance. If all the 2d polled bins are saturated, then the minimum post load rule is applied among them. Ties are always broken in favour of smaller bins (ie, slower machines). The following theorem gives the performance of OBL , when the additional information of the input size is also provided offline: Theorem 2. For lnln n - lnlnln n &gt; d 2, when the size of the input sequence m n is given as offline information , OBL has a strict competitive ratio that drops double-exponentially with d.In particular the cost of OBL is (whp) at most OBL (m) (1 + o(1))8 n d2 d+3 1 /(2 d+1 -1) + 1 ADV(m) Proof: Let + 1 m be the first ball in the system that hits only saturated bins by its 2d polls. Our purpose is to determine the value of a in the designed performance guarantee (a), so that the probability of ball +1 existing to be polynomially small. As stated before, the technique that we shall employ is a layered induction argument on the number of balls that are passed to the right of group F r , r [d]. For the assignment of the balls that end up in group F d+1 we use a slightly modified version of the always go left scheduler of [18] that gives an upper bound on the maximum load in group F d+1 (we denote this by L max [F d+1 ]). This upper bound on L max [F d+1 ] holds with high probability. This assignment is only used when it produces a smaller maximum load than the brute assignment of all the balls ending up in F d+1 to the d fastest bins of the group. We shall consider a notion of time that corresponds to the assignments of newly arrived balls into the system: At time t m, the t th ball of the input sequence is thrown into the system and it has to be immediately assigned to one of its 2d candidates. Consider the polls on behalf of a ball to be ordered according to the groups from which they are taken. Observe then that each ball t is assigned to the first unsaturated bin that it hits from the first d groups, or to a bin in group F d+1 . Thus, [d+1], each ball that has been assigned to group F up to (and including) ball , has definitely failed to hit an unsaturated bin in all the groups F 1 , . . . , F -1 . For any ball t m and [d], let Q t () denote the number of balls that have been assigned to group F up to time t (ie, right after the assignment of the t th ball), while ~ Q t () denotes the balls that have been assigned to the right of group F , that is, to bins of [n] \ =1 F . Thus, ~ Q t () = d+1 =+1 Q t (). Let also S t () denote the set of saturated bins in F at time t. Then, [d], t , IP[t hits a saturated bin in F ] = C[S t-1 ()] C[F ] C[S ()] C[F ] . Observe now that [d], Q () = iF q (i) iS ( ) C i q (i) C i &gt; C[S ()](a) C[S ()] C[F ] &lt; Q () (a) C[F ] P (1) 130 Recall that up to time , we can be sure that Q () (a) C[F ] (because all these balls are assigned to unsaturated bins), which in turn assures that P 1. Inequality (1) implies that, had we known ~ Q (-1), then the number ~ Q () of balls before ball + 1 that go to the right of set F would be stochastically dominated by the number of successes in ~ Q (-1) Bernoulli trials with success probability P . We shall denote this number by B( ~ Q ( - 1), P ). 5 In the following lemma, we apply the Chernoff-Hoeffding bound on these Bernoulli trials to get an upper bound on the amount of balls that have to be assigned beyond group F , for [d], as a function of the number of balls that have been assigned beyond group F -1 : Lemma 4.1. [d] and for an arbitrary constant &gt; 1, with probability at least 1 - n ~ Q () max 2 +1 [ ~ Q ( - 1)] 2 (a)n , 2 ln n ~ Q ( - 1) . Proof: Let's assume that we already know the number ~ Q ( - 1) of balls that have already failed in F 1 , . . . , F -1 . Then ~ Q () is stochastically dominated by the random variable B( ~ Q ( - 1), P ). By applying Chernoff-Hoeffding bounds ([11], p. 198) on these Bernoulli trials, we get that IP[B( ~ Q ( - 1), P ) ~ Q ( - 1) (P + )] exp(-2 ~ Q ( - 1) 2 ), &gt; 0 IP B( ~ Q ( - 1), P ) ~ Q ( - 1) P + ln n 2 ~ Q ( -1) n , &gt; 0 where we have set = ln n 2 ~ Q ( -1) . But recall that P = Q ( ) ( a)C[F ] and also that C[F ] n 2 . Thus we conclude that with probability at least 1 - n , ~ Q () B( ~ Q ( - 1), P ) ~ Q ( - 1) 2 Q ( ) ( a)n + ln n 2 ~ Q ( -1) Q () = ~ Q ( - 1) - ~ Q () ~ Q () ~ Q ( - 1) 2 [ ~ Q ( - 1) - ~ Q ()] (a)n + ln n 2 ~ Q ( - 1) ~ Q () 1 + 2 ~ Q ( - 1) (a)n ~ Q ( - 1) 2 ~ Q ( - 1) (a)n + ln n 2 ~ Q ( - 1) ~ Q () 2 [ ~ Q ( - 1)] 2 + (a)n ln n 2 ~ Q ( - 1) (a)n + 2 ~ Q ( - 1) from which we get the desired bound. Consider now the following finite sequence ^ Q() = max 2 +1 [ ^ Q(-1)] 2 ( a)n , 2 ln n ^ Q ( - 1) , [d] ^ Q(0) = m 5 For the integrity of the representation we set ~ Q (0) = m - 1. We then bound the number of balls that end up in group F d+1 by the d th term of this sequence: Lemma 4.2. With probability at least 1 - dn , at most ^ Q(d) balls end up in group F d+1 . Proof: The proof of this lemma is relatively simple and due to lack of space it is differed to the full version of the paper. Consequently we estimate a closed form for the first terms of the bounding sequence ^ Q(), = 0, . . . , d that was determined earlier: Lemma 4.3. The first log log m ln n log( a/8) +2 terms of the bounding sequence ^ Q(), = 0, . . . , d are given by the closed form ^ Q() = m 2 ( a 8 ) 2-1 . Proof: Assume that was the first element in the sequence for which 2 +1 [ ^ Q( - 1)] 2 (a)n &lt; 2 ln n ^ Q( - 1) (2) Then, up to term - 1 the sequence is dominated by the right-hand term of the above inequality, and thus, r &lt; , ^ Q(r) = 2 r+1 (a)n [ ^ Q(r - 1)] 2 = 2 r+1 (a)n 2 r (a)n 2 [ ^ Q(r - 2)] 4 = 2 r+1 (a)n 2 0 2 r (a)n 2 1 2 r-1 (a)n 2 2 [ ^ Q(r - 3)] 2 3 = = r =1 2 2 r+2 (a)n 2 -1 [ ^ Q(0)] 2 r = 2 r =1 ( r+2-)2 -1 1 (a)n r =1 2 -1 [ ^ Q(0)] 2 r = 2 3 2 r -r-3 ( a)n m 2 r -1 m m 2 r a 8 2 r -1 since ( a)n 8 m a 8 . We now plug in this closed form for ^ Q( 1 ) in inequality (2) to get the following: 2 ln n &gt; 2 +1 (a)n m 2 -1 a8 2 -1 -1 3 /2 m 3 /2 2 lnn &lt; (a)n 2 +1 2 3 2 ( -1) (a)n 8m 2 -1 -1 m 2 ln n &lt; 2 1 2 ( +3) a 8 2 -1 From the above and the definition of , it is easy to see that -1 = max r [d] : ln m-ln -lnln n ln 2 - 4 r + 2 r-2 ln a 8 . By setting A = log m lnln n+ln ln 2 - 4 and B = ln a 8 we get the solution A W [ B ln 2 4 exp( A ln 2) ] ln 2 + 1, where W (x) is the Lambert W Function ([7]). By approximating this function by ln x - lnln x (since x = B ln 2 4 exp(A ln 2) ) we conclude that log log m ln n log(a/8) + 3 131 Lemma 4.4. Assume that m n and d &lt; lnln n-lnlnln n (2+1) ln 2 . If we set a = 8 n d2 d+3 1 /(2 d+1 -1) the cost of OBL is upper bounded (whp) by OBL (m) (1 + o(1))(a + 1)ADV(m). Proof: The cost of OBL up to time is upper-bounded by max {L max [F d+1 ], (a + 1) opt} since in the first d groups no saturated bin ever gets another ball and all the bins are legitimate. We choose a so that the cost in the first d groups is at least as large as L max [F d+1 ] (whp), given the upper bound ^ Q(d) on the number of balls that end up in the last group. Thus the probability of + 1 existing will be then polynomially small because at least one poll in F d+1 will have to be unsaturated whp. This then implies the claimed upper bound on the performance of OBL . Due to lack of space, the complete proof of this lemma is presented in the full version of the paper. Combining the statements of all these technical lemmas, we conclude with the desired bound on the competitive ratio of OBL . A COMPETITIVE HAPS SCHEDULER In the previous section we have proposed a hops scheduler that is based on the knowledge of the size of the input sequence and then assures that its own performance is never worse than (a + 1)opt, whp. In this section we propose a haps scheduler (call it ADAPT ) whose main purpose is to "guess" the value opt of the optimum offline cost by a classical guessing argument and then let OBL do the rest of the assignments. This approach is in complete analogy with the online schedulers of the Related Machines-Full Information problem (see [6], pp. 208-210). The only difference is that OBL has a performance which holds whp, and this is why the final result also holds whp. A significant difficulty of ADAPT is exactly this guessing mechanism that will have to be based on the limited information provided to each of the new tasks. Our goal is not to assume that any kind of additional information (eg, global environment variables) is provided to the balls, other than the capacity vector and the current loads of each ball's candidates . ADAPT sacrifices one of the available polls per ball, in order to create such a good guessing mechanism. Of course, a different approach that would be based on the outcome of some of the polls (eg, a constant fraction of the polls) in order to estimate a proper online prediction of opt would be more interesting in the sense that it would not create a communication bottleneck for the tasks. Nevertheless , the purpose of ADAPT is mainly to demonstrate the possibility of constructing such an adaptive scheduler whose performance is close to that of OBL . Let's assume that the system now has n + 1 capacitated bins ( C min = C 1 C 2 C n+1 = C max ). Assume also that each new task is provided with 2d + 1 polls. Fix a number r &gt; 1. Then an r-guessing mechanism proceeds in stages: Stage ( = 0, 1, . . . ) contains a (consecutive) subsequence of tasks from the input sequence that use the same prediction = r 0 for their assignment. We set 0 = 1 C max . The following definitions refer to the notions of eligible and saturated machines upon the arrival of a new task t [m] into the system, that are used in a similar fashion as in [5] where the concept of eligible-saturated machines was introduced: Definition 5.1. Suppose that task t m belongs to stage .The set of eligible machines for t is E i [n] : 1 C i = r 0 , while a machine i [n] is considered to be saturated upon the arrival of task t, if t-1 (i) &gt; a = r a 0 , where a = 8 max 1, |E | d2 d+3 1 /(2 d+1 -1) . Notice that the static information of (E , a ), = 0, 1, . . . only depends on the capacity vector c and the number d of polls per task. Thus it can be easily computed by each of the tasks, or alternatively it can be provided a priori to all the tasks as additional offline information. ADAPT proceeds in phases and works as follows: Upon the arrival of a new ball t [m], first the fastest bin in the system is polled and the stage s(t) to which this ball belongs is determined, according to the following rule: s(t) = IN : r -1 a -1 + 1 &lt; q t-1 (n + 1) r a + 1 (obviously for stage 0 only the second inequality must hold). The remaining 2d polls of task t are taken from group E s(t) in a fashion similar to that of OBL . The assignment strategy of ADAPT is exactly the same with that of OBL , with the only difference that whenever the first 2d candidates of a task are already saturated, then this task is assigned to the fastest machine in the system, n + 1. If the latter event causes machine n + 1 to become also saturated, then by the definition of the stages this task is the last ball of the current stage and a new stage (with an r times larger prediction) starts from the next task on. Lemma 5.1. Suppose that ADAPT uses r = 9/4.Let h denote the stage at which the prediction h of ADAPT reaches opt for the first time.Then the last stage of ADAPT is at most h + 1. Proof: The case of h = 0 is easy since this implies that opt = 1/C max and then ADAPT cannot be worse than OBL which (having the right prediction) succeeds to assign all the incoming tasks (but for those that might fit in machine n +1 without making it saturated) below (a 0 + 1) opt, whp. So let's consider the case where h 1. Let E denote the set of legitimate machines in the system (ie, E = {i [n + 1] : 1/C i opt}). The amount of work inserted into the system during the whole assignment process is bounded by C[E ] opt. By definition of h, r h-1 /C max &lt; opt r h /C max . Stage h + 1 ends when machine n + 1 becomes saturated. Let W () denote the total amount of work assigned during stage , while R() denotes the amount of remaining work at the end of stage . We shall prove that the amount of remaining work at the end of stage h is not enough to make OBL fail within stage h + 1. Observe that at the end of stage h, no machine has exceeded a load of (a h + 1) h = (a h + 1)r h /C max (by definition of stages). On the other hand, during stage h + 1, each of the eligible machines i E h+1 needs a load of more than a h+1 h+1 = a h+1 r h+1 /C max in order to become saturated for this stage. We denote the additional free space of bin 132 i E h+1 by f ree i (h+1) &gt; a h+1 r h+1 C max ( a h +1) r h C max C i , while F REE(h + 1) iE h+1 f ree i (h + 1) &gt; a h+1 r h+1 C max - (a h + 1)r h C max C[E h+1 ] &gt; a h+1 r h+1 C max 1 - a h + 1 r a h+1 C[E h+1 ] &gt; (a h+1 + 1)optC[E h+1 ] ra h+1 a h+1 + 1 a h + 1 a h+1 + 1 (3) is the cumulative free space granted to the eligible bins of the system for stage h + 1. It only remains to prove that the amount of work that has to be dealt with by OBL using only the bins of [n] is less than F REE(h+1)/(a h+1 +1). Observe now that OBL uses the bins of E h+1 and has to deal with an amount of work less than W (h + 1) - f ree n+1 (h + 1) &lt; R(h)-a h+1 r h C max r a h +1 a h+1 C max &lt; C[E ]opt -a h+1 (r-2 )opt C max &lt; opt(C[E ] -(r-2)a h+1 C max ) &lt; (C[E ] -C max ) opt if we set r 17/8. This is because a h+1 a h 8 and at the end of stage h+1 bin n+1 must have already become saturated. Observe also that E h+1 E h E \ {n + 1} by definition of h. Thus, C[E h+1 ] C[E ] - C max . Thus, the amount of work that has to be served by OBL during stage h + 1 is less than opt C[E h+1 ]. Then, setting r = 9/4 in inequality (3) assures that ra h+1 a h+1 +1 a h +1 a h+1 +1 1 (recall that a h+1 a h 8) and thus the remaining work at the end of stage h is not enough to make OBL fail. The following theorem is now a straightforward consequence of the previous lemma: Theorem 3. For any input sequence of identical tasks that have to be assigned to n related machines using at most 2d + 1 polls per task, the cost of ADAPT is (whp), ADAPT () &lt; O n d2 d+3 1 /(2 d+1 -1) + 1 ADV(). CONCLUSIONS In this work we have studied the problem of exploiting limited online information for the assignment of a sequence of unit-size tasks to related machines. We have shown that the oblivious schedulers that perform asymptotically optimally in the case of identical machines, deteriorate significantly in this case. We have then determined an adaptive scheduler that actually mimics the behaviour of an ideal oblivious scheduler, in order to achieve roughly the asymptotically optimal performance similar to the case of the identical machines . As for further research, the issue of providing only limited information to online algorithms is critical in many problems for which an objective is also the minimization of the communication cost. In this category fall most of the network design problems. Another interesting line of research would be the avoidance of communication bottlenecks for such limited information online algorithms. Additionally, it would be very interesting to study the case of tasks of arbitrary sizes being injected into the system. ACKNOWLEDGMENTS The author wishes to thank Paul Spirakis for valuable discussions during the write-up of the paper and also for suggesting an appropriate terminology on the categorization of polling strategies. REFERENCES [1] J. Aspens, Y. Azar, A. Fiat, S. Plotkin, O. Waarts. Online Machine Scheduling with Applications to Load Balancing and Virtual Circuit Routing. In Journal of ACM, Vol. 44 (1997), pp. 486-504. [2] Y. Azar, A. Broder, A. Karlin, E. Upfal. Ballanced Allocations. In Proc. of 26th ACM-STOC (1994), pp. 593-602. Also in SIAM J. Computing, 29 (1999), pp. 180-200. [3] P. Berman, M. Charikar, M. Karpinski. Online Load Balancing for Related Machines. In Proc. of 5th Int. Workshop on Algorithms and Data Structures (1997), LNCS 1272, Springer-Verlag 1997, pp. 116-125. [4] P. Berenbrink, A. Czumaj, A. Steger, and B. V ocking. Balanced Allocations: The Heavily Loaded Case. In Proc. of 32 nd ACM-STOC (Portland, 2000), pp. 745-754. [5] A. Bar-Noy, A. Freund, J. Naor. NewAlgorithms for Related Machines with Temporary Jobs. In Journal of Scheduling, Vol. 3 (2000), pp. 259-272. [6] A. Borodin, R. El Yaniv. Online Computation and Competitive Analysis. Cambridge Univ. Press 1998. [7] R. Corless, G. Gonnet, D. Hare, D. Jeffrey, D. Knuth. On the Lambert W Function. In Advances in Computational Mathematics, Vol. 5 (1996), pp. 329-359. [8] A. Czumaj and B. V ocking. Tight Bounds for Worst-Case Equilibria. In Proc. of 13 th ACM-SIAM SODA (San Francisco, 2002), pp. 413-420. [9] Y. Azar. Online Load Balancing. In Online Algorithms: The State of the Art, A. Fiat, G. Woeginger, (Eds.). LNCS 1442, Springer 1998, pp. 178-195. [10] G. Gonnet. Expected Length of the Longest Probe Sequence in Hash Code Searching. In J. of ACM, Vol. 28, No. 2 (1981), pp. 289-304. [11] M. Habib, C. McDiarmid, J. Ramirez-Alfonsin, B. Reed (Eds.). Probabilistic Methods for Algorithmic Discrete Mathematics. ISBN: 3-540-64622-1, Springer-Velrag 1998. [12] E. Koutsoupias, C. Papadimitriou. Worst-case Equilibria. In Proc. of 16 th Annual Symposium on Theoretical Aspects of Computer Science (STACS), LNCS 1563, Springer-Verlag 1999, pp. 404-413. [13] M. Mavronicolas, P. Spirakis. The Price of Selfish Routing. In Proc. of 33 rd ACM-STOC (Krete, 2001), pp. 510-519. [14] M. Mitzenmacher, A. Richa, R. Sitaraman. The Power of Two Random Choices: A Survey of Techniques and Results. In Handbook of Randomized Algorithms (to appear). Also available through http://www.eecs.harvard.edu/ michaelm/. [15] A. Steger, M. Raab. "Balls into Bins" - A Simple and Tight Analysis. In Proc. of RANDOM'98, LNCS 1518, Springer Verlag, 1998, pp. 159-170. [16] J. Sgall. Online Scheduling. In Online Algorithms: The State of the Art, A. Fiat, G. Woeginger (Eds.). LNCS 1442, Springer 1998, pp. 196-231. [17] D. Shmoys, J. Wein, D. Williamson. Scheduling Parallel Machines Online. In Proc. of the 32 nd IEEE-FOCS (1991), pp. 131-140. [18] B. V ocking. HowAsymmetry Helps Load Balancing. In Proc. of 40 th IEEE-FOCS (NewYork, 1999), pp. 131-140. 133

oblivious scheduler;HOPS;related machines;Limited Information;lower bounds;online information;scheduling;Online Load Balancing;input sequence;unit-size task;Related Machines

132

Machine Learning for Information Architecture in a Large Governmental Website

This paper describes ongoing research into the application of machine learning techniques for improving access to governmental information in complex digital libraries. Under the auspices of the GovStat Project, our goal is to identify a small number of semantically valid concepts that adequately spans the intellectual domain of a collection. The goal of this discovery is twofold. First we desire a practical aid for information architects. Second, automatically derived document-concept relationships are a necessary precondition for real-world deployment of many dynamic interfaces. The current study compares concept learning strategies based on three document representations: keywords, titles, and full-text. In statistical and user-based studies, human-created keywords provide significant improvements in concept learning over both title-only and full-text representations. Categories and Subject Descriptors

INTRODUCTION The GovStat Project is a joint effort of the University of North Carolina Interaction Design Lab and the University of Maryland Human-Computer Interaction Lab 1 . Citing end-user difficulty in finding governmental information (especially statistical data) online, the project seeks to create an integrated model of user access to US government statistical information that is rooted in realistic data models and innovative user interfaces. To enable such models and interfaces, we propose a data-driven approach, based on data mining and machine learning techniques. In particular , our work analyzes a particular digital library--the website of the Bureau of Labor Statistics 2 (BLS)--in efforts to discover a small number of linguistically meaningful concepts , or "bins," that collectively summarize the semantic domain of the site. The project goal is to classify the site's web content according to these inferred concepts as an initial step towards data filtering via active user interfaces (cf. [13]). Many digital libraries already make use of content classification, both explicitly and implicitly; they divide their resources manually by topical relation; they organize content into hi-erarchically oriented file systems. The goal of the present 1 http://www.ils.unc.edu/govstat 2 http://www.bls.gov 151 research is to develop another means of browsing the content of these collections. By analyzing the distribution of terms across documents, our goal is to supplement the agency's pre-existing information structures. Statistical learning technologies are appealing in this context insofar as they stand to define a data-driven--as opposed to an agency-driven-navigational structure for a site. Our approach combines supervised and unsupervised learning techniques. A pure document clustering [12] approach to such a large, diverse collection as BLS led to poor results in early tests [6]. But strictly supervised techniques [5] are inappropriate, too. Although BLS designers have defined high-level subject headings for their collections, as we discuss in Section 2, this scheme is less than optimal. Thus we hope to learn an additional set of concepts by letting the data speak for themselves. The remainder of this paper describes the details of our concept discovery efforts and subsequent evaluation. In Section 2 we describe the previously existing, human-created conceptual structure of the BLS website. This section also describes evidence that this structure leaves room for improvement . Next (Sections 35), we turn to a description of the concepts derived via content clustering under three document representations: keyword, title only, and full-text. Section 6 describes a two-part evaluation of the derived conceptual structures. Finally, we conclude in Section 7 by outlining upcoming work on the project. STRUCTURING ACCESS TO THE BLS WEBSITE The Bureau of Labor Statistics is a federal government agency charged with compiling and publishing statistics pertaining to labor and production in the US and abroad. Given this broad mandate, the BLS publishes a wide array of information , intended for diverse audiences. The agency's website acts as a clearinghouse for this process. With over 15,000 text/html documents (and many more documents if spreadsheets and typeset reports are included), providing access to the collection provides a steep challenge to information architects. 2.1 The Relation Browser The starting point of this work is the notion that access to information in the BLS website could be improved by the addition of a dynamic interface such as the relation browser described by Marchionini and Brunk [13]. The relation browser allows users to traverse complex data sets by iteratively slicing the data along several topics. In Figure 1 we see a prototype instantiation of the relation browser, applied to the FedStats website 3 . The relation browser supports information seeking by allowing users to form queries in a stepwise fashion, slicing and re-slicing the data as their interests dictate. Its motivation is in keeping with Shneiderman's suggestion that queries and their results should be tightly coupled [2]. Thus in Fig-3 http://www.fedstats.gov Figure 1: Relation Browser Prototype ure 1, users might limit their search set to those documents about "energy." Within this subset of the collection, they might further eliminate documents published more than a year ago. Finally, they might request to see only documents published in PDF format. As Marchionini and Brunk discuss, capturing the publication date and format of documents is trivial. But successful implementations of the relation browser also rely on topical classification. This presents two stumbling blocks for system designers: Information architects must define the appropriate set of topics for their collection Site maintainers must classify each document into its appropriate categories These tasks parallel common problems in the metadata community: defining appropriate elements and marking up documents to support metadata-aware information access. Given a collection of over 15,000 documents, these hurdles are especially daunting, and automatic methods of approaching them are highly desirable. 2.2 A Pre-Existing Structure Prior to our involvement with the project, designers at BLS created a shallow classificatory structure for the most important documents in their website. As seen in Figure 2, the BLS home page organizes 65 "top-level" documents into 15 categories. These include topics such as Employment and Unemployment, Productivity, and Inflation and Spending. 152 Figure 2: The BLS Home Page We hoped initially that these pre-defined categories could be used to train a 15-way document classifier, thus automating the process of populating the relation browser altogether. However, this approach proved unsatisfactory. In personal meetings, BLS officials voiced dissatisfaction with the existing topics. Their form, it was argued, owed as much to the institutional structure of BLS as it did to the inherent topology of the website's information space. In other words, the topics reflected official divisions rather than semantic clusters. The BLS agents suggested that re-designing this classification structure would be desirable. The agents' misgivings were borne out in subsequent analysis . The BLS topics comprise a shallow classificatory structure ; each of the 15 top-level categories is linked to a small number of related pages. Thus there are 7 pages associated with Inflation. Altogether, the link structure of this classificatory system contains 65 documents; that is, excluding navigational links, there are 65 documents linked from the BLS home page, where each hyperlink connects a document to a topic (pages can be linked to multiple topics). Based on this hyperlink structure, we defined M, a symmetric 65 65 matrix, where m ij counts the number of topics in which documents i and j are both classified on the BLS home page. To analyze the redundancy inherent in the pre-existing structure , we derived the principal components of M (cf. [11]). Figure 3 shows the resultant scree plot 4 . Because all 65 documents belong to at least one BLS topic, 4 A scree plot shows the magnitude of the k th eigenvalue versus its rank. During principal component analysis scree plots visualize the amount of variance captured by each component . 0 10 20 30 40 50 60 0 2 4 6 8 10 12 14 Eigenvalue Rank Eigenvlue Magnitude Figure 3: Scree Plot of BLS Categories the rank of M is guaranteed to be less than or equal to 15 (hence, eigenvalues 16 . . . 65 = 0). What is surprising about Figure 3, however, is the precipitous decline in magnitude among the first four eigenvalues. The four largest eigenvlaues account for 62.2% of the total variance in the data. This fact suggests a high degree of redundancy among the topics. Topical redundancy is not in itself problematic. However, the documents in this very shallow classificatory structure are almost all gateways to more specific information . Thus the listing of the Producer Price Index under three categories could be confusing to the site's users. In light of this potential for confusion and the agency's own request for redesign, we undertook the task of topic discovery described in the following sections. A HYBRID APPROACH TO TOPIC DISCOVERY To aid in the discovery of a new set of high-level topics for the BLS website, we turned to unsupervised machine learning methods. In efforts to let the data speak for themselves, we desired a means of concept discovery that would be based not on the structure of the agency, but on the content of the material. To begin this process, we crawled the BLS website , downloading all documents of MIME type text/html. This led to a corpus of 15,165 documents. Based on this corpus, we hoped to derive k 10 topical categories, such that each document d i is assigned to one or more classes. 153 Document clustering (cf. [16]) provided an obvious, but only partial solution to the problem of automating this type of high-level information architecture discovery. The problems with standard clustering are threefold. 1. Mutually exclusive clusters are inappropriate for identifying the topical content of documents, since documents may be about many subjects. 2. Due to the heterogeneity of the data housed in the BLS collection (tables, lists, surveys, etc.), many doc-uments' terms provide noisy topical information. 3. For application to the relation browser, we require a small number (k 10) of topics. Without significant data reduction, term-based clustering tends to deliver clusters at too fine a level of granularity. In light of these problems, we take a hybrid approach to topic discovery. First, we limit the clustering process to a sample of the entire collection, described in Section 4. Working on a focused subset of the data helps to overcome problems two and three, listed above. To address the problem of mutual exclusivity, we combine unsupervised with supervised learning methods, as described in Section 5. FOCUSING ON CONTENT-RICH DOCUMENTS To derive empirically evidenced topics we initially turned to cluster analysis. Let A be the np data matrix with n observations in p variables. Thus a ij shows the measurement for the i th observation on the j th variable. As described in [12], the goal of cluster analysis is to assign each of the n observations to one of a small number k groups, each of which is characterized by high intra-cluster correlation and low inter-cluster correlation. Though the algorithms for accomplishing such an arrangement are legion, our analysis focuses on k-means clustering 5 , during which, each observation o i is assigned to the cluster C k whose centroid is closest to it, in terms of Euclidean distance. Readers interested in the details of the algorithm are referred to [12] for a thorough treatment of the subject. Clustering by k-means is well-studied in the statistical literature, and has shown good results for text analysis (cf. [8, 16]). However, k-means clustering requires that the researcher specify k, the number of clusters to define. When applying k-means to our 15,000 document collection, indicators such as the gap statistic [17] and an analysis of the mean-squared distance across values of k suggested that k 80 was optimal. This paramterization led to semantically intelligible clusters. However, 80 clusters are far too many for application to an interface such as the relation 5 We have focused on k-means as opposed to other clustering algorithms for several reasons. Chief among these is the computational efficiency enjoyed by the k-means approach. Because we need only a flat clustering there is little to be gained by the more expensive hierarchical algorithms. In future work we will turn to model-based clustering [7] as a more principled method of selecting the number of clusters and of representing clusters. browser. Moreover, the granularity of these clusters was un-suitably fine. For instance, the 80-cluster solution derived a cluster whose most highly associated words (in terms of log-odds ratio [1]) were drug, pharmacy, and chemist. These words are certainly related, but they are related at a level of specificity far below what we sought. To remedy the high dimensionality of the data, we resolved to limit the algorithm to a subset of the collection. In consultation with employees of the BLS, we continued our analysis on documents that form a series titled From the Editor's Desk 6 . These are brief articles, written by BLS employees. BLS agents suggested that we focus on the Editor's Desk because it is intended to span the intellectual domain of the agency. The column is published daily, and each entry describes an important current issue in the BLS domain. The Editor's Desk column has been written daily (five times per week) since 1998. As such, we operated on a set of N = 1279 documents. Limiting attention to these 1279 documents not only reduced the dimensionality of the problem. It also allowed the clustering process to learn on a relatively clean data set. While the entire BLS collection contains a great deal of non-prose text (i.e. tables, lists, etc.), the Editor's Desk documents are all written in clear, journalistic prose. Each document is highly topical, further aiding the discovery of term-topic relations. Finally, the Editor's Desk column provided an ideal learning environment because it is well-supplied with topical metadata. Each of the 1279 documents contains a list of one or more keywords. Additionally, a subset of the documents (1112) contained a subject heading. This metadata informed our learning and evaluation, as described in Section 6.1. COMBINING SUPERVISED AND UNSUPERVISED LEARNING FOR TOPIC DISCOVERY To derive suitably general topics for the application of a dynamic interface to the BLS collection, we combined document clustering with text classification techniques. Specif-ically , using k-means, we clustered each of the 1279 documents into one of k clusters, with the number of clusters chosen by analyzing the within-cluster mean squared distance at different values of k (see Section 6.1). Constructing mutually exclusive clusters violates our assumption that documents may belong to multiple classes. However, these clusters mark only the first step in a two-phase process of topic identification. At the end of the process, document-cluster affinity is measured by a real-valued number. Once the Editor's Desk documents were assigned to clusters , we constructed a k-way classifier that estimates the strength of evidence that a new document d i is a member of class C k . We tested three statistical classification techniques : probabilistic Rocchio (prind), naive Bayes, and support vector machines (SVMs). All were implemented using McCallum's BOW text classification library [14]. Prind is a probabilistic version of the Rocchio classification algorithm [9]. Interested readers are referred to Joachims' article for 6 http://www.bls.gov/opub/ted 154 further details of the classification method. Like prind, naive Bayes attempts to classify documents into the most probable class. It is described in detail in [15]. Finally, support vector machines were thoroughly explicated by Vapnik [18], and applied specifically to text in [10]. They define a decision boundary by finding the maximally separating hyperplane in a high-dimensional vector space in which document classes become linearly separable. Having clustered the documents and trained a suitable classifier, the remaining 14,000 documents in the collection are labeled by means of automatic classification. That is, for each document d i we derive a k-dimensional vector, quantifying the association between d i and each class C 1 . . . C k . Deriving topic scores via naive Bayes for the entire 15,000-document collection required less than two hours of CPU time. The output of this process is a score for every document in the collection on each of the automatically discovered topics. These scores may then be used to populate a relation browser interface, or they may be added to a traditional information retrieval system. To use these weights in the relation browser we currently assign to each document the two topics on which it scored highest. In future work we will adopt a more rigorous method of deriving document-topic weight thresholds. Also, evaluation of the utility of the learned topics for users will be undertaken. EVALUATION OF CONCEPT DISCOVERY Prior to implementing a relation browser interface and undertaking the attendant user studies, it is of course important to evaluate the quality of the inferred concepts, and the ability of the automatic classifier to assign documents to the appropriate subjects. To evaluate the success of the two-stage approach described in Section 5, we undertook two experiments. During the first experiment we compared three methods of document representation for the clustering task. The goal here was to compare the quality of document clusters derived by analysis of full-text documents, documents represented only by their titles, and documents represented by human-created keyword metadata. During the second experiment, we analyzed the ability of the statistical classifiers to discern the subject matter of documents from portions of the database in addition to the Editor's Desk. 6.1 Comparing Document Representations Documents from The Editor's Desk column came supplied with human-generated keyword metadata. Additionally , The titles of the Editor's Desk documents tend to be germane to the topic of their respective articles. With such an array of distilled evidence of each document's subject matter, we undertook a comparison of document representations for topic discovery by clustering. We hypothesized that keyword-based clustering would provide a useful model. But we hoped to see whether comparable performance could be attained by methods that did not require extensive human indexing, such as the title-only or full-text representations . To test this hypothesis, we defined three modes of document representation--full-text, title-only, and keyword only--we generated three sets of topics, T f ull , T title , and T kw , respectively. Topics based on full-text documents were derived by application of k-means clustering to the 1279 Editor's Desk documents , where each document was represented by a 1908-dimensional vector. These 1908 dimensions captured the TF.IDF weights [3] of each term t i in document d j , for all terms that occurred at least three times in the data. To arrive at the appropriate number of clusters for these data, we inspected the within-cluster mean-squared distance for each value of k = 1 . . . 20. As k approached 10 the reduction in error with the addition of more clusters declined notably, suggesting that k 10 would yield good divisions. To select a single integer value, we calculated which value of k led to the least variation in cluster size. This metric stemmed from a desire to suppress the common result where one large cluster emerges from the k-means algorithm, accompanied by several accordingly small clusters. Without reason to believe that any single topic should have dramatically high prior odds of document membership, this heuristic led to k f ull = 10. Clusters based on document titles were constructed simi-larly . However, in this case, each document was represented in the vector space spanned by the 397 terms that occur at least twice in document titles. Using the same method of minimizing the variance in cluster membership k title the number of clusters in the title-based representationwas also set to 10. The dimensionality of the keyword-based clustering was very similar to that of the title-based approach. There were 299 keywords in the data, all of which were retained. The median number of keywords per document was 7, where a keyword is understood to be either a single word, or a multi-word term such as "consumer price index." It is worth noting that the keywords were not drawn from any controlled vocabulary ; they were assigned to documents by publishers at the BLS. Using the keywords, the documents were clustered into 10 classes. To evaluate the clusters derived by each method of document representation, we used the subject headings that were included with 1112 of the Editor's Desk documents. Each of these 1112 documents was assigned one or more subject headings, which were withheld from all of the cluster applications . Like the keywords, subject headings were assigned to documents by BLS publishers. Unlike the keywords, however , subject headings were drawn from a controlled vocabulary . Our analysis began with the assumption that documents with the same subject headings should cluster together . To facilitate this analysis, we took a conservative approach; we considered multi-subject classifications to be unique. Thus if document d i was assigned to a single subject prices, while document d j was assigned to two subjects, international comparisons, prices, documents d i and d j are not considered to come from the same class. Table 1 shows all Editor's Desk subject headings that were assigned to at least 10 documents. As noted in the table, 155 Table 1: Top Editor's Desk Subject Headings Subject Count prices 92 unemployment 55 occupational safety & health 53 international comparisons, prices 48 manufacturing, prices 45 employment 44 productivity 40 consumer expenditures 36 earnings & wages 27 employment & unemployment 27 compensation costs 25 earnings & wages, metro. areas 18 benefits, compensation costs 18 earnings & wages, occupations 17 employment, occupations 14 benefits 14 earnings & wage, regions 13 work stoppages 12 earnings & wages, industries 11 Total 609 Table 2: Contingecy Table for Three Document Representations Representation Right Wrong Accuracy Full-text 392 217 0.64 Title 441 168 0.72 Keyword 601 8 0.98 there were 19 such subject headings, which altogether covered 609 (54%) of the documents with subjects assigned. These document-subject pairings formed the basis of our analysis. Limiting analysis to subjects with N &gt; 10 kept the resultant 2 tests suitably robust. The clustering derived by each document representation was tested by its ability to collocate documents with the same subjects. Thus for each of the 19 subject headings in Table 1, S i , we calculated the proportion of documents assigned to S i that each clustering co-classified. Further, we assumed that whichever cluster captured the majority of documents for a given class constituted the "right answer" for that class. For instance, There were 92 documents whose subject heading was prices. Taking the BLS editors' classifications as ground truth, all 92 of these documents should have ended up in the same cluster. Under the full-text representation 52 of these documents were clustered into category 5, while 35 were in category 3, and 5 documents were in category 6. Taking the majority cluster as the putative right home for these documents, we consider the accuracy of this clustering on this subject to be 52/92 = 0.56. Repeating this process for each topic across all three representations led to the contingency table shown in Table 2. The obvious superiority of the keyword-based clustering evidenced by Table 2 was borne out by a 2 test on the accuracy proportions. Comparing the proportion right and Table 3: Keyword-Based Clusters benefits costs international jobs plans compensation import employment benefits costs prices jobs employees benefits petroleum youth occupations prices productivity safety workers prices productivity safety earnings index output health operators inflation nonfarm occupational spending unemployment expenditures unemployment consumer mass spending jobless wrong achieved by keyword and title-based clustering led to p 0.001. Due to this result, in the remainder of this paper, we focus our attention on the clusters derived by analysis of the Editor's Desk keywords. The ten keyword-based clusters are shown in Table 3, represented by the three terms most highly associated with each cluster, in terms of the log-odds ratio. Additionally, each cluster has been given a label by the researchers. Evaluating the results of clustering is notoriously difficult. In order to lend our analysis suitable rigor and utility, we made several simplifying assumptions. Most problematic is the fact that we have assumed that each document belongs in only a single category. This assumption is certainly false. However, by taking an extremely rigid view of what constitutes a subject--that is, by taking a fully qualified and often multipart subject heading as our unit of analysis--we mitigate this problem. Analogically, this is akin to considering the location of books on a library shelf. Although a given book may cover many subjects, a classification system should be able to collocate books that are extremely similar, say books about occupational safety and health. The most serious liability with this evaluation, then, is the fact that we have compressed multiple subject headings, say prices : international into single subjects. This flattening obscures the multivalence of documents. We turn to a more realistic assessment of document-class relations in Section 6.2. 6.2 Accuracy of the Document Classifiers Although the keyword-based clusters appear to classify the Editor's Desk documents very well, their discovery only solved half of the problem required for the successful implementation of a dynamic user interface such as the relation browser. The matter of roughly fourteen thousand unclassified documents remained to be addressed. To solve this problem, we trained the statistical classifiers described above in Section 5. For each document in the collection d i , these classifiers give p i , a k-vector of probabilities or distances (depending on the classification method used), where p ik quantifies the strength of association between the i th document and the k th class. All classifiers were trained on the full text of each document, regardless of the representation used to discover the initial clusters. The different training sets were thus constructed simply by changing the 156 Table 4: Cross Validation Results for 3 Classifiers Method Av. Percent Accuracy SE Prind 59.07 1.07 Naive Bayes 75.57 0.4 SVM 75.08 0.68 class variable for each instance (document) to reflect its assigned cluster under a given model. To test the ability of each classifier to locate documents correctly, we first performed a 10-fold cross validation on the Editor's Desk documents. During cross-validation the data are split randomly into n subsets (in this case n = 10). The process proceeds by iteratively holding out each of the n subsets as a test collection for a model trained on the remaining n - 1 subsets. Cross validation is described in [15]. Using this methodology, we compared the performance of the three classification models described above. Table 4 gives the results from cross validation. Although naive Bayes is not significantly more accurate for these data than the SVM classifier, we limit the remainder of our attention to analysis of its performance. Our selection of naive Bayes is due to the fact that it appears to work comparably to the SVM approach for these data, while being much simpler, both in theory and implementation. Because we have only 1279 documents and 10 classes, the number of training documents per class is relatively small. In addition to models fitted to the Editor's Desk data, then, we constructed a fourth model, supplementing the training sets of each class by querying the Google search engine 7 and applying naive Bayes to the augmented test set. For each class, we created a query by submitting the three terms with the highest log-odds ratio with that class. Further, each query was limited to the domain www.bls.gov. For each class we retrieved up to 400 documents from Google (the actual number varied depending on the size of the result set returned by Google). This led to a training set of 4113 documents in the "augmented model," as we call it below 8 . Cross validation suggested that the augmented model decreased classification accuracy (accuracy= 58.16%, with standard error= 0.32). As we discuss below, however, augmenting the training set appeared to help generalization during our second experiment. The results of our cross validation experiment are encouraging . However, the success of our classifiers on the Editor's Desk documents that informed the cross validation study may not be good predictors of the models' performance on the remainder to the BLS website. To test the generality of the naive Bayes classifier, we solicited input from 11 human judges who were familiar with the BLS website. The sample was chosen by convenience, and consisted of faculty and graduate students who work on the GovStat project. However, none of the reviewers had prior knowledge of the outcome of the classification before their participation. For the experiment, a random sample of 100 documents was drawn from the entire BLS collection. On average each re-7 http://www.google.com 8 A more formal treatment of the combination of labeled and unlabeled data is available in [4]. Table 5: Human-Model Agreement on 100 Sample Docs. Human Judge 1 st Choice Model Model 1 st Choice Model 2 nd Choice N. Bayes (aug.) 14 24 N. Bayes 24 1 Human Judge 2 nd Choice Model Model 1 st Choice Model 2 nd Choice N. Bayes (aug.) 14 21 N. Bayes 21 4 viewer classified 83 documents, placing each document into as many of the categories shown in Table 3 as he or she saw fit. Results from this experiment suggest that room for improvement remains with respect to generalizing to the whole collection from the class models fitted to the Editor's Desk documents. In Table 5, we see, for each classifier, the number of documents for which it's first or second most probable class was voted best or second best by the 11 human judges. In the context of this experiment, we consider a first- or second-place classification by the machine to be accurate because the relation browser interface operates on a multi-way classification, where each document is classified into multiple categories. Thus a document with the "correct" class as its second choice would still be easily available to a user. Likewise, a correct classification on either the most popular or second most popular category among the human judges is considered correct in cases where a given document was classified into multiple classes. There were 72 multi-class documents in our sample, as seen in Figure 4. The remaining 28 documents were assigned to 1 or 0 classes. Under this rationale, The augmented naive Bayes classifier correctly grouped 73 documents, while the smaller model (not augmented by a Google search) correctly classified 50. The resultant 2 test gave p = 0.001, suggesting that increasing the training set improved the ability of the naive Bayes model to generalize from the Editor's Desk documents to the collection as a whole. However, the improvement afforded by the augmented model comes at some cost. In particular , the augmented model is significantly inferior to the model trained solely on Editor's Desk documents if we concern ourselves only with documents selected by the majority of human reviewers--i.e. only first-choice classes. Limiting the right answers to the left column of Table 5 gives p = 0.02 in favor of the non-augmented model. For the purposes of applying the relation browser to complex digital library content (where documents will be classified along multiple categories ), the augmented model is preferable. But this is not necessarily the case in general. It must also be said that 73% accuracy under a fairly liberal test condition leaves room for improvement in our assignment of topics to categories. We may begin to understand the shortcomings of the described techniques by consulting Figure 5, which shows the distribution of categories across documents given by humans and by the augmented naive Bayes model. The majority of reviewers put 157 Number of Human-Assigned Classes Frequency 0 1 2 3 4 5 6 7 0 5 10 15 20 25 30 35 Figure 4: Number of Classes Assigned to Documents by Judges documents into only three categories, jobs, benefits, and occupations . On the other hand, the naive Bayes classifier distributed classes more evenly across the topics. This behavior suggests areas for future improvement. Most importantly, we observed a strong correlation among the three most frequent classes among the human judges (for instance, there was 68% correlation between benefits and occupations). This suggests that improving the clustering to produce topics that were more nearly orthogonal might improve performance CONCLUSIONS AND FUTURE WORK Many developers and maintainers of digital libraries share the basic problem pursued here. Given increasingly large, complex bodies of data, how may we improve access to collections without incurring extraordinary cost, and while also keeping systems receptive to changes in content over time? Data mining and machine learning methods hold a great deal of promise with respect to this problem. Empirical methods of knowledge discovery can aid in the organization and retrieval of information. As we have argued in this paper, these methods may also be brought to bear on the design and implementation of advanced user interfaces. This study explored a hybrid technique for aiding information architects as they implement dynamic interfaces such as the relation browser. Our approach combines unsupervised learning techniques, applied to a focused subset of the BLS website. The goal of this initial stage is to discover the most basic and far-reaching topics in the collection. Based jobs prices spending costs Human Classifications 0.00 0.15 jobs prices spending costs Machine Classifications 0.00 0.10 Figure 5: Distribution of Classes Across Documents on a statistical model of these topics, the second phase of our approach uses supervised learning (in particular, a naive Bayes classifier, trained on individual words), to assign topical relations to the remaining documents in the collection. In the study reported here, this approach has demon-strated promise. In its favor, our approach is highly scalable. It also appears to give fairly good results. Comparing three modes of document representation--full-text, title only, and keyword--we found 98% accuracy as measured by collocation of documents with identical subject headings. While it is not surprising that editor-generated keywords should give strong evidence for such learning, their superiority over full-text and titles was dramatic, suggesting that even a small amount of metadata can be very useful for data mining. However, we also found evidence that learning topics from a subset of the collection may lead to overfitted models. After clustering 1279 Editor's Desk documents into 10 categories , we fitted a 10-way naive Bayes classifier to categorize the remaining 14,000 documents in the collection. While we saw fairly good results (classification accuracy of 75% with respect to a small sample of human judges), this experiment forced us to reconsider the quality of the topics learned by clustering. The high correlation among human judgments in our sample suggests that the topics discovered by analysis of the Editor's Desk were not independent. While we do not desire mutually exclusive categories in our setting, we do desire independence among the topics we model. Overall, then, the techniques described here provide an encouraging start to our work on acquiring subject metadata for dynamic interfaces automatically. It also suggests that a more sophisticated modeling approach might yield 158 better results in the future. In upcoming work we will experiment with streamlining the two-phase technique described here. Instead of clustering documents to find topics and then fitting a model to the learned clusters, our goal is to expand the unsupervised portion of our analysis beyond a narrow subset of the collection, such as The Editor's Desk. In current work we have defined algorithms to identify documents likely to help the topic discovery task. Supplied with a more comprehensive training set, we hope to experiment with model-based clustering, which combines the clustering and classification processes into a single modeling procedure. Topic discovery and document classification have long been recognized as fundamental problems in information retrieval and other forms of text mining. What is increasingly clear, however, as digital libraries grow in scope and complexity, is the applicability of these techniques to problems at the front-end of systems such as information architecture and interface design. Finally, then, in future work we will build on the user studies undertaken by Marchionini and Brunk in efforts to evaluate the utility of automatically populated dynamic interfaces for the users of digital libraries. REFERENCES [1] A. Agresti. An Introduction to Categorical Data Analysis. Wiley, New York, 1996. [2] C. Ahlberg, C. Williamson, and B. Shneiderman. Dynamic queries for information exploration: an implementation and evaluation. In Proceedings of the SIGCHI conference on Human factors in computing systems, pages 619626, 1992. [3] R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. ACM Press, 1999. [4] A. Blum and T. Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the eleventh annual conference on Computational learning theory, pages 92100. ACM Press, 1998. [5] H. Chen and S. Dumais. Hierarchical classification of web content. In Proceedings of the 23rd annual international ACM SIGIR conference on Research and development in information retrieval, pages 256263, 2000. [6] M. Efron, G. Marchionini, and J. Zhang. Implications of the recursive representation problem for automatic concept identification in on-line governmental information. In Proceedings of the ASIST Special Interest Group on Classification Research (ASIST SIG-CR), 2003. [7] C. Fraley and A. E. Raftery. How many clusters? which clustering method? answers via model-based cluster analysis. The Computer Journal, 41(8):578588, 1998. [8] A. K. Jain, M. N. Murty, and P. J. Flynn. Data clustering: a review. ACM Computing Surveys, 31(3):264323, September 1999. [9] T. Joachims. A probabilistic analysis of the Rocchio algorithm with TFIDF for text categorization. In D. H. Fisher, editor, Proceedings of ICML-97, 14th International Conference on Machine Learning, pages 143151, Nashville, US, 1997. Morgan Kaufmann Publishers, San Francisco, US. [10] T. Joachims. Text categorization with support vector machines: learning with many relevant features. In C. N edellec and C. Rouveirol, editors, Proceedings of ECML-98, 10th European Conference on Machine Learning, pages 137142, Chemnitz, DE, 1998. Springer Verlag, Heidelberg, DE. [11] I. T. Jolliffe. Principal Component Analysis. Springer, 2nd edition, 2002. [12] L. Kaufman and P. J. Rosseeuw. Finding Groups in Data: an Introduction to Cluster Analysis. Wiley, 1990. [13] G. Marchionini and B. Brunk. Toward a general relation browser: a GUI for information architects. Journal of Digital Information, 4(1), 2003. http://jodi.ecs.soton.ac.uk/Articles/v04/i01/Marchionini/. [14] A. K. McCallum. Bow: A toolkit for statistical language modeling, text retrieval, classification and clustering. http://www.cs.cmu.edu/~mccallum/bow, 1996. [15] T. Mitchell. Machine Learning. McGraw Hill, 1997. [16] E. Rasmussen. Clustering algorithms. In W. B. Frakes and R. Baeza-Yates, editors, Information Retrieval: Data Structures and Algorithms, pages 419442. Prentice Hall, 1992. [17] R. Tibshirani, G. Walther, and T. Hastie. Estimating the number of clusters in a dataset via the gap statistic, 2000. http://citeseer.nj.nec.com/tibshirani00estimating.html. [18] V. N. Vapnik. The Nature of Statistical Learning Theory. Springer, 2000. 159

information architecture;Information Architecture;BLS;digital libraries;document classification;Machine Learning;topic discovery;document representation;Interface Design;clustering;relational browser

133

Machine Learning in DNA Microarray Analysis for Cancer Classification

The development of microarray technology has supplied a large volume of data to many fields. In particular, it has been applied to prediction and diagnosis of cancer, so that it expectedly helps us to exactly predict and diagnose cancer. To precisely classify cancer we have to select genes related to cancer because extracted genes from microarray have many noises. In this paper, we attempt to explore many features and classifiers using three benchmark datasets to systematically evaluate the performances of the feature selection methods and machine learning classifiers. Three benchmark datasets are Leukemia cancer dataset, Colon cancer dataset and Lymphoma cancer data set. Pearson's and Spearman's correlation coefficients, Euclidean distance, cosine coefficient, information gain, mutual information and signal to noise ratio have been used for feature selection. Multi-layer perceptron, k-nearest neighbour, support vector machine and structure adaptive selforganizing map have been used for classification. Also, we have combined the classifiers to improve the performance of classification. Experimental results show that the ensemble with several basis classifiers produces the best recognition rate on the benchmark dataset.

Introduction The need to study whole genome such as Human Genomic Project (HGP) is recently increasing because fragmentary knowledge about life phenomenon with complex control functions of molecular-level is limited. DNA chips have been developed during that process because understanding the functions of genome sequences is essential at that time. The development of DNA microarray technology has been produced large amount of gene data and has made it easy to monitor the expression patterns of thousands of genes simultaneously under particular experimental environments and conditions (Harrington et al. 2000). Also, we can analyze the gene information very rapidly and precisely by managing them at one time (Eisen et al. 1999). Microarray technology has been applied to the field of accurate prediction and diagnosis of cancer and expected that it would help them. Especially accurate classification of cancer is very important issue for treatment of cancer. Many researchers have been studying many problems of cancer classification using gene expression profile data and attempting to propose the optimal classification technique to work out these problems (Dudoit et al. 2000, Ben-Dor et al. 2000) as shown in Table . Some produce better results than others, but there have been still no comprehensive work to compare the possible feature selection methods and classifiers. We need a thorough effort to give the evaluation of the possible methods to solve the problems of analyzing gene expression data. The gene expression data usually consist of huge number of genes, and the necessity of tools analysing them to get useful information gets radical. There is research that systematically analyzes the results of test using a variety of feature selection methods and classifiers for selecting informative genes to help classification of cancer and classifying cancer (Ryu et al. 2002). However, the results were not verified enough because only one benchmark dataset was used. Due to the reason, it is necessary to analyse systematically the performance of classifiers using a variety of benchmark datasets. In this paper, we attempt to explore many features and classifiers that precisely classify cancer using three recently published benchmark dataset. We adopted seven feature selection methods and four classifiers, which are commonly used in the field of data mining and pattern recognition. Feature selection methods include Pearson's and Spearman's correlation coefficients, Euclidean distance, cosine coefficient, information gain, mutual information and signal to noise ratio. Also, classification methods include multi-layer perceptron (MLP), k-nearest neighbour (KNN), support vector machine (SVM) and structure adaptive self organizing map (SOM). We also attempt to combine some of the classifiers with majority voting to improve the performance of classification. Backgrounds DNA arrays consist of a large number of DNA molecules spotted in a systemic order on a solid substrate. Depending on the size of each DNA spot on the array, DNA arrays can be categorized as microarrays when the diameter of DNA spot is less than 250 microns, and macroarrays when the diameter is bigger than 300 microns. The arrays with the small solid substrate are also referred to as DNA chips. It is so powerful that we can investigate the gene information in short time, because at least hundreds of genes can be put on the DNA microarray to be analyzed. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Sam ple Ge ne s ALL AML 1000 2000 3000 4000 5000 6000 7129 Gene expression data DNA microarray Image scanner Fig. 1. General process of acquiring the gene expression data from DNA microarray DNA microarrays are composed of thousands of individual DNA sequences printed in a high density array on a glass microscope slide using a robotic arrayer as shown in Fig. 1. The relative abundance of these spotted DNA sequences in two DNA or RNA samples may be assessed by monitoring the differential hybridization of the two samples to the sequences on the array. For mRNA samples, the two samples are reverse-transcribed into cDNA, labeled using different fluorescent dyes mixed (red-fluorescent dye Cy5 and green-fluorescent dye Cy3). After the hybridization of these samples with the arrayed DNA probes, the slides are imaged using scanner that makes fluorescence measurements for each dye. The log ratio between the two intensities of each dye is used as the gene expression data (Lashkari et al. 1997, Derisi et al. 1997, Eisen et al. 1998). ) 3 Cy ( Int ) 5 Cy ( Int log _ 2 = expression gene (1) where Int(Cy5) and Int(Cy3) are the intensities of red and green colors. Since at least hundreds of genes are put on the DNA microarray, it is so helpful that we can Table . Relevant works on cancer classification Method Authors Dataset Feature Classifier Accuracy [%] Leukemia 94.1 Furey et al. Colon Signal to noise ratio SVM 90.3 Li et al. 2000 Leukemia Model selection with Akaike information criterion and Bayesian information criterion with logistic regression 94.1 Lymphoma 84.6~ Li et al. 2001 Colon Genetic Algorithm KNN 94.1~ Leukemia 91.6 Colon Nearest neighbor 80.6 Leukemia 94.4 Colon SVM with quadratic kernel 74.2 Leukemia 95.8 Ben-Dor et al. Colon All genes, TNoM score AdaBoost 72.6 Leukemia 95.0~ Lymphoma Nearest neighbor 95.0~ Leukemia 95.0~ Lymphoma Diagonal linear discriminant analysis 95.0~ Leukemia 95.0~ Dudoit et al. Lymphoma The ratio of between-groups to within-groups sum of squares BoostCART 90.0~ Leukemia 94.2 Lymphoma 98.1 Colon Logistic discriminant 87.1 Leukemia 95.4 Lymphoma 97.6 Colon Principal component analysis Quadratic discriminant analysis 87.1 Leukemia 95.9 Lymphoma 96.9 Colon Logistic discriminant 93.5 Leukemia 96.4 Lymphoma 97.4 Nguyen et al. Colon Partial least square Quadratic discriminant analysis 91.9 investigate the genome-wide information in short time. 2.2 Related Works It is essential to efficiently analyze DNA microarray data because the amount of DNA microarray data is usually very large. The analysis of DNA microarray data is divided into four branches: clustering, classification, gene identification, and gene regulatory network modeling. Many machine learning and data mining methods have been applied to solve them. Information theory (Fuhrman et al. 2000) has been applied to gene identification problem. Also, boolean network (Thieffry et al. 1998), Bayesian network (Friedman et al. 2000), and reverse engineering method (Arkin et al. 1997) have been applied to gene regulatory network modeling problem. Several machine learning techniques have been previously used in classifying gene expression data, including Fisher linear discriminant analysis (Dudoit et al. 2000), k nearest neighbour (Li et al. 2001), decision tree, multi-layer perceptron (Khan et al. 2001, Xu et al. 2002), support vector machine (Furey et al. 2000, Brown et al. 2000), boosting, and self-organizing map (Golub et al. 1999). Also, many machine learning techniques were have been used in clustering gene expression data (Shamir 2001). They include hierarchical clustering (Eisen et al. 1998), self-organizing map (Tamayo et al. 1999), and graph theoretic approaches (Hartuv et al. 2000, Ben-Dor et al. 1999, Sharan et al. 2000) The first approach, classification method, is called supervised method while the second approach, clustering method, is called unsupervised method. Clustering methods do not use any tissue annotation (e.g., tumor vs. normal) in the partitioning step. In contrast, classification methods attempt to predict the classification of new tissues, based on their gene expression profiles after training on examples (training data) that have been classified by an external "supervision" (Ben-Dor et al. 2000). Table shows relevant works on cancer classification. Machine Learning for DNA Microarray We define machine learning for DNA microarray that selects discriminative genes related with classification from gene expression data, trains classifier and then classifies new data using learned classifier. The system is as shown in Fig. 2. After acquiring the gene expression data calculated from the DNA microarray, our prediction system has 2 stages: feature selection and pattern classification stages. The feature selection can be thought of as the gene selection, which is to get the list of genes that might be informative for the prediction by statistical, information theoretical methods, etc. Since it is highly unlikely that all the 7,129 genes have the information related to the cancer and using all the genes results in too big dimensionality, it is necessary to explore the efficient way to get the best feature. We have extracted 25 genes using seven methods described in Section 3.1, and the cancer predictor classifies the category only with these genes. Given the gene list, a classifier makes decision to which category the gene pattern belongs at prediction stage. We have adopted four most widely used classification methods and an ensemble classifier as shown in Fig. 2. Feature selection Tumor Normal Cancer predictor Pearson's correlation coefficient Spearman's correlation coefficient Euclidean distance Cosine coefficient Information gain Mutual information Signal to noise ratio 3-layered MLP with backpropagation k-nearest neighbor Support vector machine Structure adaptive self-organizing map Ensemble classifier Microarray Expression data Fig. 2. Cancer classification system 3.1 Gene Selection Among thousands of genes whose expression levels are measured, not all are needed for classification. Microarray data consist of large number of genes in small samples. We need to select some genes highly related with particular classes for classification, which is called informative genes (Golub et al. 1999). This process is referred to as gene selection. It is also called feature selection in machine learning. Using the statistical correlation analysis, we can see the linear relationship and the direction of relation between two variables. Correlation coefficient r varies from 1 to +1, so that the data distributed near the line biased to (+) direction will have positive coefficients, and the data near the line biased to (-) direction will have negative coefficients. Suppose that we have a gene expression pattern g i (i = 1 ~ 7,129 in Leukemia data, i = 1 ~ 2,000 in Colon data, i = 1 ~ 4,026 in Lymphoma data) . Each g i is a vector of gene expression levels from N samples, g i = (e 1 , e 2 , e 3 , ..., e N ). The first M elements (e 1 , e 2 , ..., e M ) are examples of tumor samples, and the other N-M (e M+1 , e M+2 , ..., e N ) are those from normal samples. An ideal gene pattern that belongs to tumor class is defined by g ideal_tumor = (1, 1, ..., 1, 0, ..., 0), so that all the elements from tumor samples are 1 and the others are 0. In this paper, we have calculated the correlation coefficient between this g ideal and the expression pattern of each gene. When we have two vectors X and Y that contain N elements, r Pearson and r Spearman are calculated as follows: ( ) ( ) = N Y Y N X X N Y X XY r Pearson 2 2 2 2 (2) ( ) ( ) 1 6 1 2 2 = N N D D r y x Spearman (3) where, D x and D y are the rank matrices of X and Y, respectively. The similarity between two input vectors X and Y can be thought of as distance. Distance is a measure on how far the two vectors are located, and the distance between g ideal_tumor and g i tells us how much the g i is likely to the tumor class. Calculating the distance between them, if it is bigger than certain threshold, the gene g i would belong to tumor class, otherwise g i belongs to normal class. In this paper, we have adopted Euclidean distance (r Eclidean ) and cosine coefficient (r Cosine ) represented by the following equations: ( ) = 2 Y X r Eclidean (4) = 2 2 Y X XY r Cosine (5) We have utilized the information gain and mutual information that are widely used in many fields such as text categorization and data mining. If we count the number of genes excited ( ) ( i g P ) or not excited ( ) ( _ i g P ) in category c j ( ) ( j c P ), the coefficients of the information gain and mutual information become as follows: ) ( ) ( ) , ( log ) , ( ) ( ) ( ) , ( log ) , ( ) , ( i j j i i i i j j i j i j i g P c P c g P c g P g P c P c g P c g P c g IG + = (6) ) ( ) ( ) , ( log ) , ( i j j i j i g P c P c g P c g MI = (7) Mutual information tells us the dependency relationship between two probabilistic variables of events. If two events are completely independent, the mutual information is 0. The more they are related, the higher the mutual information gets. Information gain is used when the features of samples are extracted by inducing the relationship between gene and class by the presence frequency of the gene in the sample. Information gain measures the goodness of gene using the presence and absence within the corresponding class. For each gene g i , some are from tumor samples, and some are from normal samples. If we calculate the mean and standard deviation from the distribution of gene expressions within their classes, the signal to noise ratio of gene g i , SN(g i ), is defined by: ) ( ) ( ) ( ) ( ) ( i normal i tumor i normal i tumor i g g g g g SN = (8) 3.2 Classification Many algorithms designed for solving classification problems in machine learning have been applied to recent research of prediction and classification of cancer with gene expression data. General process of classification in machine learning is to train classifier to accurately recognize patterns from given training samples and to classify test samples with the trained classifier. Representative classification algorithms such as multi-layer perceptron, k-nearest neighbour, support vector machine, and structure-adaptive self-organizing map are applied to the classification. 1) MLP Error backpropagation neural network is a feed-forward multilayer perceptron (MLP) that is applied in many fields due to its powerful and stable learning algorithm (Lippman et al. 1987). The neural network learns the training examples by adjusting the synaptic weight of neurons according to the error occurred on the output layer. The power of the backpropagation algorithm lies in two main aspects: local for updating the synaptic weights and biases, and efficient for computing all the partial derivatives of the cost function with respect to these free parameters (Beale 1996). The weight-update rule in backpropagation algorithm is defined as follows: ) 1 ( ) ( + = n w x n w ji ji j ji (9) where ) (n w ji is the weight update performed during the nth iteration through the main loop of the algorithm, is a positive constant called the learning rate, j is the error term associated with j, x ji is the input from node i to unit j, and 0 &lt;1 is a constant called the momentum. 2) KNN k-nearest neighbor (KNN) is one of the most common methods among memory based induction. Given an input vector, KNN extracts k closest vectors in the reference set based on similarity measures, and makes decision for the label of input vector using the labels of the k nearest neighbors. Pearson's coefficient correlation and Euclidean distance have been used as the similarity measure. When we have an input X and a reference set D = {d 1 , d 2 , ..., d N }, the probability that X may belong to class c j , P(X, c j ) is defined as follows: j j i kNN d i j b c d P d X c X P i = ) , ( ) , Sim( ) , ( (10) where Sim(X, d i ) is the similarity between X and d i and b j is a bias term. 3) SASOM Self-organizing map (SOM) defines a mapping from the input space onto an output layer by unsupervised learning algorithm. SOM has an output layer consisting of N nodes, each of which represents a vector that has the same dimension as the input pattern. For a given input vector X, the winner node m c is chosen using Euclidean distance between x and its neighbors, m i . i i c m x m x = min (11) )} ( ) ( { ) ( ) ( ) ( ) 1 ( t m t x t n t t m t m i ci i i + = + (12) Even though SOM is well known for its good performance of topology preserving, it is difficult to apply it to practical classification since the topology should be fixed before training. A structure adaptive self-organizing map (SASOM) is proposed to overcome this shortcoming (Kim et al. 2000). SASOM starts with 44 map, and dynamically splits the output nodes of the map, where the data from different classes are mixed, trained with the LVQ learning algorithm. Fig. 3 illustrates the algorithm of SASOM. Input data No Yes Map generated Initialize map as 4X4 Train with Kohonen's algorithm Find nodes whose hit_ratio is less than 95.0% Split the nodes to 2X2 submap Train the split nodes with LVQ algorithm Remove nodes not participted in learning Stop condition satisfied? Structure adaptation Fig. 3. Overview of SASOM 4) SVM Support vector machine (SVM) estimates the function classifying the data into two classes (Vapnik 1995, Moghaddam et al. 2000). SVM builds up a hyperplane as the decision surface in such a way to maximize the margin of separation between positive and negative examples. SVM achieves this by the structural risk minimization principle that the error rate of a learning machine on the test data is bounded by the sum of the training-error rate and a term that depends on the Vapnik-Chervonenkis (VC) dimension. Given a labeled set of M training samples (X i , Y i ), where X i R N and Y i is the associated label, Y i {-1, 1}, the discriminant hyperplane is defined by: = + = M i i i i b X X k Y X f 1 ) , ( ) ( (13) where k( .) is a kernel function and the sign of f(X) determines the membership of X. Constructing an optimal hyperplane is equivalent to finding all the nonzero i (support vectors) and a bias b. We have used SVM light module and SVM RBF in this paper. 5) Ensemble classifier Classification can be defined as the process to approximate I/O mapping from the given observation to the optimal solution. Generally, classification tasks consist of two parts: feature selection and classification. Feature selection is a transformation process of observations to obtain the best pathway to get to the optimal solution. Therefore, considering multiple features encourages obtaining various candidate solutions, so that we can estimate more accurate solution to the optimal than any other local optima. When we have multiple features available, it is important to know which of features should be used. Theoretically, as many features we may concern, it may be more effective for the classifier to solve the problems. But features that have overlapped feature spaces may cause the redundancy of irrelevant information and result in the counter effect such as overfitting. Therefore, it is more important to explore and utilize independent features to train classifiers, rather than increase the number of features we use. Correlation between feature sets can be induced from the distribution of feature numbers, or using mathematical analysis using statistics. Meanwhile, there are many algorithms for the classification from machine learning approach, but none of them is perfect. However, it is always difficult to decide what to use and how to set up its parameters. According to the environments the classifier is embedded, some algorithm works well and others not. It is because, depending on the algorithms, features and parameters used, the classifier searches in different solution space. These sets of classifiers produce their own outputs, and enable the ensemble classifier to explore more wide solution space. We have applied this idea to a classification framework as shown in Fig. 4. If there are k features and n classifiers, there are kn feature-classifier combinations. There are kn C m possible ensemble classifiers when m feature-classifier combinations are selected for ensemble classifier. Then classifiers are trained using the features selected, finally a majority voting is accompanied to combine the outputs of these classifiers. After classifiers with some features are trained independently produce their own outputs, final answer will be judged by a combining module, where the majority voting method is adopted. Class i Class 1 Feature extractor Majority Voting Input pattern ... ... Classifier n1 Classifier nk ... Classifier a1 Classifier ak ... Classifier a2 Classifier n2 Feature 1 Feature k Feature 2 ... ... ... .. . Fig 4. Overview of the ensemble classifier Experimental Results There are several microarray datasets from published cancer gene expression studies, including leukemia cancer dataset, colon cancer dataset, lymphoma dataset, breast cancer dataset, NCI60 dataset, and ovarian cancer dataset. Among them three datasets are used in this paper. The first dataset and third dataset involve samples from two variants of the same disease and second dataset involves tumor and normal samples of the same tissue. Because the benchmark data have been studied in many papers, we can compare the results of this paper with others. 1) Leukemia cancer dataset Leukemia dataset consists of 72 samples: 25 samples of acute myeloid leukemia (AML) and 47 samples of acute lymphoblastic leukemia (ALL). The source of the gene expression measurements was taken form 63 bone marrow samples and 9 peripheral blood samples. Gene expression levels in these 72 samples were measured using high density oligonucleotide microarrays (Ben-Dor et al. 2000). 38 out of 72 samples were used as training data and the remaining were used as test data in this paper. Each sample contains 7129 gene expression levels. 2) Colon cancer dataset Colon dataset consists of 62 samples of colon epithelial cells taken from colon-cancer patients. Each sample contains 2000 gene expression levels. Although original data consists of 6000 gene expression levels, 4000 out of 6000 were removed based on the confidence in the measured expression levels. 40 of 62 samples are colon cancer samples and the remaining are normal samples. Each sample was taken from tumors and normal healthy parts of the colons of the same patients and measured using high density oligonucleotide arrays (Ben-Dor et al. 2000). 31 out of 62 samples were used as training data and the remaining were used as test data in this paper. 3) Lymphoma cancer dataset B cell diffuse large cell lymphoma (B-DLCL) is a heterogeneous group of tumors, based on significant variations in morphology, clinical presentation, and response to treatment. Gene expression profiling has revealed two distinct tumor subtypes of B-DLCL: germinal center B cell-like DLCL and activated B cell-like DLCL (Lossos et al. 2000). Lymphoma dataset consists of 24 samples of GC B-like and 23 samples of activated B-like. 22 out of 47 samples were used as training data and the remaining were used as test data in this paper. 4.2 Environments For feature selection, each gene is scored based on the feature selection methods described in Section 3.1, and the 25 top-ranked genes are chosen as the feature of the input pattern. For classification, we have used 3-layered MLP with 5~15 hidden nodes, 2 output nodes, 0.01~0.50 of learning rate and 0.9 of momentum. KNN has been used with k=1~8. Similarity measures used in KNN are Pearson's correlation coefficient and Euclidean distance. SASOM has been used by 44 map with rectangular topology, 0.05 of initial learning rate, 1000 of initial learning length, 10 of initial radius, 0.02 of final learning rate, 10000 of final learning length and 3 of final radius. We have used SVM with linear function and RBF function as kernel function. In RBF, we have changed 0.1~0.5 gamma variable. 4.3 Analysis of results Table shows the IDs of genes overlapped by Pearson's correlation coefficient, cosine coefficient, Euclidean distance in each dataset. Among these genes there are some genes overlapped by other feature selection methods. For example, gene 2288 of leukemia has been third-ranked in information gain. The number of overlapped genes of leukemia dataset is 17. The number of overlapped genes of colon dataset is 9. The number of overlapped genes of lymphoma dataset is 19. These overlapped genes are very informative. In particular, Zyxin, gene 4847 of leukemia, has been reported as informative (Golub et al. 1999), but there are no genes appeared commonly in every method. Table . The IDs of genes overlapped by Pearson's correlation coefficient, cosine coefficient, and Euclidean distance 461 1249 1745 1834 2020 2043 2242 2288 3258 3320 4196 4847 5039 6200 6201 Leukemia 6373 6803 187 619 704 767 1060 Colon 1208 1546 1771 1772 36 75 76 77 86 86 678 680 1636 1637 2225 2243 2263 2412 2417 Lymphoma 2467 3890 3893 3934 Fig. 5 shows the expression level of genes chosen by Pearson's correlation coefficient method in Leukemia dataset. 1~27 samples are ALL and 28~38 samples are AML. The differences of brightness between AML and ALL represent that genes chosen by Pearson's correlation coefficient method divide samples into AML and ALL. The results of recognition rate on the test data are as shown in Tables , , and . Column is the list of feature selection methods: Pearson's correlation coefficient (PC), Spearman's correlation coefficient (SC), Euclidean distance (ED), cosine coefficient (CC), information gain (IG), mutual information (MI), and signal to noise ratio (SN). KNN Pearson and MLP seem to produce the best recognition rate among the classifiers on the average. KNN Pearson is better than KNN cosine . SVM is poorer than any other classifiers. 0 0.2 0.4 0.6 0.8 1 Sample Ge n e 27 38 5 10 15 20 25 Fig. 5. Expression level of genes chosen by r Pearson in Leukemia dataset Table . Recognition rate with features and classifiers (%) in Leukemia dataset SVM KNN MLP SASOM Linear RBF Cosine Pearson PC 97.1 76.5 79.4 79.4 97.1 94.1 SC 82.4 61.8 58.8 58.8 76.5 82.4 ED 91.2 73.5 70.6 70.6 85.3 82.4 CC 94.1 88.2 85.3 85.3 91.2 94.1 IG 97.1 91.2 97.1 97.1 94.1 97.1 MI 58.8 58.8 58.8 58.8 73.5 73.5 SN 76.5 67.7 58.8 58.8 73.5 73.5 Mean 85.3 74.0 72.7 72.7 84.5 85.3 Table . Recognition rate with features and classifiers (%) in Colon dataset SVM KNN MLP SASOM Linear RBF Cosine Pearson PC 74.2 74.2 64.5 64.5 71.0 77.4 SC 58.1 45.2 64.5 64.5 61.3 67.7 ED 67.8 67.6 64.5 64.5 83.9 83.9 CC 83.9 64.5 64.5 64.5 80.7 80.7 IG 71.0 71.0 71.0 71.0 74.2 80.7 MI 71.0 71.0 71.0 71.0 74.2 80.7 SN 64.5 45.2 64.5 64.5 64.5 71.0 Mean 70.1 62.7 66.4 66.4 72.7 77.4 Table . Recognition rate with features and classifiers (%) in Lymphoma dataset SVM KNN MLP SASOM Linear RBF Cosine Pearson PC 64.0 48.0 56.0 60.0 60.0 76.0 SC 60.0 68.0 44.0 44.0 60.0 60.0 ED 56.0 52.0 56.0 56.0 56.0 68.0 CC 68.0 52.0 56.0 56.0 60.0 72.0 IG 92.0 84.0 92.0 92.0 92.0 92.0 MI 72.0 64.0 64.0 64.0 80.0 64.0 SN 76.0 76.0 72.0 76.0 76.0 80.0 Mean 69.7 63.4 62.9 63.4 69.1 73.1 Fig. 6 shows the comparison of the average performance of features. Although the results are different between datasets, information gain is the best, and Pearson's correlation coefficient is the second. Mutual information and Spearman's correlation coefficient are poor. The difference of performance in datasets might be caused by the characteristics of data. 0 20 40 60 80 100 Leukemia Colon Ly mphoma R e c o g n it io n ra t e [ % ] PC SC ED CC IG MI SN Fig. 6. Average performance of feature selection methods Recognition rates by ensemble classifiers are shown in Table . Majority voting-3 means the ensemble classifier using majority voting with 3 classifiers, and majority voting-all means the ensemble classifier using majority voting with all 42 feature-classifier combinations. Fig. 7 shows the comparison of the performance of the best classifier of all possible 42 C 3 ensemble classifiers, ensemble classifier-3 and ensemble classifier-all. The best result of Leukemia is obtained by all classifier except SASOM. The result of the best classifier is the same as that of the best ensemble classifier using majority voting with 3 classifiers. In other datasets, the performance of ensemble classifier surpasses the best classifier. In all datasets, ensemble classifier using majority voting with all classifiers are the worst. Table . Recognition rate by ensemble classifier Majority voting-3 Majority voting-all Leukemia 97.1 91.2 Colon 93.6 71.0 Lymphoma 96.0 80.0 60 70 80 90 100 Leukemia Colon Lymphoma R ecognit ion rat e [ %] The best classifier Majority voting-3 Majority voting-all Fig. 7. Comparison of the performance of the best classifier, the best ensemble classifier-3, and ensemble classifier-all Table . Classifiers of the best ensemble classifier of all possible 42 C 3 ensemble classifiers in Colon dataset Classifier Feature selection method MLP KNN cosine KNN cosine Cosine coefficient Euclidean distance Pearson's correlation coefficient MLP KNN cosine KNN Pearson Cosine coefficient Euclidean distance Pearson's correlation coefficient MLP KNN cosine SASOM Cosine coefficient Euclidean distance Pearson's correlation coefficient MLP KNN cosine KNN pearson Mutual information Euclidean distance Pearson's correlation coefficient MLP KNN cosine KNN pearson Information gain Euclidean distance Pearson's correlation coefficient MLP MLP KNN pearson Cosine coefficient Pearson's correlation coefficient Euclidean distance KNN pearson KNN pearson SASOM Euclidean distance Mutual information Pearson's correlation coefficient KNN pearson KNN pearson SASOM Euclidean distance Information gain Pearson's correlation coefficient Table shows the classifiers of the best ensemble classifier of all possible 42 C 3 ensemble classifiers in Colon dataset where its recognition rate is 93.6%. If we observe the classifiers of the best ensemble classifier in Fig. 10, we find features more important to affect the result than classifiers. In other words, in ensemble classifiers there must be classifiers with Euclidean distance and Pearson's correlation coefficient. The other classifier is the one with cosine coefficient, mutual information or information gain. This fact is also prominent in Lymphoma dataset. Most of the classifiers of the best ensemble classifiers are classifiers with information gain, signal to noise ratio and Euclidean distance, or the classifiers with information gain, signal to noise ratio and Pearson's correlation coefficient. As shown in Fig. 8~11, Euclidean distance, Pearson's correlation coefficient and cosine coefficient are highly correlated in Colon dataset. Table shows genes ranked by Euclidean distance, Pearson's correlation coefficient and cosine coefficient and the value of genes by each method. The bold faced figures mean the overlapped genes of those features. There are some overlapped genes among them as shown in Table . This indicates overlapped genes of highly correlated features can discriminate classes and the other genes not overlapped among combined features can supplement to search the solution spaces. For example, gene 1659 and gene 550 are high-ranked in both of Pearson's correlation coefficient and cosine coefficient, and gene 440 is high-ranked in both of Euclidean distance and cosine coefficient. This subset of two features might paly an important role in classification. This paper shows that the ensemble classifier works and we can improve the classification performance by combining complementary common sets of classifiers learned from three independent features, even when we use simple combination method like majority voting. Fig. 8. Correlation of Euclidean distance and Pearson's correlation coefficient in Colon dataset Fig. 9. Correlation of Euclidean distance and cosine coefficient in Colon dataset Fig. 10. Correlation of Pearson's correlation coefficient and cosine coefficient in Colon dataset Fig. 11. Correlation of Euclidean distance, Pearson's correlation coefficient and cosine coefficient in Colon dataset Table . Genes ranked by Euclidean distance, Pearson's correlation coefficient and cosine coefficient Rank Euclidean Pearson Cosine 1 619(2.262385) 2 767(2.335303) 3 704(2.374358) 4 187(2.388404) 5 207(2.410640) 6 887(2.473033) 7 635(2.474971) 8 1915(2.498611) 9 1046(2.506833) 10 1208(2.512257) 11 482(2.520699) 12 1771(2.525080) 13 1993(2.529032) 14 62(2.546894) 15 1772(2.547455) 16 1194(2.549244) 17 1594(2.551892) 18 199(2.557360) 19 1867(2.587469) 20 959(2.589989) 21 440(2.593881) 22 480(2.594514) 23 1546(2.604907) 24 399(2.613609) 25 1060(2.614100) 619(0.681038) 1771(0.664378) 1659(0.634084) 550(0.631655) 187(0.626262) 1772(0.621581) 1730( 0.615566) 1648(0.614949) 365(0.614591) 1208(0.603313) 1042(0.602160) 1060(0.601712) 513(0.596444) 767(0.594119) 1263(0.591725) 138(0.587851) 1826(0.584774) 1546(0.582293) 141(0.579073) 1227(0.574537) 704(0.569022) 1549(0.562828) 1489(0.561003) 1724(0.559919) 1209(0.559778) 619(0.895971) 1772(0.875472) 767(0.874914) 1771(0.873892) 1659(0.870115) 187(0.867285) 704(0.866679) 1208(0.866029) 550(0.864547) 1546(0.856904) 251(0.855841) 1915(0.855784) 440(0.855453) 1263(0.854854) 1060(0.854829) 965(0.854137) 1648(0.854119) 1942(0.853586) 513(0.852270) 1042(0.851993) 1993(0.851753) 365(0.851205) 1400(0.849531) 207(0.849084) 271(0.848481) Concluding Remarks We have conducted a thorough quantitative comparison among the 42 combinations of features and classifiers for three benchmark dataset. Information gain and Pearson's correlation coefficient are the top feature selection methods, and MLP and KNN are the best classifiers. The experimental results also imply some correlations between features and classifiers, which might guide the researchers to choose or devise the best classification method for their problems in bioinformatics. Based on the results, we have developed the optimal feature-classifier combination to produce the best performance on the classification. We have combined 3 classifiers among 42 classifiers using majority voting. We could confirm that ensemble classifier of highly correlated features works better than ensemble of uncorrelated features. In particular, we analyzed the improvement of the classification performance for Colon dataset. Moreover, our method of combining classifiers is very simple, and there are many methods of combining classifiers in machine learning and data mining fields. We will have to apply more sophisticated methods of combining classifiers to the same dataset to confirm the results obtained and get better results. Acknowledgements This paper was supported by Brain Science and Engineering Research Program sponsored by Korean Ministry of Science and Technology. References Alon, U., Barkai, N., et al. (1999): Broad patterns of gene expression revealed by clustering analysis of tumor and normal colon tissues probed by oligonucleotide arrays. Proc. of the Natl. Acad. of Sci. USA, 96:6745-6750. Arkin, A., Shen, P. and Ross, J. (1997): A test case of correlation metric construction of a reaction pathway from measurements. Science, 277:1275-1279. Beale, H. D. (1996): Neural Network Design. 1-47, PWS Publish Company. Ben-Dor, A., Shamir, R. and Yakhini, Z. (1999): Clustering gene expression patterns. Journal of Computational Biology, 6:281-297. Ben-Dor, A., Bruhn, L., Friedman, N., Nachman, I., Schummer, M. and Yakhini, N. (2000): Tissue classification with gene expression profiles. Journal of Computational Biology, 7:559-584. Brown, M. P. S., Grundy, W. N., Lin, D., Cristianini, N., Sugnet, C. W., Furey, T. S., Ares, M. Jr. and Haussler, D. (2000): Knowledge-based analysis of microarray gene expression data by using support vector machines. Proc. of the Natl. Acad. of Sci. USA, 97:262-267, 2000. Derisi, J., Iyer, V. and Brosn, P. (1997): Exploring the metabolic and genetic control of gene expression on a genomic scale. Science, 278:680-686. Dudoit, S., Fridlyand, J. and Speed, T. P. (2000): Comparison of discrimination methods for the classification of tumors using gene expression data. Technical Report 576, Department of Statistics, University of California, Berkeley. Eisen, M. B., Spellman, P. T., Brown, P. O. and Bostein, D. (1998): Cluster analysis and display of genome-wide expression patterns. Proc. of the Natl. Acad. of Sci. USA, 95:14863-14868. Eisen, M. B. and Brown, P. O. (1999): DNA arrays for analysis of gene expression. Methods Enzymbol, 303: 179-205. Friedman, N., Linial, M., Nachman, I. and Pe'er, D. (2000): Using Bayesian networks to analyze expression data. Journal of Computational Biology, 7:601-620. Fuhrman, S., Cunningham, M. J., Wen, X., Zweiger, G., Seilhamer, J. and Somogyi, R. (2000): The application of Shannon entropy in the identification of putative drug targets. Biosystems, 55:5-14. Furey, T. S., Cristianini, N., Duffy, N., Bednarski, D. W., Schummer, M. and Haussler, D. (2000): Support vector machine classification and validation of cancer tissue samples using microarray expression data. Bioinformatics, 16(10):906-914. Golub, T. R., Slonim, D. K., Tamayo, P., Huard, C., GaasenBeek, M., Mesirov, J. P., Coller, H., Loh, M. L., Downing, J. R., Caligiuri, M. A., Blomfield, C. D., and Lander, E. S. (1999): Molecular classification of cancer: Class discovery and class prediction by gene-expression monitoring. Science, 286:531-537. Harrington, C. A., Rosenow, C., and Retief, J. (2000): Monitoring gene expression using DNA microarrays. Curr. Opin. Microbiol., 3:285-291. Hartuv, E., Schmitt, A., Lange, J., Meier-Ewert, S., Lehrach, H. and Shamir, R. (2000): An algorithm for clustering cDNA fingerprints. Genomics, 66(3):249-256. Khan, J., Wei, J. S., Ringner, M., Saal, L. H., Ladanyi, M., Westermann, F., Berthold, F., Schwab, M., Antonescu, C. R., Peterson, C. And Meltzer, P. S. (2001): Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks. Nature Medicine, 7(6):673-679. Kim, H. D. and Cho, S.-B. (2000): Genetic optimization of structure-adaptive self-organizing map for efficient classification. Proc. of International Conference on Soft Computing, 34-39, World-Scientific Publishing. Lashkari, D., Derisi, J., McCusker, J., Namath, A., Gentile, C., Hwang, S., Brown, P., and Davis, R. (1997): Yeast microarrays for genome wide parallel genetic and gene expression analysis. Proc. of the Natl. Acad. of Sci. USA, 94:13057-13062. Lippman, R. P. (1987): An introduction to computing with neural nets. IEEE ASSP Magazine, 4-22. Li, L., Weinberg, C. R., Darden, T. A. and Pedersen, L. G. (2001): Gene selection for sample classification based on gene expression data: study of sensitivity to choice of parameters of the GA/KNN method. Bioinformatics, 17(12):1131-1142. Li, W. and Yang, Y. (2000): How many genes are needed for a discriminant microarray data analysis. Critical Assessment of Techniques for Microarray Data Mining Workshop. Lossos, I. S., Alizadeh, A. A., Eisen, M. B., Chan, W. C., Brown, P. O., Bostein, D., Staudt, L. M., and Levy, R. (2000): Ongoing immunoglobulin somatic mutation in germinal center B cell-like but not in activated B cell-like diffuse large cell lymphomas. Proc. of the Natl. Acad. of Sci. USA, 97(18):10209-10213. Moghaddam, B. and Yang, M.-H. (2000): Gender classification with support vector machines. Proc. of 4th IEEE Intl. Conf. on Automatic Face and Gesture Recognition, 306-311. Nguyen, D. V. and Rocke, D. M. (2002): Tumor classification by partial least squares using microarray gene expression data. Bioinformatics, 18(1):39-50. Quinlan, J. R. (1986): The effect of noise on concept learning. Machine Learning: An Artificial Intelligence Approach, Michalski, R. S., Carbonell, J. G. and Mitchell, T. M. (eds). San Mateo, CA: Morgan Kauffmann, 2:149-166. Ryu, J. and Cho, S. B. (2002): Towards optimal feature and classifier for gene expression classification of cancer. Lecture Note in Artificial Intelligence, 2275:310-317. Shamir, R. and Sharan, R. (2001): Algorithmic approaches to clustering gene expression data. Current Topics in Computational Biology. In Jiang, T., Smith, T., Xu, Y. and Zhang, M. Q. (eds), MIT press. Sharan, R. and Shamir, R. (2000): CLICK: A clustering with applications to gene expression analysis. Proc. Of the Eighth International Conference in Computational Molecular Biology (ISBM), 307-316. Tamayo, P. (1999): Interpreting patterns of gene expression with self-organizing map: Methods and application to hematopoietic differentiation. Proc. of the National Academy of Sciences of the United States of America, 96: 2907-2912. Thieffry, D. and Thomas, R. (1998): Qualitative analysis of gene networks. Pacific Symposium on Biocomputing, 3:66-76. Vapnik, V. N. (1995): The Nature of Statistical Learning Theory, New York: Springer. Xu, Y., Selaru, M., Yin, J., Zou, T. T., Shustova, V., Mori, Y., Sato, F., Liu, T. C., Olaru, A., Wang, S., Kimos, M. C., Perry, K., Desai, K., Greenwood, B. D., Krasna, M. J., Shibata, D., Abraham, J. M. and Meltzer, S. J. (2002): Artificial neural networks and gene filtering distinguish between global gene expression profiles of Barrett's esophagus and esophageal cancer. Cancer Research, 62:3493-3497.

classification;MLP;SASOM;gene expression profile;SVM;KNN;Biological data mining;ensemble classifier;feature selection

134

Machine Learning in Low-level Microarray Analysis

Machine learning and data mining have found a multitude of successful applications in microarray analysis, with gene clustering and classification of tissue samples being widely cited examples. Low-level microarray analysis often associated with the pre-processing stage within the microarray life-cycle has increasingly become an area of active research, traditionally involving techniques from classical statistics. This paper explores opportunities for the application of machine learning and data mining methods to several important low-level microarray analysis problems: monitoring gene expression, transcript discovery, genotyping and resequencing . Relevant methods and ideas from the machine learning community include semi-supervised learning, learning from heterogeneous data, and incremental learning.

INTRODUCTION DNA microarrays have revolutionized biological research over the short time since their inception [2; 27; 28; 29]. Although most widely used for parallel measurement of gene expression [27; 28], microarrays are starting to find common application in other areas of genomics and transcriptomics, including genomic re-sequencing [30; 31], genotyping [32; 33], and transcript discovery [34]. Research labs armed with microarrays have been able to partake in a range of studies, including finding gene function [35; 36; 37]; correcting mistaken database annotations [36; 7]; performing linkage analyses; determining specific genes involved in biological pathways; identifying genes that are important at certain times of development (or that are turned on/off over a course of treatment); elucidating gene regulatory networks [13]; diagnosing disease in tissue sam-Figure 1: The relationship between low-level and high-level microarray analysis. ples [38; 39; 40; 41]; tioners' misdiagnoses [38]. The common thread among these high-level microarray analysis problems is that they answer sophisticated questions of direct biological interest to medical researchers (such as "which genes are being co-expressed under treatment X?"), where the raw data used are estimates of biologically meaningful parameters (such as the expression level estimates for thousands of genes). In contrast to these so-called high-level problems, low-level microarray analysis [19] is concerned with the preceding step in the microarray assay cycle (Figure 1) given raw data straight from a scanner which has no direct biological interpretation , clean and summarize this data to produce the biologically meaningful parameter estimates (such as expression level estimates) that are later used in high-level analyses . In low-level analysis, more consideration is generally given to the behavior of the underlying molecular biology, microarray technology, and experimental design than in high-level analysis . This makes generative methods readily applicable in low-level problems, facilitating the formulation of confidence SIGKDD Explorations. and even identifying medical practi statements such as p-values in gene expression calls. Hence, while high-level problems have been tackled with discriminative approaches, such as those found in machine learning and data mining, in addition to classical statistical methods, the low-level analysis community has traditionally called upon only the latter. In this paper we argue that low-level microarray analysis poses a number of interesting problems for the data mining and machine learning community, distinct to the traditional high-level microarray problems. These problems are relevant to the long-term success of DNA microarrays and are already topics of active research in the low-level microarray analysis community. It is our hope that this position paper motivates and enables further machine learning research in the area. Although we will focus on high density oligonucleotide microarrays, particularly those of the Affymetrix GeneChip variety, the underlying concepts and opportunities remain the same for related technologies. Throughout the paper, we distinguish machine learning from statistics. While these disciplines are closely related and serve as foundations for inference in microarray analysis, the distinction does have content. In our view, classical statistics is generative , dealing with relatively low-dimensional data and parameter spaces, while machine learning is often discriminative in nature and explicitly addresses computational issues in high-dimensional data analysis. Section 2 reviews relevant background ideas from machine learning. For an overview of the background molecular biology and microarray technology, see the guest editorial elsewhere in this issue. The low-level problems of absolute and differential expression level summarization, expression detection , and transcript discovery are reviewed in Section 3, along with suggested applications of machine learning approaches to these problems. Sections 4 and 5 similarly cover microarray-based genotyping and re-sequencing. Finally, Section 6 concludes the paper. BACKGROUND MACHINE LEARNING We assume familiarity with the notions of unsupervised learning (clustering) and supervised learning (classification and regression). As many of the low-level analysis problems discussed below are amenable to learning from partially labeled data, learning from heterogeneous data, and incremental learning, we briefly review these paradigms here. 2.1 Learning from Partially Labeled Data Given an i.i.d. labeled sample {(x i , y i )} n i=1 drawn from the unknown and fixed joint distribution F (x, y), and an i.i.d. unlabeled sample {x i } m i=n+1 drawn from the marginal distribution F (x), the problem of learning from partially labeled data [22; 20] is to use the data in choosing a function ^ g m (X) approximating E(Y |X) where (X, Y ) F . This problem has been motivated by a number of applications where only limited labeled data is present, say due to expense, while unlabeled data is plentiful [16]. This is particularly the case in the areas of text classification, medical research, and computer vision [42], within which much of the research into learning from partially labeled data has occurred. This problem, also called the labeled-unlabeled data problem [42], has been explored under a number of closely-related guises. Some of the earliest approaches used so-called hybrid learners [6], where an unsupervised learning algorithm assigns labels to the unlabeled data, thereby expanding the labeled dataset for subsequent supervised learning. The term multimodal learning is sometimes used to refer to partially labeled learning in the computer vision literature [17]. Co-training is a form of partially labeled learning where the two datasets may be of different types and one proceeds by using the unlabeled data to bootstrap weak learners trained on the labeled data [16]. More recently, semi-supervised learning [25] and transductive learning [26] have gained popularity. Equivalent to partially labeled learning, semi-supervised learning includes a number of successful algorithms, such as those based on the support vector machine (SVM) [25; 8]. Transductive learners , on the other hand, aim to predict labels for just the unlabeled data at hand, without producing the inductive approximation ^ g m . This approach can be used to generalize the aforementioned hybrid learners, whose unsupervised step typically ignores the labeled data. In particular , it is shown in [26] that direct transduction is more effective than the traditional two-step approach of induction followed by deduction. A number of transductive schemes have been proposed, such as those based on the SVM [4; 25], a graph-based transductive learner [9], and a leave-one-out error ridge regression method [26]. Joachims [25] describes an approximate solver for the semi-supervised SVM which utilizes a fast SVM optimizer as an inner loop. The story is not all good. [10] tells us that while unlabeled data may be useful, labeled examples are exponentially more valuable in a suitable sense. [43] tells us that unlabeled data may lead the transductive SVM to maximize the wrong margin , and in [42] it is shown that unlabeled data may in fact degrade classifier performance under certain conditions relating the risk and empirical risk. Nonetheless, learning from partially labeled data has enjoyed great success in many theoretical and empirical studies [16; 42; 44; 43]. We are especially interested in partially labeled learning as an approach to the low-level microarray analysis problems discussed in Sections 35, where we have relatively few labeled examples but an abundant source of unlabeled data. [45] is a recent example of partially labeled learning applied to high-level microarray analysis. There, the problem of predicting gene function is tackled using a semi-supervised scheme trained on a two-component dataset of DNA microarray expression profiles and phylogenetic profiles from whole-genome sequence comparisons. This leads us to the next relevant idea from machine learning. 2.2 Learning from Heterogeneous Data Learning from heterogeneous data is the process of learning from training data, labeled or not, that can be partitioned into subsets, each of which contains a different type of data structure or originates from a different source. This notion is equivalent to the methods of data fusion [5]. Research into learning from heterogeneous data tends to be quite domain-specific and has enjoyed increasing interest from the bioinformatics community in particular (e.g., [18]). [46] presents a kernel-based framework for learning from heterogeneous descriptions of a collection of genes, proteins or other entities. The authors demonstrate the method's superiority to the homogeneous case on the problem of predicting yeast protein function using knowledge of amino acid sequence, protein complex data, gene expression data, and known protein-protein interactions. SIGKDD Explorations. Volume 5,Issue 2 - Page 131 [37] proposes an SVM method for classifying gene function from microarray expression estimates and phylogenetic profiles . This is achieved through the construction of an explicitly heterogeneous kernel: first separate kernels are constructed for each data type, taking into account high-order within-type correlations, then these kernels are combined, ignoring high-order across-type correlations. Our interest in learning from heterogeneous data arises because several sources of knowledge relevant to low-level microarray analysis are available, and incorporating such problem domain knowledge has been shown to improve the performance of learning algorithms in the past. 2.3 Incremental Learning Incremental learning is focused on learning from data presented sequentially, where the model may be required to make predictions on unseen data during training. This is in contrast to cases where all training occurs before any predictions are made (batch learning ), and is similar to online learning [24]. A number of incremental learning algorithms have been proposed and applied in the literature. For example, several incremental support vector machines have been studied [24; 21; 47]. In [48], incremental learning is applied to distributed video surveillance. SVM algorithm parameter selection is investigated in [47]. [21] applies an incremental SVM to detecting concept drift the problem of varying distributions over long periods of data gathering and to adaptive classification of documents with respect to user interest. An exact incremental SVM is proposed in [24], where decremental unlearning of incremental training data is possible. This can be used to efficiently evaluate the computationally-expensive leave-one-out error measure. Due to the relatively small sizes of datasets typically available in low-level microarray analysis, there is great potential for learners that can incrementally incorporate new data gathered in the lab, thereby improving estimator performance specific to that lab's patterns of microarray assay. EXPRESSION ANALYSIS The most successful application of DNA microarray technology to date has been to gene expression analysis. Tra-ditionally , this has involved estimating gene expression levels (Section 3.1), an area that is being addressed through successful statistical methods and active statistics research. However, the task of determining transcription activity over entire chromosomes (Section 3.2) is less well developed and offers serious opportunities for machine learning. 3.1 Gene Expression Monitoring 3.1.1 The Problem Traditional microarrays measure mRNA target abundance using the scanned intensities of fluorescence from tagged molecules hybridized to substrate-attached probes [29]. The brighter the intensity within a cell of identical probes, the more hybridization there has been to those probes (Figure 2a). The scanned intensity, then, roughly corresponds to target abundance. Since probes are limited in length while targets may be thousands of bases long, the GeneChip uses a set of probes to detect each target nucleic acid. The probes are spread out Figure 2: Probe-level features for expression level summarization : (a) a cell of probes; (b) target transcript, perfect match probe and mis-match probe sequences; and (c) scanned and image-analyzed probe-level intensities. along a 600 base pair region close to the 3' end of the transcript . To measure the effects of cross-hybridization, or un-intended hybridization of target A to the probes intended for target B, a system of probe pairs is used. In each pair, a perfect match (PM) probe contains the target's exact complementary sequence, while a mismatch (MM) probe replaces the middle base of the perfect match probe with its Watson-Crick complement. In this way, a target is probed by a probe set of 11-20 PM-MM probe pairs. The aim is roughly for the PMs to measure signal plus noise and for the MMs to measure just noise, so that the signal is revealed using some function of (PM - MM). Figure 2b depicts the probe set arrangement , while Figure 2c gives an example of the scanned intensities. We may now define the expression level summarization problem. Low-level Problem 1. Given a probe set's intensities (possibly after background correction and normalization), the expression level summarization problem is to estimate the amount of target transcript present in the sample. While the expression level summary aims to estimate gene expression level from the features of Figure 2, expression detection is concerned with determining the presence of any gene expression at all. Low-level Problem 2. Given a probe set's intensities, possibly normalized, the expression detection problem is to predict whether the target transcript is present (P) or absent (A) in the sample, or otherwise call marginal (M) if it is too difficult to tell. In addition to the P/M/A detection call, we wish to state a confidence level in the call, such as a p-value. Detection calls are not as widely utilized as expression level estimates. They are often used, for example, to filter out genes with negligible expression before performing computationally -expensive high-level analyses, such as clustering on gene expression profiles. SIGKDD Explorations. Volume 5,Issue 2 - Page 132 The previous two problems dealt with estimates based on a single probe-set read from a single array. Comparative studies , on the other hand, involve assaying two arrays, one the baseline and the other the experiment, followed by computation of a single comparative estimate. Low-level Problem 3. Given two sets of intensities, possibly normalized, for the same probe set on two arrays: a. The differential expression level summarization problem is to estimate the relative abundance of target transcript on each array. b. The comparison call problem is to predict whether the expression of the target has increased, not changed, or decreased from one chip to the other. As in Low-level Problem 2, a statement of confidence in the call should be supplied. The log-ratio of expression levels for a target is sometimes known as the relative expression level [3] and is closely related to the notion of fold change (which is sign(log-ratio) 2log-ratio). Comparison calls are sometimes referred to as change calls. An advantage of working with these comparative estimates is that probe-specific affinities (one cause of undesired variation) are approximately cancelled out by taking ratios [3]. All of these problems are complicated by exogenous sources of variation which cloud the quantities we are interested in. [49] proposes a breakdown of the sources of variation in microarray experiments into intrinsic noise (variation inherent in the experiment's subjects), intermediate noise (arising for example from laboratory procedures), and measurement error (variation due to the instrumentation, such as array manufacture, scanning, or in silico processing). 3.1.2 Current Approaches At the level of microarray design, sophisticated probe modeling and combinatorial techniques are used to reduce probe-specific effects and cross-hybridization. However, much of the unwanted variation identified above must still be tackled during low-level analysis. This means that care must be taken with the relevant statistical issues. For example, in experimental design, we must trade off between biological replicates (across samples) and technical replicates (one sample across chips). Background correction and normalization , for reducing systematic variation within and across replicate arrays, also surface as major considerations [19; 11]. Three popular approaches to Low-level Problem 1 [11] are the Affymetrix microarray suite (MAS) 5.0 signal measure [14; 3; 1], the robust multi-array average (RMA) [50; 11] and the model-based expression index (MBEI) [51]. MAS5 first performs background correction by subtracting a background estimate for each cell, computed by partitioning the array into rectangular zones and setting the background of each zone to that zone's second-percentile intensity. Next MAS5 subtracts an "ideal mismatch value" from each PM intensity and log-transforms the adjusted PMs to stabilize the variance. A robust mean is computed for the resulting values using a biweight estimator, and finally this value is scaled using a trimmed mean to produce the signal estimate. RMA proceeds by first performing quantile normalization [52], which puts the probe intensity distributions across replicate arrays on the same scale. RMA then models the PMs Figure 3: An ROC curve: (0, 0) and (1, 1) correspond to the "always negative" and "always positive" classifiers respectively . The closer to the ideal point (0, 1) the better. Neither of the two families A or B dominates the other. Instead , one or the other is better according to the desired trade-off between FP and TP. as background plus signal, where the signal is exponentially and the background normally distributed MM intensities are not used in RMA. A robust additive model is used to model the PM signal (in log-space) as the sum of the log scale expression level, a probe affinity effect, and an i.i.d. error term. Finally, median polish estimates the model parameters and produces the log-scale expression level summary . MBEI fits P M i,j -M M i,j = i j + i,j , using maximum likelihood to estimate the per-gene expression levels i . Here the j are probe-specific affinities and the i,j are i.i.d. normal errors. Although it may seem that expression detection is just a matter of thresholding expression level estimates, this has proven not to be the case [53]. It is known that expression level estimators often have difficulty at low levels of expression , while detection algorithms are designed with this setting in mind. The most widely used detection algorithm for the GeneChip is a method based on a Wilcoxon signed-rank test [54; 3; 55]. This algorithm corresponds to a hypothesis test of H 0 : median( P M i -M M i P M i +M M i ) = versus H 1 : median( P M i -M M i P M i +M M i ) &gt; , where is a small positive constant. These hypotheses correspond to absence and presence of expression, respectively . The test is conducted using a p-value for a sum of signed ranks R i = P M i -M M i P M i +M M i - . The p-value is thresholded so that values in [0, 1 ), [ 1 , 2 ), and [ 2 , 1] result in present, marginal, and absent calls, respectively. Here 0 &lt; 1 &lt; 2 &lt; 0.5 control the trade-off between false positives (FP) and true positives (TP). Recently, a number of alternate rank sum-based algorithms have been proposed [53]. One in particular a variant on the MAS5 method where scores are set to R i = log P M i M M i has been shown to outperform MAS5 detection in a range of real-world situations. One aspect of the study in [53] of particular interest is the use of the Receiver Operating Characteristic (ROC) Convex Hull method [56] for comparing competing classifiers on a spike-in test set. ROC curves (see Figure 3) characterize the classification performance of a family of classifiers parameterized by a tun-SIGKDD Explorations. Volume 5,Issue 2 - Page 133 able parameter that controls the FP-TP trade-off. For example , as the level of a hypothesis test is decreased, the rate of false positive rejections decreases (by definition), while the rate of false negative acceptances will typically go up. An ROC curve encodes this trade-off, extending the notion of contingency table to an entire curve. It is a more expressive object than accuracy, which boils performance down to one number [56; 57]. Comparing ROC curves has traditionally been achieved by either choosing the "clear winner" (in the rare case of domination [57]), or choosing the maximizer of the Area Under Curve (AUC). Although AUC works in some cases, it gives equal credit to performance over all misclassification cost and class size settings usually an undesirable strategy if any domain knowledge is available. The ROC Convex Hull method, on the other hand, relates expected-cost optimality to conditions on relative misclassification cost and class size, so that the typical case of semi-dominance (as in Figure 3) can be handled in a principled way rather than selecting p-value thresholds by hand, end-users are provided with the right classifier and thresholds by the method. This use of the ROCCH method demonstrates a surprising application of machine learning to low-level microarray analysis. Many of these absolute expression algorithms have their comparative analogues. For example, MAS5 produces the signal log ratio with an associated confidence interval, using a biweight algorithm [14; 3]. MAS5 also implements a comparison call based on the Wilcoxon signed-rank sum test, just as in the absolute MAS5 detection algorithm above [55]. While the Affymetrix microarray suite is the software package bundled with the GeneChip, the Bioconductor project [15] an open-source set of R [12] packages for bioinformatics data analysis has been gaining popularity and implements most of the methods discussed here. 3.1.3 Open Problems While Low-level Problem 1 involves prediction of continuous expression levels (non-negative real values) given a vector of (non-negative real) perfect match and mismatch intensities, with total length between 22 and 40, Low-level Problem 2 is a 3-class classification problem with call confidence levels. Open Problem 1. In the respective settings of Low-level Problems 13: a. What machine learning techniques are competitive with algorithms based on classical statistical methods for expression level estimation? b. Which machine learning classifiers are competitive for expression detection? c. What machine learning methods achieve high performance on the comparative analogues of the previous two problems , posed on the appropriate product space of microarray measurements? Comparisons for expression level estimators might be made based on bias and variance, computational efficiency, and biological relevance of learned models. The ROCCH method is ideal for detector comparison. Issues of background correction and normalization across multiple arrays must likely also be addressed to enable competitiveness with the state of the art. Research into applying semi-supervised, heterogeneous data and incremental learners to gene expression monitoring is directly motivated by the proportion of labeled to unlabeled data available, the existence of GeneChip domain knowledge , and the endemic nature of microarray assays that are continually performed in individual research labs. Biologists could augment the limited labeled probe-level data available with relatively abundant unlabeled data. Labeled data can be procured, for example, from bacterial control experiments with known concentrations, called spike-in assays, and bacterial control probe sets that are present in some GeneChips for calibration purposes. The former source of labeled data is the more useful for this problem, as it provides examples with a range of labels. Unfortunately, spike-in studies are rare because they are not of independent scientific interest: they are only performed for low-level microarray research. For the few spike-in assays that are available, only a small number of targets are spiked in at an equally small number of concentrations (typically 10). Unlabeled data, in contrast, could be taken from the large collection of available biologically relevant assays; each one providing tens of thousands of data points. Beyond probe intensities, other data sources could include probe sequences and probe-affinity information derived from probe models. Such information is closely related to the hybridization process and might be of use in expression level estimation: both target and non-specific hybridization are known to be probe-dependent. Although labeled data from spike-in studies are of greatest utility for learning [10], the quantity of unlabeled data produced by a series of biologically interesting microarray assays in any given lab suggests a semi-supervised incremental approach. Since the ROCCH involves taking a pointwise maximum over the individual noisy ROC curves, it incorporates a possibly large degree of uncertainty. It should be possible to extend the results of [53] to quantify this property. Open Problem 2. Can the ROC Convex Hull method of [56] be extended to provide confidence intervals for its conditions on expected-cost optimality? 3.2 Transcript Discovery 3.2.1 The Problem The applications to expression monitoring described above are all related to addressing questions about pre-defined transcripts. More precisely, the vast majority of expression analysis is performed using probes interrogating only a small sub-sequence of each transcript. This has clearly been a useful approach, but there are at least two potential drawbacks. One is that we can only monitor the expression of genes known to exist at the time of the array's design. Even in a genome as well-studied as that of the human, new transcripts are routinely discovered. Another is that in directly monitoring only a sub-sequence of the transcript, it will often be impossible to distinguish between alternatively spliced forms of the same gene (which may have very different functional roles). An alternative approach is to use arrays with probes tiled uniformly across genomic sequence, without regard to current knowledge of transcription. Such genome tiling arrays have been used to monitor expression in all the non-repetitive sequence of human chromosomes 21 and 22 [34], and more widespread use is underway. SIGKDD Explorations. Volume 5,Issue 2 - Page 134 The problems arising in the analysis of data from genome tiling arrays are essentially the same as those for the expression monitoring arrays described above: estimation of expression level, detection of presence, and detection of differential expression. There is, however, the additional challenge of determining the number of distinct transcripts and their location within the tiled genomic region. Low-level Problem 4. The problem of transcript discovery can be viewed in two steps: a. Determining the exon structure of genes within a tiled region; and b. Determining which exons should be classified together as part of a single gene's transcript. 3.2.2 Current Approaches A simple heuristic approach is taken in [34], in which PM-MM probe pairs are classified as positive or negative based on thresholds applied to the difference and ratio of the PM and MM values. Positions classified as positive and located close to other positive positions are grouped together to form predicted exons. A more effective approach [58] is based on the application of a Wilcoxon signed-rank test in a sliding window along the genomic sequence, using the associated Hodges-Lehmann estimator for estimation of expression level. Grouping into exons is achieved by thresholding on present call p-values or estimated expression level, then defining groups of probes exceeding the threshold to be exons. 3.2.3 Open Problems The problem of detecting exons based on probe intensities (Low-level Problem 4a) is very similar to the problem of absolute expression detection (Low-level Problem 2). For example, the exon detection method of [58] and the MAS5 expression detection algorithm [55] are both built around the Wilcoxon signed-rank test. The problem of finding exons has been addressed as described, but the methods are heuristic and there is plenty of room for improvement. Associating exons to form transcripts (Low-level Problem 4b) has been addressed in a large experiment across almost 70 experimental pairs using a heuristic correlation-based method; again, this presents an opportunity for research into more effective methods. Open Problem 3. Are there machine learning methods that are able to out-perform current classical statistical methods in transcript discovery as defined in Low-Level Problem 4? One possibility which appears well suited to the problem is the use of hidden Markov models where the underlying un-observed Markov chain is over states representing expressed versus non-expressed sequence. The distribution of the observed probe intensities would depend on the underlying hidden state. Another possible approach, considering the success which has been demonstrated in predicting genes from sequence data alone, would also be to integrate array-derived data with sequence information in prediction of transcripts. GENOTYPING Descriptions of genome sequencing efforts such as the human genome project often lend the impression that there is a unique genomic sequence associated with each species. This is a useful and approximately correct abstraction. But in fact, any two individuals picked at random from a species population will have differing nucleotides at a small fraction of the corresponding positions in their genomes. Such single-nucleotide polymorphisms, or SNPs, help form the basis of genetically-determined variation across individuals. Biologists estimate that about one position in 1,000 in the human genome is a SNP. With over 3 billion bases of genomic DNA, we see that SNPs number in the several millions. Although there are other kinds of individual genomic variation, such as insertions, deletions, and duplications of DNA segments, our focus here is SNPs. Further complicating the picture is the fact that humans are diploid organisms--each person possesses two complete but different copies of the human genome, one inherited from the mother and one from the father. Now consider a polymorphic position, or locus, at which two different bases occur in the population, say G and T. These variants are called the alleles at the locus, so in this case we are describing a biallelic SNP. A given individual will have inherited either a G or T in the paternal genome, and the same is true of the maternal genome. Thus there are three possible genotypes, or individual genetic signatures, at this SNP: they are de-noted GG, TT, and GT. We do not distinguish the last case from TG, since there is no inherent ordering of the paternal and maternal genomes at a given polymorphic position. We refer generically to the alleles of a biallelic SNP as A and B. Biological evidence suggests that essentially all SNPs are biallelic in humans. The genotyping problem, then, is to establish an individual's genotype as AA, BB, or AB for as many SNPs as possible in the human genome. The completion of the human genome project means that one has recourse to the full genomic sequence surrounding a SNP to help solve the genotyping problem. Furthermore, various large-scale public projects to locate SNPs and identify their alleles exist, notably The SNP Consortium (TSC); the data they generate may also be utilized for genotyping. The major drawback to traditional genotyping protocols are their lack of parallelism, with consequent expense in terms of material and labor. In contrast, Kennedy et al. [33] describe whole-genome sampling analysis (WGSA), which enables massively parallel genotyping via genotyping microarrays . For the Affymetrix Mapping 10k Array, which genotypes approximately 10,000 SNPs across the human genome, each SNP actually has 56 corresponding probes, collectively termed a miniblock. The miniblock has 7 probe quartets for the SNP's flanking region on the forward strand and another 7 probe quartets for the reverse complement strand, so 4 7 2 yields 56 probes. Each probe quartet in turn corresponds to a 25-mer in which the SNP is at one of 7 offsets from the central position. The four probes within a probe quartet differ in the base they put at the SNP: a perfect match to the A allele, a perfect match to the B allele, and mismatches for each. Low-level Problem 5. Given a SNP's 56-vector of miniblo-SIGKDD Explorations. Volume 5,Issue 2 - Page 135 ck probe intensities, the genotype calling problem is to predict the individual's corresponding alleles as AA, BB or AB. Write PM(A), PM(B), MM(A), and MM(B) for the probe intensities within a quartet. We would then hope that an AA individual has PM(A) &gt; MM(A) but PM(B) MM(B), for all probe quartets on both strands. For a BB individual, we hope to find just the opposite effect, and an AB individual should have both PM(A) &gt; MM(A) and PM(B) &gt; MM(B). The mismatch probes in each quartet act as controls , establishing the level of nonspecific hybridization for their corresponding perfect match probes. The presence of multiple probe quartets allows for the determination of genotype even when one strand and/or some offsets do not yield reliable hybridization, say for biochemical reasons. 4.2 Current Approaches Low-level Problem 5 is a three-class classification problem. In many machine learning applications, the metric of interest for competing classifiers is predictive accuracy, in this case the probability of correctly genotyping a new individual's SNP based on the miniblock vector. However, in the kinds of genetic studies which take large numbers of genotypes as input, there is usually an explicit requirement that genotype predictions have a prespecified accuracy, often 99%. To attain such accuracy, it is usually permissible for some fraction of genotypable SNPs to be no-calls; that is, the classifier can refuse to predict a genotype for some miniblocks. When comparing genotypers, our interest therefore lies in the trade-off between the rate of no-calls and the accuracy attained on those SNPs which are called. For example, some studies consider the punt rate, or lowest no-call rate which yields a prespecified accuracy level on the called SNPs. A simple unsupervised approach to training a genotyper is to ignore available labels during training, instead using these labels to subsequently assess the trade-off between accuracy and no-call rate for the trained model. This is the strategy pursued by MPAM (modified partitioning around medoids) [59], the discriminative clustering genotyper used for the Affymetrix 10k array. An alternative approach, using a parametric generative model for the clustering, will be described elsewhere. It resembles ABACUS, a model studied in the context of re-sequencing microarrays [31] (see Section 5). 4.3 Open Problems Open Problem 4. Are there machine learning methods that are able to meet typical accuracy and punt-rate specifications on the genotype calling problem? In order to choose a genotyper using supervised learning, we need labels (true genotypes) along with corresponding miniblock reads from genotyping arrays. Unfortunately, there is no large-scale set of publicly available genotypes. Instead , one makes do with modestly-sized sets of genotypes available commercially from companies using smaller-scale techniques. Of course, no genotyping method is error-free, so in practice one measures concordance with reference genotypes . If the concordance is high enough, the remaining cases of disagreement between a candidate genotyper and the reference genotypes can be resolved via the older labor-intensive methods. The incomplete nature of reference genotype data leads naturally to the setting of semi-supervised learning. Rather than falling back to unsupervised methods such as those described above, we may consider employing more general semi-supervised learners as described in Section 2.1. Additionally, the methods of [23] could be used to incorporate low-level physical parametric models of hybridization into a kernel-based classifier. RE-SEQUENCING As explained in Section 4, within a single species genomic sequence will vary slightly from one individual to the next. While Low-level Problem 5 focuses on the determination of genotype at a position known in advance to be polymorphic, the problem described in this section concerns locating such polymorphic sites in the first place. The usual starting point is a newly-sequenced genome, such as the recently-finished human genome. It is often the case that, based on previous research, an investigator will be interested in detailed study of variation in a particular genomic region (say on the order of tens or hundreds of kilo-bases ) and wants to re-sequence this region in a large number of individuals. Such re-sequencing allows for identification of the small subset of polymorphic locations. Here we consider the more recent challenges of microarray-based re-sequencing of diploid genomic DNA. A typical re-sequencing array uses eight probes to interrogate each base of the monitored sequence. These eight probes comprise two quartets, one for the forward strand and one for the reverse. Each quartet is formed of 25-mer probes perfectly complementary to the 25 bases of the reference sequence centered on the interrogated base, but with all four possible bases used at the central position. Low-level Problem 6. The goal of the re-sequencing problem is to start with a set of probe intensities and classify each position as being one of A, C, G, T, AC, AG, AT, CG, CT, GT, or N, where N represents a `no call' (due to sample failure or ambiguous data). The intuition is that for a homozygous position, one of the four probes should be much brighter relative to the others on each strand, and for a heterozygous position, two probes corresponding to the two bases of a SNP should be brighter on each strand. Of particular interest are positions in which the called base is heterozygous, or homozygous and different to the reference sequence, as such positions exhibit polymorphism and are candidate positions for explaining phenotypic differences between individuals. At face value, this classification problem is much harder than the genotyping problem. There are fewer probes to start with (a miniblock of 8 rather than 40 or more) and more categories (11 as opposed to 3 or 4) into which to classify. 5.2 Current Approaches The most recent analysis of the kind of re-sequencing array discussed here [31] is based on modeling pixel intensities within each probe as independent random variables with a common mean and variance. The model for a homozygous base is that, on each strand, the probe correspond-three probes have another. The means and variance are estimated by maximum likelihood, and the likelihood of the SIGKDD Explorations. ing to the base has one mean and variance, and the other Volume 5,Issue 2 - Page 136 model is evaluated. The model for each of the six heterozygous possibilities is similar, except two probes correspond to each heterozygote model and the other two are background. The likelihoods (overall and for each strand) are converted to scores and, provided the maximum score exceeds some threshold, the best-scoring model is chosen as the base call. A number of other filters that deal with the signal absence, signal saturation, sample failure, and so on are applied, as is an iterative procedure to account for bias in the background probes. This method, called ABACUS, was found to make base calls at over 80% of all bases, with an estimate accuracy in excess of 99% at the bases which were called. 5.3 Open Problems A good base-calling method for re-sequencing arrays already exists in ABACUS, but there remains room for improvement . A recent and improved implementation [60] of the ABACUS method on a new genomic region found the overall sequencing accuracy to be on the order of 99.998%, but the accuracy on heterozygote calls to be about 96.7%. Biologists would value highly an improvement in heterozygote call accuracy. Open Problem 5. Can a supervised learning method be used to call bases in re-sequencing arrays with accuracy, in particular heterozygote accuracy, in excess of the accuracies achieved by the more classic statistical approaches used to date? Considering the ongoing efforts of SNP detection projects, there is an abundance of labeled data available, so the problem seems quite amenable to machine learning approaches. As with the genotyping problem, it would be desirable to have a measure of confidence associated with base calls. It may also be useful to take into account the sequences of the 25-mer probes, as there are known sequence-specific effects on the probe intensities. CONCLUSIONS We have described a variety of low-level problems in microarray data analysis and suggested the applicability of methods from several areas of machine learning. Some properties of these problems which should be familiar to machine learning researchers include high-dimensional observations with complicated joint dependencies (probe intensities ), partially labeled data sets (expression levels, genotypes ), data from disparate domains (microarray assays, probe sequences, phylogenetic information), and sequential observations (ongoing experimental work at individual labs). We pointed out the suitability of semi-supervised, heterogeneous , and incremental learning in these settings. It is worth remarking that analogous problems arise with other high-throughput technologies, such as cDNA and long oligonucleotide microarrays, mass spectrometry, and fluorescence-activated cell sorting. There are other issues in low-level analysis we did not cover. Here we mention two of these. Image analysis is the problem of going from raw pixel values in the scanned image of a microarray to a set of pixel intensities for each feature placed on the probe, and then to single-number probe intensities. The surface of the GeneChip contains detectable grid points which facilitate rotation and translation of the image to a canonical alignment; subsequent mapping of each pixel to a feature is semi- or fully automated and has not previously raised major analysis issues. However, work is being done on aggressive reduction of feature sizes to a scale where this mapping procedure could become a central concern. On the more theoretical side, probe models based on the physics of polymer hybridization have recently been the focus of considerable interest. These models reflect a significant increase in the use of biological knowledge for estimating target abundance and present an opportunity for application of machine learning techniques which can exploit parametric distributions in high-dimensional data analysis, such as graphical models. We close by observing that a fuller awareness of low-level microarray analysis issues will also benefit machine learning researchers involved with high-level problems: the inevitable information reduction from earlier stage to later could well conceal too much of what the unfiltered array data reveal about the biological issue at hand. Familiarity with initial normalization and analysis methods will allow the high-level analyst to account for such a possibility when drawing scientific conclusions. ACKNOWLEDGMENTS We thank Rafael Irizarry, Ben Bolstad, Francois Collin and Ken Simpson for many useful discussions and collaboration on low-level microarray analysis. REFERENCES [1] Affymetrix. Affymetrix Microarray Suite Guide. Affymetrix Inc., Santa Clara, CA, 2001. version 5.0. [2] M. Schena. DNA Microarrays: A Practical Approach. Oxford University Press, 1999. [3] Affymetrix. Statistical algorithms description document . Whitepaper, Affymetrix Inc., Santa Clara, CA, 2002. [4] A. Gammerman, V. Vovk, and V. Vapnik. Learning by transduction. In Fourteenth Conference on Uncertainty in Artificial Intelligence, pages 148155. Morgan Kaufmann Publishers, 1998. [5] P. K. Varshney. Scanning the issue: Special issue on data fusion. Proceedings of the IEEE, 85:35, 1997. [6] R. O. Duda and P. E. Hart. Pattern Classification and Scene Analysis. John Wiley and Sons, New York, 1973. [7] T. Gaasterland and S. Bekiranov. Making the most of microarray data. Nature Genetics, 24:204206, 2000. [8] K. P. Bennett and A. Demiriz. Semi-supervised support vector machines. In Advances in Neural Information Processing Systems 11, pages 368374, Cambridge, MA, 1999. MIT Press. [9] T. Joachims. Transductive learning via spectral graph partitioning. In Proceedings of the International Conference on Machine Learning (ICML), 2003. [10] V. Castelli and T. Cover. On the exponential value of labeled samples. Pattern Recognition Letters, 16:105 111, 1995. SIGKDD Explorations. Volume 5,Issue 2 - Page 137 [11] R. A. Irizarry. Science and Statistics: A Festschrift for Terry Speed, volume 40 of Lecture NotesMonograph Series, chapter Measures of gene expression for Affymetrix high density oligonucleotide arrays, pages 391402. Institute of Mathematical Statistics, 2003. [12] R. Ihaka and R. Gentleman. R: A language for data analysis and graphics. Journal of Computational and Graphical Statistics, 5(3):299314, 1996. [13] N. Friedman. Probabilistic models for identifying regulation networks. Bioinformatics, 19:II57, October 2003. [14] E. Hubbell, W. M. Liu, and R. Mei. Robust estimators for expression analysis. Bioinformatics, 18:15851592, 2002. [15] Bioconductor Core. An overview of projects in computing for genomic analysis. Technical report, The Bioconductor Project, 2002. [16] A. Blum and T. Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the Workshop on Computational Learning Theory. Morgan Kaufmann Publishers, 1998. [17] L. Wu, S. L. Oviatt, and P. R. Cohen. Multimodal integration - a statistical view. IEEE Transactions on Mul-timedia , 1:334 341, 1999. [18] A. J. Hartemink and E. Segal. Joint learning from multiple types of genomic data. In Proceedings of the Pacific Symposium on Biocomputing 2004, 2004. [19] G. K. Smyth, Y. H. Yang, and T. P. Speed. Functional Genomics: Methods and Protocols, volume 224 of Methods in Molecular Biology, chapter Statistical issues in cDNA microarray data analysis, pages 111136. Hu-mana Press, Totowa, NJ, 2003. [20] A. Blum and S. Chawla. Learning from labeled and unlabeled data using graph mincuts. In International Conference on Machine Learning (ICML), 2001. [21] R. Klinkenberg and T. Joachims. Detecting concept drift with support vector machines. In P. Langley, editor , Proceedings of ICML-00, 17th International Conference on Machine Learning, pages 487494, Stanford, CA, 2000. Morgan Kaufmann Publishers. [22] M. Szummer and T. Jaakkola. Partially labeled classification with Markov random walks. In Neural Information Processing Systems (NIPS), 2001. [23] T. S. Jaakkola and D. Haussler. Exploiting generative models in discriminative classifiers. In Advances in Neural Information Processing Systems 11: Proceedings of the 1998 Conference, pages 487493. MIT Press, 1998. [24] G. Cauwenberghs and T. Poggio. Incremental and decremental support vector machine learning. In NIPS, pages 409415, 2000. [25] T. Joachims. Transductive inference for text classification using support vector machines. In I. Bratko and S. Dzeroski, editors, Proceedings of the 16th Annual Conference on Machine Learning, pages 200209. Morgan Kaufmann, 1999. [26] O. Chapelle, V. Vapnik, and J. Weston. Advances in Neural Information Processing Systems 12, chapter Transductive inference for estimating values of functions . MIT Press, 2000. [27] M. Schena, D. Shalon, R. W. Davis, and P. O. Brown. Quantitative monitoring of gene expression patterns with a complementary DNA microarray. Science, 270:467470, 1995. [28] D. J. Lockhart, H. Dong, M. C. Byrne, M. T. Follet-tie , M. V. Gallo, M. S. Chee, M. Mittmann, C. Wang, M. Kobayashi, H. Horton, and E. L. Brown. Expression monitoring by hybridization to high-density oligonucleotide arrays. Nature Biotechnology, 14:16751680, 1996. [29] R. J. Lipshutz, S. P. A. Fodor, T. R. Gingeras, and D. H. Lockhart. High density synthetic oligonucleotide arrays. Nature Genetics, 21:2024, 1999. Supplement. [30] J. B. Fan, D. Gehl, L. Hsie, K. Lindblad-Toh, J. P. Laviolette, E. Robinson, R. Lipshutz, D. Wang, T. J. Hudson, and D. Labuda. Assessing DNA sequence variations in human ests in a phylogenetic context using high-density oligonucleotide arrays. Genomics, 80:351 360, September 2002. [31] D. J. Cutler, M. E. Zwick, M. M. Carrasquillo, C. T. Yohn, K. P. Tobin, C. Kashuk, D. J. Mathews, N. A. Shah, E. E. Eichler, J. A. Warrington, and A. Chakravarti. High-throughput variation detection and genotyping using microarrays. Genome Research, 11:19131925, November 2001. [32] J. B. Fan, X. Chen, M. K. Halushka, A. Berno, X. Huang, T. Ryder, R. J. Lipshutz, D. J. Lockhart , and A. Chakravarti. Parallel genotyping of human SNPs using generic high-density oligonucleotide tag arrays . Genome Research, 10:853860, June 2000. [33] G. C. Kennedy, H. Matsuzaki, D. Dong, W. Liu, J. Huang, G. Liu, X. Su, M. Cao, W. Chen, J. Zhang, W. Liu, G. Yang, X. Di, T. Ryder, Z. He, U. Surti, M. S. Phillips, M. T. Boyce-Jacino, S. P. A. Fodor, and K. W. Jones. Large-scale genotyping of complex DNA. Nature Biotechnology, October 2003. [34] P. Kapranov, S. E. Cawley, J. Drenkow, S. Bekiranov, R. L. Strausberg, S. P. A. Fodor, and T. R. Gingeras. Large-scale transcriptional activity in chromosomes 21 and 22. Science, 296:916919, 2002. [35] M. Brown, W. N. Grundy, D. Lin, N. Cristianini, C. Sugnet, M. Ares Jr, and D. Haussler. Support vector machine classification of microarray gene expression data. Technical Report UCSC-CRL-99-09, Department of Computer Science, University of California at Santa Cruz, 1999. [36] M. P. S. Brown, W. N. Grundy, D. Lin, N. Cristianini, C. W. Sugnet, T. S. Furey, M. Ares Jr, and D. Haussler. Knowledge-based analysis of microarray gene expression data by using support vector machines. Proceedings of the National Academy of Sciences, 97:262267, 1997. SIGKDD Explorations. Volume 5,Issue 2 - Page 138 [37] P. Pavlidis, J. Weston, J. Cai, and W. N. Grundy. Gene functional classification from heterogeneous data. In Proceedings of the Fifth International Conference on Computational Molecular Biology, pages 242248, 2001. [38] T. S. Furey, N. Cristianini, N. Duffy, D. W. Bednarski, M. Schummer, and D. Haussler. Support vector machine classification and validation of cancer tissue samples using microarray expression data. Bioinformatics, 16:906914, 2000. [39] S. Mukherjee, P. Tamayo, D. Slonim, A. Verri, T. Golub, J. P. Mesirov, and T. Poggio. Support vector machine classification of microarray data. Technical Report 182, Center for Biological and Computational Learning Massachusetts Institute of Technology, 1998. [40] S. Ramaswamy, P. Tamayo, R. Rifkin, S. Mukherjee, C. Yeang, M. Angelo, C. Ladd, M. Reich, E. Latulippe, J. P. Mesirov, T. Poggio, W. Gerald, M. Loda, E. S. Lander, and T. R. Golub. Multiclass cancer diagnosis using tumor gene expression signatures. Proceedings of the National Academy of Sciences, 98, 2001. [41] C. Yeang, S. Ramaswamy, P. Tamayo, S. Mukherjee , R. M. Rifkin, M. Angelo, M. Reich, E. Lander, J. Mesirov, and T. Golub. Molecular classification of multiple tumor types. Bioinformatics, 1:17, 2001. [42] F. G. Cozman and I. Cohen. Unlabeled data can degrade classification performance of generative classifiers . In Fifteenth International Florida Artificial Intelligence Society Conference, pages 327331, 2002. [43] T. Zhang and F. J. Oles. A probability analysis on the value of unlabeled data for classification problems. In Proceedings of the International Conference on Machine Learning, pages 11911198, 2000. [44] K. Nigam, A. McCallum, S. Thrun, and T. Mitchell. Text classification from labeled and unlabeled documents using EM. Machine Learning, 39:103134, 2000. [45] T. Li, S. Zhu, Q. Li, and M. Ogihara. Gene functional classification by semi-supervised learning from heterogeneous data. In Proceedings of the ACM Symposium on Applied Computing, 2003. [46] G. R. G. Lanckriet, M. Deng, N. Cristianini, M. I. Jordan , and W. S. Noble. Kernel-based data fusion and its application to protein function prediction in yeast. In Proceedings of the Pacific Symposium on Biocomputing 2004, 2004. [47] A. Shilton, M. Palaniswami, D. Ralph, and A. C. Tsoi. Incremental training in support vector machines. In Proceedings of the International Joint Conference on Neural Networks, 2001. [48] C. P. Diehl. Toward Efficient Collaborative Classification for Distributed Video Surveillance. PhD thesis, Department of Electrical and Computer Engineering, Carnegie Mellon University, 2000. [49] J. H. Maindonald, Y. E. Pittelkow, and S. R. Wilson. Science and Statistics: A Festschrift for Terry Speed, volume 40 of IMS Lecture NotesMonograph Series, chapter Some Considerations for the Design of Microarray Experiments, pages 367390. Institute of Mathematical Statistics, 2003. [50] R. A. Irizarry, B. Hobbs, F. Collin, Y. D. Beazer-Barclay , K. J. Antonellis, U. Scherf, and T. P. Speed. Exploration, normalization, and summaries of high density oligonucleotide array probe level data. Bio-statistics , 4:249264, 2003. [51] C. Li and W. H. Wong. Model-based analysis of oligonucleotide arrays: Expression index computation and outlier detection. Proceedings of the National Academy of Science, 98:3136, 2001. [52] B. M. Bolstad, R. A. Irizarry, M. Astrand, and T. P. Speed. A comparison of normalization methods for high density oligonucleotide array data based on bias and variance. Bioinformatics, 19:185193, 2003. [53] B. I. P. Rubinstein and T. P. Speed. Detecting gene expression with oligonucleotide microarrays, 2003. manuscript in preparation. [54] W. Liu, R. Mei, D. M. Bartell, X. Di, T. A. Webster, and T. Ryder. Rank-based algorithms for analysis of microarrays. Proceedings of SPIE, Microarrays: Optical Technologies and Informatics, 4266, 2001. [55] W. M. Liu, R. Mei, X. Di, T. B. Ryder, E. Hubbell, S. Dee, T. A. Webster, C. A. Harrington, M. H. Ho, J. Baid, and S. P. Smeekens. Analysis of high density expression microarrays with signed-rank call algorithms . Bioinformatics, 18:15931599, 2002. [56] F. Provost and T. Fawcett. Analysis and visualization of classifier performance: Comparison under imprecise class and cost distributions. In Third International Conference on Knowledge Discovery and Data Mining, Menlo Park, CA, 1997. [57] T. Fawcett, F. Provost, and R. Kohavi. The case against accuracy estimation for comparing induction algorithms . In Fifteenth International Conference on Machine Learning, 1998. [58] D. Kampa and et al. Novel RNAs identified from a comprehensive analysis of the transcriptome of human chromosomes 21 and 22. Manuscript in preparation. [59] W.-M. Liu, X. Di, G. Yang, H. Matsuzaki, J. Huang, R. Mei, T. B. Ryder, T. A. Webster, S. Dong, G. Liu, K. W. Jones, G. C. Kennedy, and D. Kulp. Algorithms for large scale genotyping microarrays. Bioinformatics, 2003. In press. [60] Affymetrix. GeneChip CustomSeq resequencing array: Performance data for base calling algorithm in GeneChip DNA analysis software. Technical note, Affymetrix Inc., Santa Clara, CA, 2003. http://www.affymetrix.com/support/tech nical/technotes/customseq technote.pdf. SIGKDD Explorations. Volume 5,Issue 2 - Page 139

low-level analysis;data mining;transductive learning;learning from heterogeneous data;heterogeneous data;semi-supervised learning;incremental learning;transcript discovery;microarray;gene expression estimation;statistics;genotyping;Low-level microarray analysis;re-sequencing

135

Measuring Cohesion of Packages in Ada95

Ada95 is an object-oriented programming language. Pack-ages are basic program units in Ada 95 to support OO programming, which allow the specification of groups of logically related entities. Thus, the cohesion of a package is mainly about how tightly the entities are encapsulated in the package. This paper discusses the relationships among these entities based on dependence analysis and presents the properties to obtain these dependencies. Based on these, the paper proposes an approach to measure the package cohesion, which satisfies the properties that a good measure should have.

INTRODUCTION Cohesion is one of the most important software features during its development. It tells us the tightness among the components of a software module. The higher the cohesion of a module, the more understandable, modifiable and maintainable the module is. A software system should have high cohesion and low coupling. Researchers have developed several guidelines to measure cohesion of a module [1, 3, 4]. Since more and more applications are object-oriented, the approaches to measure cohesion of object-oriented (OO) programs have become an important research field. Generally, each object-oriented programming language provides facilities to support OO features, such as data abstraction, encapsulation and inheritance. Each object consists of a set of attributes to represent the states of objects and a set of operations on attributes. Thus, in OO environment, the cohesion is mainly about how tightly the attributes and operations are encapsulated. There are several approaches proposed in literature to measure OO program cohesion [2, 5, 6, 7, 11, 12]. Most approaches are based on the interaction between operations and attributes. The cohesion is measured as the number of the interactions. Generally only the references from operations to attributes are considered. And few care about the interactions of attributes to attributes and operations to operations at the same time. This might lead to bias when measuring the cohesion of a class. For example, when designing the trigonometric function lib class, we might set a global variable to record the temporal result. The variable is referred in all the operations of the class. According to methods based on the interaction between operations and attributes [6, 7], the cohesion is the maximum 1. In fact, there are no relations among the operations if the calls are not taken into account. In this view, its cohesion is 0. The difference is caused by considering only the references from operations to attributes, while not considering the inter-operation relations. In our previous work, we have done some research in measuring OO program cohesion [10, 13, 14]. Our approach overcomes the limitations of previous class cohesion measures, which consider only one or two of the three facets. Since the OO mechanisms in different programming languages are different from each other, this paper applies our measure to Ada packages. The remaining sections are organized as follows. Section 2 introduces the package in Ada 95. Section 3 discusses the basic definitions and properties for our measure. Based on the definitions and properties, Section 4 proposes approaches to measure package cohesion. Conclusion remarks are given in the last section. PACKAGES IN ADA 95 In Ada 95[ISO95], packages and tagged types are basic program units to support OO programming. A package allows the specification of groups of logically related entities. Typically, a package contains the declaration of a type along with the declarations of primitive subprograms of the type, which can be called from outside the package, while its inner workings remain hidden from outside users. In this paper, we distinguish packages into four groups. PG1: Packages that contain any kind of entities except tagged types. PG2: Packages that only contain the declaration of one tagged type along with those primitive subprograms of the type. There are two subgroups in PG2: - PG2-1: The type is an original tagged type. - PG2-2: The type is a derived type. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SIGAda'03, December 711, 2003, San Diego, California, USA. Copyright 2003 ACM 1-58113-476-2/03/0012...$5.00. 62 PG3: Combination of PG1 and PG2. PG4: Generic packages. After a generic package is instantiated, it belongs to one of the former three groups. Thus, only cohesion measure of PG1, PG2 and PG3 is discussed in the paper. DEFINITIONS In this section, we will present our definitions in the form of PG1. The cohesion of a package from PG1 is mainly about how tightly the objects and subprograms are encapsulated in the package. In this paper, the relationships among objects and subprograms are defined as three dependencies: inter-object, inter-subprogram and subprogram-object dependence. Definition 1 In the package body or a subprogram of the package, if the definition (modification) of object A uses (refer, but not modify) object B directly or indirectly, or whether A can been defined is determined by the state of B, then A depends on B, denoted by A B. Generally, if B is used in the condition part of a control statement (such as if and while), and the definition of A is in the inner statement of the control statement, the definition of A depends on B's state. Definition 2 If object A is referred in subprogram P, P depends on A, denoted by P A. Definition 3 There are two types of dependencies between subprograms: call dependence and potential dependence. If P is called in M, then M call depends on P, denoted by M P. If the object A used in M is defined in P, the A used in M depends on the A defined in P, denoted by M A A, P, where (A, A) is named as a tag. For each call edge, add a tag (*, *) for unification. i.e. if P Q, P *,* Q. To obtain these dependencies, we introduce four sets for each subprogram M: IN(M) is an object set, each element of which is an object referred before modifying its value in M; OUT(M) is an object set, each element of which is an object modified in M. DEP_A (M) is a dependence set which represents the dependencies from the objects referred in M to the objects defined outside M. Each element has the form &lt;A, B&gt;, where A and B are objects of the package. DEP_A_OUT(M) is a dependence set which records the dependencies from the objects referred in M to the objects defined outside M when exiting M. In general, the intermediate results are invisible outside, and an object might be modified many times in a subprogram. We introduce DEP_A_OUT to improve the precision. Obviously, DEP_A_OUT(M) DEP_A (M). Property 1 A IN(M), A OUT(P) M A A, P. Property 2 &lt;A, B&gt; DEP_A(M), B OUT(P) M B A, P. Property 3 M B A, P, &lt;B, C&gt;(&lt;B, C&gt; DEP_A_OUT(P), C OUT(Q)) M C A, Q. In our previous work [8, 9], we have proposed methods to analyze dependencies among statements for Ada programs. And these dependencies can be easily transformed to the dependencies proposed in this paper. Due to the space limitation, we do not discuss them in detail here. To present our cohesion measure in a united model, we introduce package dependence graph to describe all types of dependencies. Definition 4 The package dependence graph (PGDG) of a package PG is a directed graph, PGDG = &lt;N, E, T&gt;, where N is the node set and E is the edge set, T is the tag set. N = N O N P , N O is the object node set, each of which represents a unique object; N P is the subprogram node set, each of which represents a unique subprogram. PGDG consists of three sub-graphs: Inter-Object Dependence Graph (OOG), OOG = &lt;N O , E O &gt;, where N O is the object node set (the name of a node is the name of the object it represents); E O is the edge set, if A B, then edge &lt;A, B&gt; E O . Inter-Subprogram Dependence Graph (PPG), PPG = &lt;N P , E P , T&gt;, where N P is the subprogram node set; E P is the edge set which represents the dependencies between subprograms; T (V V) is the tag set, where V is the union of objects and {*}. Subprogram-Object Dependence Graph (POG), POG = &lt;N, E PO &gt;, where N is the node set which represents objects and subprograms; E PO is the edge set representing dependencies between subprograms and objects. If P A, &lt;P, A&gt; E PO. Example1 shows the package Tri, which contains three objects: temp, temp1 and temp2, and four subprograms: sin, cos, tg and ctg. Figure 1 shows the PGDG of the package Tri in Example1 (all the Tags on PPG are (*, *), because there are only call dependencies in this example. We omit the Tags for convenience). Example1: package Tri. package Tri is temp, temp1, temp2: real; function sin (x: real) return real; function cos (x: real) return real; function tg (x: real) return real; function ctg (x: real) return real; end Tri; package body Tri is function sin (x: real) return real is begin temp:=...; return temp; end sin; ... 63 function tg (x: real) return real is begin temp1:=sin(x);temp2:=cos(x); temp:=temp1/temp2; return temp; end tg; ... end Tri; 3.2 Extended Definitions Since there is no object in the package of PG2, the definitions of Section 3 can not be applied to these packages directly. Therefore, this section will extend the definitions of Section 3.1 to a more general model by the following steps: For PG1, if there is an embedded package, the package is taken as an object. For PG2, take the components of the type as objects of the package. Let A, B be object of a type T, M, P primitive subprograms, and Com1 and Com2 are components of T. Then A, B (A.Com1 B.Com2) Com1 Com2. A, P (P A.Com) P Com. A, B, M, P (M 2 . , 1 . Com B Com A P) M 2 , 1 Com Com P. For PG3, take the types as objects of the package. To present our measure in a unified model, we add powers for different objects. PW(O) = others O O PG Cohesioin O O Cohesion 1 object type a is )) ( ( object package a is ) ( where Cohesion (O) is the cohesion of O, PG (O) returns the package containing O. MEASURING PACKAGE COHESION According to the PGDG, this section will propose our method to measure the package cohesion. In the following discussions, we assume package PG contains n objects and m subprograms, where m, n 0. 4.1 Measuring Inter-Object Cohesion Inter-object cohesion is about the tightness among objects in a package. To measure this cohesion, for each object A, introduce a set A_DEP to record the objects on which A depends, i.e. O_DEP(A) = {B| A B, A B}. Let = ) ( _ ) ( ) ( _ A DEP O B B PW A DEP PW . Then, we define the inter-object cohesion as: = ) , _ ( PG O O Cohesion &gt; = = = 1 1 ) ( _ 1 1 ) ( 0 0 1 n n A DEP PW n n A PW n n i i where 1 ) ( _ n A DEP PW represents the degree on which A depends on other objects. If n=0, there is no object in the package, we set it to 0. If n=1, there is one and only one object in the package, the cohesion is its power. 4.2 Measuring Subprogram-Object Cohesion Subprogram-object cohesion is the most important field in measuring cohesion. Until now, there have been several approaches proposed in literature, such as Chae's methods [6, 7]. But most approaches are based on the POG. As we have mentioned above, all these methods describe the object reference in a simple way and subprograms are connected by the objects referred. Whether there are related among these subprograms are not described exactly. Thus, these approaches should be improved to describe these relations. For completeness, we use Co(Prev) to represent a previous cohesion measure, which satisfies Briand's four properties. For each subprogram P, we introduce another two sets: P_O and P_O_OUT. Where P_O(P) records all the objects referred in P. Figure. 1. PGDG of class Tri temp temp1 temp2 (a) OOG sin cos tg ctg (c) PPG sin cos tg ctg temp temp1 temp2 (b) POG 64 P_O_OUT(P) records the objects referred in P, but these objects relate to objects referred by other subprogram, i.e., P_O_OUT(P)={A| B, M (P A B , M M B A, P) A,B '*'}. Let = ) ( _ ) ( _ _ ) ( ) ( ) ( P O P A i P OUT O P A i i i A PW A PW P Then, we define the subprogram-object cohesion as: = ) , _ ( PG O P Cohesion = = = = Others P Prev Co m m A PW A PW n m m i i P O P A i i 1 i ) ( _ ) ( ) ( 1 1 ) ( ) ( 0 0 0 If P_O(P) = , i.e. no objects are referred in P, we set ) (P =0. If the objects referred in P are not related to other subprograms, these objects can work as local variables. It decreases the cohesion to take a local variable for a subprogram as an object for all subprograms. If there is no object or subprogram in the package, no subprogram will depend on others. Thus, 0 ) , _ ( = PG O P Cohesion . 4.3 Measuring Inter-Subprogram Cohesion In the PGDG, although subprograms can be connected by objects, this is not necessary sure that these subprograms are related. To measure the inter-subprogram cohesion, we introduce another set P_DEP(P) = {M| P M} for each P. The inter-subprogram cohesion Cohesion(P_P, PG) is defined as following: = ) , _ ( PG P P Cohesion &gt; = = = 1 1 ) ( _ 1 1 1 0 0 1 m m P DEP P m m m m i i where 1 | ) ( _ | m P DEP P represents the tightness between P and other subprograms in the package. If each subprogram depends on all other subprograms, Cohesion(P_P, PG) = 1. If all subprograms have no relations with any other subprogram, Cohesion(P_P, PG) = 0. 4.4 Measuring Package Cohesion After measuring the three facets independently, we have a discrete view of the cohesion of a package. We have two ways to measure the package cohesion: 1) Each measurement works as a field, the package cohesion is 3-tuple, Cohesion(PG) = &lt; Cohesion(O_O, PG), Cohesion(P_O, PG), Cohesion(P_P, PG)&gt;. 2) Integrate the three facets as a whole = ) (PG Cohesion = = = Others PG Cohesion k m n PG P P Cohesion k m i i i 3 1 ) ( 0 , 0 ) , _ ( * 0 0 where k ( ] 1 , 0 ; k 1 , k 2 , k 3 &gt;0, and k 1 + k 2 + k 3 =1. Cohesion 1 (PG) = Cohesion(O_O, PG) Cohesion 2 (PG) = Cohesion(P_P, PG) Cohesion 3 (PG) = Cohesion(P_O, PG) If n=0, m 0, the package cohesion describes only the tightness of the call relations, thus we introduce a parameter k to constrain it. For the example shown in Figure 1, the cohesion of Tri describes as follows: Cohesion(O_O, Tri)= 1/3 Cohesion(P_O, Tri)=0 Cohesion(P_P, Tri)=1/3 Let k 1 = k 2 = k 3 = 1/3, Co(Prev)=1, then Cohesion(Tri)= 2/9. Briand et al. [3, 4] have stated that a good cohesion measure should be (1) Non-negative and standardization. (2) Minimum and maximum. (3) Monotony. (4) Cohesion does not increase when combining two modules. These properties give a guideline to develop a good cohesion measure. According to the definitions, it is easy to prove our measure satisfies these properties. 4.5 Cohesion for PG2-2 In the hierarchies of types, the derived type inherits the components and primitive subprograms of the super types. Generally, inheritance will increase the coupling and decrease the 65 cohesion. For the package from PG2-2, we will discuss its cohesion in four cases: Case 1: Take the package independently. Case 2: Take all the primitive subprograms and components (contains those from super type) into consideration. Case 3: If the primitive subprograms of the derived type might access the components (or subprogram) of the super type, take these components (or subprogram) as those of the derived type. Case 4: Take the super type as an object of the derived type. The shortcoming of Case 1 is that: It only measures the cohesion of the additional components and primitive subprograms of the derived type, not the complete type. The primitive subprograms in the super type can not access the components of the derived type except dispatched subprograms. Consequently, in Case 2 or 3, the deeper the hierarchy of types is, the smaller the cohesion. And it is hard to design a package which cohesion is big enough. Although we present four cases in this section, none is good enough to describe the cohesion for a package from PG2-2. To measure the cohesion of a derived type, much more aspects should be considered. RELATED WORKS There have several methods proposed in literatures to measure class cohesion. This section gives a brief review of these methods. (1) Chidamber's LCOM1 [0, 2 ) 1 ( m m ], it measures the cohesion by similar methods and non-similar methods. It is a reverse cohesion measure. The bigger the measure, the lower the cohesion. (2) The PPG in Hitz's LCOM2 is represented by an undirected graph. LCOM2 is the number of sub-graphs connected. When there is one and only one sub-graph, he introduces connectivity to distinguish them. (3) Briand's RCI is the ratio of the number of edges on POG to the max interaction between subprograms and objects. (4) Henderson's LCOM3 can be described as follows. ) ( 3 C LCOM = m m A n n j j = 1 | ) ( | 1 1 where (A)= {M| AP_O(M)}, A is attribute and M is method. (5) Chae's CO [6] introduces glue methods, and Xu-Zhou's CO [13] introduces cut set (glue edges) to analyze the interact pattern. These two measures are more rational than other measures. From the introductions above, we can see that All these methods consider the attribute reference in a simple way. Whether the methods are related or not are not described exactly. LCOM1, LCOM2 and LCOM3 are non-standard, because their up-bounds are related to the number of methods in the class. LCOM1 is non-monotonous. The measuring results might be inconsistent with intuition in some cases RCI has the basic four properties proposed by Briand. But it does not consider the patterns of the interactions among its members, neither LCOM1 and LCOM2 nor LCOM3. Chae's CO overcomes most limitations of previous measures. But it is non-monotonous [13]. Xu-Zhou's CO improves Chae's cohesion measure, and makes its result more consistent with intuition. The chief disadvantage of both measures is that they can be applied to connected POG; otherwise the result will always be 0. LCOM1 and LCOM2 measure the cohesion among methods in a class. We can improve the similar function using the dependencies among methods proposed in this paper. LCOM3, Chae and Xu-Zhou's CO measure the cohesion among methods and attributes in a class. In this paper we improve them by introducing ) (M for each method M. CONCLUSION This paper proposes an approach to measure the cohesion of a package based on dependence analysis. In this method, we discussed the tightness of a package from the three facets: inter-object , subprogram-object and inter-subprogram. These three facets can be used to measure the package cohesion independently and can also be integrated as a whole. Our approach overcomes the limitations of previous class cohesion measures, which consider only one or two of the three facets. Thus, our measure is more consistent with the intuition. In the future work, we will verify and improve our measure by experiment analysis When measuring package cohesion, the following should be paid attentions. (1) In the hierarchies of types, the primitive subprograms of super type might access the objects of the derived type by dispatching. Therefore, when measuring the cohesion of PG2, it is hard to determine whether the accession of derived typed is considered or not. (2) We can determine polymorphic calls in an application system. However it is impossible for a package, which can be reused in many systems. (3) How to deal with some special subprograms, such as access subprograms, since such subprograms can access some special objects in the package. (4) How to apply the domain knowledge to cohesion measure. In all, if a package can be applied to many applications, the cohesion is mainly about itself without considering the application environments. Otherwise, it is the cohesion in the special environments. 66 ACKNOWLEDGMENTS This work was supported in part by the National Natural Science Foundation of China (NSFC) (60073012), National Grand Fundamental Research 973 Program of China (G1999032701), and National Research Foundation for the Doctoral Program of Higher Education of China (20020286004). REFERENCES [1] Allen, E.B., Khoshgoftaar, T.M. Measuring Coupling and Cohesion: An Information-Theory Approach. in Proceedings of the Sixth International Software Metrics Symposium. Florida USA, IEEE CS Press, 1999, 119-127. [2] Bansiya, J.L., et al. A Class Cohesion Metric for Object-oriented Designs. Journal of Object-oriented Programming, 1999, 11(8): 47-52. [3] Briand, L.C., Morasca, S., Basili, V.R. Property-Based Software Engineering Measurement. IEEE Trans. Software Engineering, Jan. 1996, 22(1): 68-85. [4] Briand, L.C., Daly, J., Wuest, J. A Unified Framework for Cohesion Measurement in Object-Oriented Systems. Empirical Software Engineering, 1998, 3(1): 65-117. [5] Briand, L.C., Morasca, S., Basili, V.R. Defining and Validating Measures for Object-Based High-Level Design. IEEE Trans. Software Engineering, 1999, 25(5): 722-743. [6] Chae, H.S., Kwon, Y.R., Bae, D.H. A Cohesion Measure for Object-Oriented Classes. Software Practice & Experience, 2000, 30(12): 1405-1431. [7] Chae, H.S., Kwon, Y.R. A Cohesion Measure for Classes in Object-Oriented Systems. in Proceedings of the Fifth International Software Metrics Symposium. Bethesda, MD USA, 1998, IEEE CS Press, 158-166. [8] Chen, Z., Xu, B., Yang, H. Slicing Tagged Objects in Ada 95. in Proceedings of AdaEurope'2001, LNCS 2043: 100-112 . [9] Chen, Z., Xu, B., Yang, H., Zhao, J. Static Dependency Analysis for Concurrent Ada 95 Programs. in Proceedings of AdaEurope 2002, LNCS 2361, 219-230. [10] Chen, Z., Xu, B. Zhou, Y., Zhao, J., Yang, H. A Novel Approach to Measuring Class Cohesion Based on Dependence Analysis. in Proceedings of ICSM 2002, IEEE CS Press, 377-383 [11] Chidamber, S.R., Kemerer, C.F. A Metrics Suite for Object-Oriented Design. IEEE Trans. Software Engineering, 1994, 20(6): 476-493. [12] Hitz, M., Montazeri, B. Measuring Coupling and Cohesion in Object-Oriented Systems. in Proceedings of International Symposium on Applied Corporate Computing, Monterrey, Mexico, October 1995: 25-27. [13] Xu, B., Zhou, Y. Comments on A cohesion measure for object-oriented classes. Software Practice & Experience, 2001, 31(14): 1381-1388. [14] Zhou, Y., Guan, Y., Xu, B. On Revising Chae's Cohesion Measure for Classes. J. Software. 2001, 12(Suppl.): 295-300 (in Chinese) 67

Object-Oriented;Ada95;cohesion;dependence;Measurement;OO programming;measure;Cohesion

136

Measurement of e-Government Impact: Existing practices and shortcomings

Public administrations of all over the world invest an enormous amount of resources in e-government. How the success of e-government can be measured is often not clear. E-government involves many aspects of public administration ranging from introducing new technology to business process (re-)engineering. The measurement of the effectiveness of e-government is a complicated endeavor. In this paper current practices of e-government measurement are evaluated. A number of limitations of current measurement instruments are identified. Measurement focuses predominantly on the front (primarily counting the number of services offered) and not on the back-office processes. Interpretation of measures is difficult as all existing measurement instruments lack a framework depicting the relationships between the indicators and the use of resources. The different measures may fit the aim of the owners of the e-governmental services, however, due to conflicting aims and priorities little agreement exists on a uniform set of measures, needed for comparison of e-government development. Traditional methods of measuring e-government impact and resource usage fall short of the richness of data required for the effective evaluation of e-government strategies.

INTRODUCTION Public institutions as well as business organizations use the Internet to deliver a wide range of information and services at an increasing level of sophistication [24]. However, Web sites and related business processes and information systems are so complex that it is difficult for governments to determine adequate measures for evaluating the efficiency and effectiveness of the spending of their public money. Moreover only measuring the front of public websites is a too narrow view on e-government. E-government involves the collaboration and communication between stakeholders and integration of cross-agency business processes. An important aim of having a well-funded theory on measuring e-government is to allow comparison or benchmarking. By examining the results of such benchmarks we might be able to distinct good from bad practices and to give directives to designers of e-governmental services. Moreover is should help to identify how effective public money is spend. Thus evaluating the relationship between results and resources used. Comparison can have different goals. Selecting between options is just one of them. Principally we can distinct three types of comparison: 1) comparison of alternatives (e.g. to make a choice between alternative solutions), 2) vertical comparison over time (e.g. in order to measure improvement of versions) and 3) horizontal comparison (e.g. benchmarking different solutions). Whatever type of comparison we choose, we can only compare if we have a set of preferred outcomes. Many measurement instruments currently applied are not described in such way that the preferences underlying the instrument have been made explicitly. In this research we describe different measurement instruments used and we will discuss their strengths and weaknesses. The goal of this paper is to evaluate how current measurement instruments the impact of e-government. The combined research methodology of literature research and case study was chosen to answer the goal of this research. Case study research can be characterized as qualitative and observatory, using predefined research questions [36]. Case studies were used to examine existing measurement instruments in the Netherlands. The e-government monitor of Accenture and a European Union measurement instrument was also evaluated as thes instruments are used by Dutch agencies to benchmark their situation. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. ICEC'04, Sixth International Conference on Electronic Commerce Edited by: Marijn Janssen, Henk G. Sol, and Ren W. Wagenaar Copyright 2004 ACM 1-58113-930-6/04/10...$5.00 481 THE NATURE OF E-GOVERNMENT Before analyzing instruments developed to measure e-government development, it is necessary to understand the nature of e-government . The organization of public administration involves many, heterogeneous types of business processes. When law or regulations are changed these processes and their supportive systems have to be adapted. Within one policy layer the process of adaptation starts with legislation drafting (adapting the law or regulations) followed by a chain of processes varying from translating these law texts into specifications, design of processes and supporting systems, development of these processes and systems and finally implementation and use (for a recent `legal engineering' approach see [34]). A complicating factor is that more than one governmental layer exists and often interaction between these layers occurs. Especially the need to adapt legislation from the European Union is a more then ever dominant factor that even further complicates this process is. In Figure 1 the fragmented nature of the government is shown. Legislation and service provisioning efforts are distributed over the European, State, Region and local level. The situation is even more complicated as within each level many agencies of various types exists. At a local level municipalities, water boards, chambers of commerce, local taxes and many other public and public-private agencies exists. As such many agencies influence the impact of e-government and measurement should include them all. Europe Europe Europe Europe Europe state Europe region Europe Europe local int e r oper ata b ility int e roper a t ability int e r opera t a bility policy and standardization navigation, service provisioning, helpdesk and appeal Aggregation of local initaitves, navigagtion, geogprahical relevance standardization, policies, culture, enforcement businesses demand ? ? ? ? Figure 1: Fragmented nature of public administration Two main types of interactions influence this fragmented landscape, the policy-making, implementation, execution and enforcement of new legislation and the business and citizens searching for information and services. We will denote the creation, implementation, execution, enforcement and maintenance of laws as production cycle in this paper. Governments are looking for ways to increase their efficiency, effectiveness, decrease the administrative burden and reduce the lead times for adopting new legislations. The consequences of new laws at production phase (drafting) are only roughly known. Only after implementing the new regulations in the processes and supporting systems the full meaning of applying these regulations becomes clear. Certainly when the interpretations and translation into practical applications take place at local government level it often will not be possible to timely inform the businesses or citizens, who are affected by the new law, pro-actively, as no information about concrete effects on their cases is available. A complicating factor is that many times a administrative process is supported by heterogeneous information systems and many different parties can be involved in adapting those systems. Most of the companies' ERP software components need to be updated quite frequently to be in, and keep being in accordance with small changes in administrative legislation. The same holds for Human Resources software, Time reporting software, Tax reporting software and, more indirectly, Logistics software. All have to be updated due to changes in legislation. It does not require extensive explanation to stress the need for smart public-private networks from a production chain perspective. Nowadays businesses expect that the governments reduce the administrative burden for businesses. Governments can achieve this goal by creating a smart, service oriented, public administration. To be able to provide these integrated services access to different legal sources or better to the formal models that could be derived from those models is needed (see e.g. [7]). Standardization initiatives like the Metalex standard for describing legal sources (see www.metalex.nl ) and the Power method for modeling normative systems (see www.belastingdienst.nl/epower ) are essential first steps towards this. They provide a basis for interoperable and contextual better understandable and accessible legal sources that could easier be connected to the context of business activities. From the demand perspective, citizens and businesses find it very hard to access relevant legislation, local procedures and rules, policy documents etc. Governmental bodies are engaged in a flurry of policy and law making activities. Not only is this a complex myriad of legal issues, but the information is produced at different levels of public administration, including local, regional, national and European union. A commonly accepted requirement however is that online state legislative information is equally accessible to all (Fage & Fagan, 2004) and of course in the first place it should be accessible. Many governments currently are searching for ways to make their information accessible and retrievable. This involves issues regarding terminology, explaining the type of legislative document, understandable and easy-to-use search interfaces and accessing the official status of online documents. A central question for researchers working in the field of e-government is how to measure the progress made in complying to the requirements mentioned before. In this paper we examine some examples of measurement instruments that were developed to measure progress in e-Government. But before describing these instruments we will discuss some literature on measuring eGovernment first. LITERATURE REVIEW There is a diversity of literature focusing on measurements. Stage models are often used to positioning and evaluate the current status of e-government development. Services literature focusing 482 on the measurement to of perceived quality, satisfaction of complex, multi-service organizations. Last there is the research focusing on developing suitable `yardstick', performance indicators acting as surrogates for measuring the performance. 3.1 Stage models Many researchers on e-business address the stages of Web site development. Green [17] suggests three stages: attracting, transforming, and utilization of media technology. Fink et al. (2001) focuses on attracting, enhancing and retaining client relationships using the Web site applications. Moon [27]proposes a five stage model. Layne and Lee [23] propose four stages of a growth model towards e-government. In stage one, governments create a `state website' by establishing a separate Internet department. There is no integration between the processes in the front and back offices. The web sites are focused on creating a web-presence and providing information. Transaction possibilities for customers are restricted to printing online forms and sending them by traditional mail. At stage two, transaction, there is two-way communication. Customers transact with government on-line by filling out forms and government responds by providing confirmations, receipts, etc. The number of transactions is relatively small and the organization becomes confronted with translating information from and back to the front office and the other way around. Often a working database is established in the front office for supporting immediate transactions. The data in this database is periodically derived from and exported to the various databases in the back office. At stage three, vertical integration, the focus is moving towards transformation of government services, rather than automating and digitizing existing processes. Most information systems are localized and fragmented. Government organizations often maintain separate databases that are not connected to other governmental agencies at the same level or with similar agencies at the local or federal level. A natural progression according to Layne and Lee is the integration of scattered information systems at different levels (vertical) of government services. Layne and Lee (2001) expect that vertical integration within the similar functional walls but across different levels of government will happen first, because the gap between levels of government is much less than the difference between different functions. Information systems are connected to each other or, at least, can communicate with each other. The problem with the description of governmental systems is that they don't make a distinction between the (legal and administrative `knowledge' contained in those systems and the data (of citizens) to which that knowledge is applied (e.g. to derive if someone is entitled to receive a subsidy). Especially the sharing of data is limited and for very good reasons too! Both the desire to guarantee a certain level of privacy and the vulnerability for misuse of data have been reasons for the legislator to limit storage, reusing and sharing data between different agencies (not to speak of passing data of citizens from the government to private institutions). The challenge consequently is how to realize the full potential of information technology, from the customer's perspective, since this can only be achieved by horizontally integrating government services across different functional walls (`silos') in Layne and Lee's stage four, horizontal integration. The question is how to achieve this without the need for having databases across different functional areas to communicate with each. The situation that information obtained by one agency can be used for CRM purposes by other agencies by sharing information is for reasons mentioned earlier, undesirable. The knowledge contained in these information systems however can be shared and reused thus allowing better integrated government services. 3.2 Service literature The concepts of perceived quality and satisfaction are two of the fundamental pillars of recent evaluation studies. Bign et al. [5] conclude that in most cases the fundamental unit of analysis has been an isolated service, and the fact that several services may be offered by an individual organization has not been taken into account. Indeed multi-service organizations, where the customer has access to several services, have not been so intensively dealt with. The problems facing these organizations in terms of measurement and management of perceived quality and of satisfaction are more complex than in those organizations where only one service is offered, but less complex then situations where a service has to be assembled from many other service suppliers. When measuring the quality of such integrated service it is necessary to take into consideration not only the perceived quality of each of the elementary services, but also the perceived overall quality of the constituting ones. Bign et al. [5] found that the scale used to determine the perceived quality of the core services of hospitals, and universities, is composed of five dimensions, as proposed by Parasuraman et al. [28]: tangibility, reliability, responsiveness, confidence and empathy. 3.3 Performance indicators Performance indicators serve as surrogates to deduce the quality of the e-government performance [20]. Lee [24] provides measurements based on development states and a modification of Simeon's [32]components. (1) The following five items determine the affect (Attracting) of a homepage on the Web site: 1. Design of logo and tagline (quick summary of what a Web site is all about). 2. Graphics (e.g. layout, color and figures of a homepage). 3. Institution's self-advertising (e.g. banner, button, and interstitials). 4. Services for attracting (e.g. quiz, lottery, e-card, maps, weather, channels, download service).C 5. Contents for attracting (e.g. entertainments, culture, tourism, game, kids, health, gallery). (2) Informing consists of nine items developed by modifying Simeon's (1999) components: local links, contents for publicity, contents for learning, reports, descriptions on the institution, descriptions on online administrative services, projects, contact information and counseling. (3) Community consists of ten items: online forum, events, partner links (or ads), e-Magazine (or newsletter or Webcast), message boards, users' participation (e.g. articles, photos, personal links), focus of news, vision (or values), domain identity and community services (or online support for community meeting or networking). A good example of the latter is the ``Citizen discussion room'' of Ulsan City (eulsan.go.kr). 483 (4) Delivering as a variable is determined by the presence or absence of features like: search engine, mailing list, framework, multimedia, password system, FAQ, chat, downloadable publications and update indication. (5) Innovation. Public institutions have to utilize the Internet for actual service innovation. Hence, two variables indicating innovation results are selected: transformation level of existing services and frequency of new innovative services. These are each rated on a five-point scale: ``(1) never; (2) only descriptions; (3) online request; (4) partial; (5) full processing'' for the first item and'' (1) never to (5) many new systems'', for the second item. Such quantification is possible because the introduction of new innovative systems on public sector Web sites is growing, for example, Docket Access, View Property Assessments, and Request for Proposals of Philadelphia (phila.gov) and Citizen Assessment Systems, Citizen Satisfaction Monitor, Online Procedures Enhancement (OPEN) System of Seoul (metro.seoul.kr). These measures focus mainly on components visible to users and do not take into account back-office components like integration. Van der Merwe and Bekker [26] classify website evaluation criteria in 5 groups as shown in table1. Many of their criteria seem to be inspired by their e-Commerce orientation but many of the criteria will be applicable to e-Government as well. Table 1: Web site evaluation criteria groups according to Merwe and Bekker [26] Phase Criteria group This criteria group evaluates/measures Interface Graphic design principles The effective use of color, text, backgrounds, and other general graphic design principles Graphics and multimedia The effectiveness of the graphics and multimedia used on the site Style and text Whether or not the text is concise and relevant, and the style good Flexibility and compatibility The degree to which the interface is designed to handle exceptions, for example, text-only versions of pages Navigation Logical structure The organization and menu system of the site Ease of use The ease of navigation to find the pages that the user is looking for Search engine The search engine's ability to find the correct pages easily and provide clear descriptions of the search results Navigational necessities Other important aspects of navigation like the absence of broken links and ``under-construction'' pages Content Product/service-related information Whether or not the products/services are described precisely and thoroughly Agency and contact Information Whether or not it is easy to find information on the company, its employees and its principals Information quality The currency and relevance of the content on the site Interactivity How much input the user has on the content displayed on the site Reliability Stored customer profile The registering process and how the company uses the stored customer profile Order process The effectiveness and ease of use of the online order process After-order to order receipt The company's actions from order placement until the order is delivered Customer service How the company communicates and helps its online customers Technical Speed Different aspects of the loading speed of the site Security Security systems and the ways used by the company to protect customers' privacy on the site Software and database Flexibility in terms of different software used. Also looks at the data software and data communication systems used on the site System design The correct functioning of the site and how well it integrates with internal and external systems MEASUREMENTS INSTRUMENTS FOUND IN PRACTICE Hazlett and Hill discuss [19] discuss the current level of government measurement. Huang and Chao (2001) note that while the development and management of web sites are becoming essential elements of public sector management, little is known about their effectiveness. Indeed, Kaylor et al. (2001) note that research into the effectiveness of e-Government efforts tends to concentrate on content analysis or measures of usage. These may not be wholly appropriate metrics with which to gauge success. Aspects of service relevant in this context may, according to Voss (2000) include: consumer perceptions of security and levels of trust; response times (bearing in mind that Internet consumers may well be accustomed to quick responses); navigability of the Web site; download time; fulfillment of service promised; timely updating of information; site effectiveness and functionality. Reinforcing a point made above, Voss (2000) takes the view that e-service channels should be regarded as providing 484 support for front-line service delivery and not replacements for front-line service. However, such channels do enable change in the nature of front-line service delivery and of human involvement in service. 4.1 OVERHEID.NL The Dutch Government has recently published "Overheid.nl Monitor 2003", its fifth annual e-government progress report. While highlighting a number of encouraging developments, the report concludes that much remains to be done in areas such as user-friendliness, transactional services and e-democracy. Overheid.nl focused on all government agencies, and mentions the following agencies explicitly. Municipalities Ministries Provinces Water boards A screenshot of the online monitor of the measurement of municipalities is shown in the figure below. Figure 2: Screenshot of Overheid.nl "Overheid.nl Monitor 2003: developments in electronic government" is based on a periodical large-scale survey of government websites, which was carried out in October 2003 by Advies Overheid.nl on behalf of the Dutch Ministry of the Interior and Kingdom Relations. The survey assessed 1,124 government websites according to five criteria: user-friendliness, general information, government information, government services, and scope for participation (interactive policy making). The website user friendliness measurement etc, is very thorough in this survey. The e-service measurement is less well defined. The services investigated for the survey are listed clearly for several layers of government but they seem to be limited to the so called `Dutch service product catalogue', set of typical municipal products and services. Figure 3: Sample listing of services measured and ways of accessing them investigated Additionally, researchers measured the e-mail response time of government websites and assessed user satisfaction via a survey of 3,000 users. The report states that, although e-government services are developing on schedule and are becoming more sophisticated, there is still much room for improvement. On the positive side, the report finds that: E-government is developing on schedule. The 2003 target of providing 35% of government services electronically was widely achieved, with 39% of services for citizens and 44% of services for businesses e-enabled by October 2003. However, the report also identifies a number of shortcomings and areas where improvement is needed: Practically no full electronic transactions are available. In this respect, the report considers that development of such services will depend both on future solutions for user identification and authentication and on back-office re-engineering. Although the use of e-services is growing, the development of e-government is still mainly supply-driven and the penetration of government websites remains unknown. "Only if we assess the penetration of government websites and the level of their use can we take a truly demand-driven approach", the report says. The items related to municipalities are connected to the functionalities within an implemented product and services catalogue. D1 Is a product catalogue or another form of systematically offered services been used? D2 -if so, does at least contain 150 products? D3 -if so, does is at least contain 50 forms? i.e.: can one request for or comment on at least 50 product by using a form that is contained in the product catalogue which can be filled in, printed and send in by the users ...? D4 -if so, can these be accessed per life event? D5 -if so, can these be accessed per theme? D6 -if so, can these be accessed by using a lexicon (a-z)? D7 -if so, does it contain a specific search function? (fill in the search term) Besides this, four commonly municipal products are mentioned that can be supplied more or less in a digital form. (choices: no info; info; down loadable form; up loadable form; transaction) D8a Request for building permission D8b Request for Cutting trees permission D8c Request for extract from GBA D8d Report of change of address / removal (no transaction possible) 485 Users' satisfaction with e-government services is still significantly lower than with services delivered through traditional channels. E-democracy tools and citizen engagement through electronic means remain embryonic. According to the report, this is due not only to a lack of demand but also to a poorly organized supply side, with inadequate moderation, unappealing consultation subjects and missing follow-up. In addition to identifying progress accomplished and remaining issues, the report makes a number of recommendations that should help reach the objectives of the Andere Overheid ("Other Government") action programme, presented in December 2003. Such recommendations include the following: E-government services must become more user-friendly and easier to find. Metadata standards should be defined to make them easier to find through search engines. FAQs and lists of the most searched terms and documents should also be made more widely available. E-government services must be permanently improved: even once all government websites are fully functional, government should still constantly aim to improve e-government and consult target groups about new services they might require, says the report. E-government must be further developed through service integration across government bodies, which is currently still in its infancy. According to the report, the Dutch supply-driven approach has so far sought solutions within the limits and administrative competencies of single bodies. Emphasis must be shifted from the breadth of services to their depth. Rather than aiming to run every electronic product and service conceivable, government bodies should aim to integrate services as deeply as possible, especially those in frequent and popular demand, the report says. This implies developing seamless links from the front to the back office and fostering a more customer-minded culture. From the perspective of the policymakers we may conclude that the benchmark takes into account individual agencies and their websites, number of services and to a certain degree also service levels, but the aim is to integrate horizontally, something which is not measured by www.overheid.nl . 4.2 WEBDAM.NL In the year 2000 the ministry of interior decided that all municipalities should be online by 2002 and 25% of the services provisioning should be support by websites. In order to help municipalities to achieve this, the Webdam project was started in March 2000 aiming at stimulating municipalities to develop websites. These websites should deliver better and improved services provisioning over the Internet as citizens expected. One of the activities in the Webdam project has been the development of a website that allowed municipalities to share knowledge and make better use of the Internet. To further stimulate municipalities Webdam started a Top 50 for municipalities' websites, using criteria such as design, content, service level and communication. Assessment and ranking of each municipality is performed by representatives coming from three groups; civil servants, citizens and experts. Webdam uses a panel of experts to judge the public agencies' web pages. The stakeholders include the following groups: 1. Webdam employees (experts) 2. Public servants municipality under study 3. Citizens These stakeholders judges the web pages based on fie main groups 1. Layout, 2. Content, 3. Communication, 4. services and, 5. plus/minus remarks Each group has a minimum and maximum score. The total is aggregated to determine a ranking. Who determines score? Figure 4: Screenshot of webdam Webdam focuses exclusively on the front-office, the aspects directly visibility to the citizens using the web pages. No connection to the size of the municipality, the number of citizens and other expect influencing the available resources of the municipality. is made 4.3 Accenture e-gov monitor The yearly research conduct by Accenture [1][2][3] has a profound influence on governments. An increase of decrease in ranking of this report results in discussions about the future of e-government . Accenture researchers in each of the 23 selected countries described the typical services that national government should provide using the Internet in order to fulfill the obvious needs of citizens and businesses. They accessed and assessed the websites of national government agencies to determine the quality and maturity of services, and the level at which business can be conducted electronically with government. In total, 169 national government services across nine major service sectors were investigated by Accenture during a two weeks lasting study (!) in 2002 using the web in 23 countries. The nine service sectors researched were Human Services, Justice & Public Safety, Revenue, Defence, Education, Transport & Motor Vehicles, Regulation & Democracy, Procurement and Postal. The main "indicator" of the eGovernment level chosen by Accenture 486 is what they call: service maturity. Service maturity is an indication for the level to which a government has developed an online presence. It takes into account the numbers of services for which national governments are responsible that are available online (Service Maturity Breadth), and the level of completeness with which each service is offered (Service Maturity Depth). Service Maturity Overall is the product of Service Maturity Breadth and Service Maturity Depth. Service maturity is decomposed into the following aspects: Publish - Passive/Passive Relationship. The user does not communicate electronically with the government agency and the agency does not communicate (other than through what is published on the website) with the user. Interact - Active/Passive Interaction. The user must be able to communicate electronically with the government agency, but the agency does not necessarily communicate with the user. Transact - Active/Active Interaction. The user must be able to communicate electronically with the government agency, and the agency must be able to respond electronically to the user. the degree to which the services are organized around the citizen, as opposed to around internal government structures. In 2004 Accenture again investigated 12 service sectors and 206 services in yet again two weeks. They were: agriculture; defence; e-Democracy; education; human services; immigration, justice and security; postal; procurement; regulation; participation; revenue and customs; and transport. Little is said by Accenture about the metrics involved. They have performed the survey for five years now and the perspective chosen is that of a citizen accessing a government service using on line means. For this article it is interesting to note the final remarks in the 2004 report: governments are at the end of what can be achieved with traditional methods; they are developing strategies to cope with horizontal integration between agencies. 4.4 Regional innovation scorecard (ris) One of the ambitions of the EU is to become the most competitive and dynamic knowledge-based economy of the world (Lissabon agenda). To measure this, the European Regional Innovation Scorecard (RIS), a scorecard used for monitoring and comparing the innovation in regions, has been developed [13]. The scorecard is seen as acknowledged instrument to compare regions in their ability to foster economic growth. The Largest province of The Netherlands, The Province of South Holland, explicitly states on their website: "The province of Noord-Brabant ranks third on the European regional innovation scoreboard. Zuid-Holland will have to make a considerable effort in the coming years if it is to reach the top-20". They also state that "the scoreboard is regarded as extremely relevant because it is generally accepted as the a leading European benchmark for innovation dynamics" [30]. Surprisingly the same regional authority does not pay any attention to the contribution of their own eGovernment services to that level of innovation and economic growth. There is no mentioning of eGovernment or government services or anything close to it in the whole policy documents related to innovation at this Province. This becomes less surprising when the indicators of the RIS and the EIS are viewed more closely. The RIS uses the following indicators. (1) population with tertiary education (2) lifelong learning, (3) employment in medium/high-tech manufacturing, (4) employment in high-tech services, (5) public R&D expenditures, (6) business R&D expenditures, (7) EPO high-tech patent applications, (8) all EPO patent applications, and (9) five indicators using unpublished CIS-2 data (10) the share of innovative enterprises in manufacturing and services innovation expenditures as a percentage of turnover in both (11) manufacturing and (12) services, (13) the share of sales of new-to-the-firm products in manufacturing. These indicators are based on Eurostat exercises [13]. From this analysis it becomes obvious that this Province considers innovation as a development process "outside" the government and its own performance. The basic assumption made by the Dutch Provinces is that governments can stimulate innovation in the economy without being part of the regional economy. The main political driver for efficient eGovernment is economic growth and jobs and the main driver for economic growth is considered to be innovation. The metrics of the benchmarks do no coincide. EVALUATION The evaluation instruments described before are just examples of the overwhelming number measurement instruments currently used. Table 2 summarizes the described instruments. Although we focused on a limited number of instruments these instruments are very likely representative for the other measurement instruments. The following observations can be made about the measurement instruments: Most instruments focus on the performance of a single agency; Measurement focus on front, which is directly visible, and not on the business process and information system in the back. This is surprisingly as these aspects influence the performance to a large extend; Short-term focus, not many indicators are focused on measuring the long-term performance; Interpretation of measures is difficult as all existing measurement instruments lack a framework depicting the relationships between the indicators and the use of resources. Different stakeholders may come to different interpretations of the status of e-government. From a theoretical point of view we conclude after examining many other existing instruments that these instruments lack a 487 clear connection to any theoretical framework on e-Government and a well-described set of preferences that can be used for comparison. Even if we would consider that these measurements instruments were developed independent of each other it is astonishing that these instruments show that little overlap both in features as in measurement method. Table 2: Summary of measurement instruments studied Governmental performance is dependent on a complex of interlinked processes and dependencies between these processes, the stakeholders involved including civil servants and their departments. The legal and political context which is very dominant in a governmental setting furthermore increases complexity. Sometimes obvious improvements in a service provision chain may be blocked because data may not be shared due to data protection regulations. The system of checks an balances that is fundamental to governments' functioning and essential for maintaining citizens' trust in the government can be troublesome if we want to redesign inefficient processes. A combination of factors such as the volume of regulations and the lack of understanding of their original aims, the lack of formal process models that could help to get insights in the dependencies between processes and explain how the legal requirements are translated into process requirements and the lack of formally described legal models, don't really help if we want to explicitly formulate the criteria that determine e-Government success. These criteria determine e-Government success (or failure) are exactly the ones that should be the ones in our measurement instruments. But even if we would have had a better theory on performance of e-Government processes and we would have had well funded measurement instruments, interpretation of the outcomes of applying those instruments would be problematic, especially within the political context within these instruments are generally used. Bruin [10] showed that when the distance between the interpreters and providers of information is bigger, it is more difficult to interpret information. Politicians do not always steer on rational grounds but suppose they would then their control system (or policy making process see figure 5) would include a comparison and control function. We stated before that comparison is based upon a set of preferences. Public services can consequently be evaluated using competing norms and values [18]. A court for example might be asked to deal with cases as efficiently as possible, to maximize the number of cases dealt with, within a certain time period on the one hand, while on the other hand the sentence should be carefully made, funded with the right arguments and understandable. Performance measurement instruments that lack an explicit set of preferences (or norms for comparison) might give a wrong view on reality if looked at with other preferences in mind. CONCLUSIONS AND FURTHER RESEARCH We investigate the current e-government measurement practice in the Netherlands and investigated some theoretical work in this field. Our analyzes shows a messy picture on the measurement of e-government. Many measurement instruments take a too simplistic view and focus on measuring what is easy to measure. Many of the instruments focus on measuring the visible front of e-government and ignore the performance of the cross-agency business processes. None of the instruments focus on measuring multi-service organizations. The instruments focus on one (type of) agency and do not provide an overall picture. Interpretation of measures is difficult as all existing measurement instruments lack a framework depicting the relationships between the indicators and the use of resources. The different measures may fit the aim of the owners of the e-governmental services, however, due to conflicting aims and priorities little agreement exists on a uniform set of measures, needed for comparison of e-government development. Different stakeholders may come to different interpretations of the status of e-government. As such the existing instruments provide a picture of the status of e-government that may not useful as surrogates for deducing the e-government performance Traditional methods of measuring e-government impact and resource usage fall short of the richness of data required for the effective evaluation of e-government strategies. A good theoretical framework for measuring the impact of e-government and the use of resources is still lacking. Due to this fact and the many reports that are produced on e-Government developments, based on different measurement instruments that used different criteria, we can hardly learn from existing practices. It would be Measurement instrument Focus Update frequency Source data Characteristics of the method Overheid.nl All public agency websites Yearly Experts Ranking based on web site features Webdam Municipality websites Monthly (continuous) Expert panel consisting of 3 types of representatives: 1) Civil servants, 2) citizens and 3) experts Ranking based on web site features Accenture Comparison of countries Yearly Accenture researchers based on judgment of a selected services Ranking based inventory of services Regional innovation scorecard European regions Eurostat Ranking based on economic quantitative indicators 488 beneficial for both citizens as for governments if such a theoretical framework would be developed and a more or less standardized measurement instrument could become available. This would allow governments and designers to compare different e-government approaches and learn from them and learning from our experiences certainly fits within the ambition of the European Union to become the most competitive and dynamic knowledge-based economy of the world. REFERENCES [1] Accenture (2001). Governments Closing Gap Between Political Rhetoric and eGovernment Reality, http://www.accenture.com/xdoc/en/industries/government/20 01FullReport.pdf . [2] Accenture (2002). eGovernment Leadership -Realizing the Vision, http://www.accenture.com/xd/xd.asp?it=enWeb&xd=industri es/government/gove_welcome.xml [3] Accenture (2003). eGovernment Leadership: Engaging the Customer, http://www.accenture.com/xd/xd.asp?it=enweb&xd=industri es/government/gove_capa_egov.xml [4] Armour, F.J. Kaisler, S.H. and Liu, S.Y. (1999). A big-picture look at Enterprise Architectures, IEEE IT Professional, 1(1): 35-42. [5] Bign, E., Moliner, M.A., and Snchez, J. (2003) Perceived Quality and satisfaction in multiservice organizations. The case of Spanish public services. Journal of Services Marketing, 17(4), pp. 420-442. [6] Boer, A. Engers, T. van and R. Winkels (2003). Using Ontologies for Comparing and Harmonizing Legislation. In Proceedings of the International Conference on Artificial Intelligence and Law (ICAIL), Edinburgh (UK), ACM Press. [7] Alexander Boer, Radboud Winkels, Rinke Hoekstra, and Tom M. van Engers. Knowledge Management for Legislative Drafting in an International Setting. In D. Bourcier, editor, Legal Knowledge and Information Systems. Jurix 2003: The Sixteenth Annual Conference., pages 91-100 , Amsterdam, 2003. IOS Press. [8] Bons R., Ronald M.Lee and Tan, Yua-Hua, (1999). A Formal Specification of Automated Auditing of Trustworthy Trade Procedures for Open Electronic Commerce. Hawaii International Conference on System Sciences (HICCS). [9] Buckland and F. Gey (1994). The relationship between recall and precision. Journal of the American Society for Information Science, 45(1):12-19. [10] Bruin, H. de (2002). Performance measurement in the public sector: strategies to cope with the risks of performance measurement. The International Journal of Public Sector Management, vol. 15, no. 7, pp. 578-594. [11] Checkland, P. (1981). Systems Thinking, Systems Practice. Wiley, Chichester. [12] Coase, R. (1937). The Nature of the Firm. Economia, 4: 386-405 . [13] European Commissions (2002). 2003 European Innovation Scoreboard: European Trend Chart on Innovation. Innovation/SMEs Programme. [14] European Commission (2004). Green paper on Public private partnerships and community law on public contracts and concessions, European Commission, no. 327. [15] Fagan, J.C. & Fagan, B. (2004). An accessibility study of state legislative web sites. Government Information Quarterly, 21: 65-85. [16] Galliers, R.D. (1992). Information Systems Research. Issues, methods and practical guidelines. Alfred Waller, Fawley, England. [17] Green, S.H.B. (1998), Cyberspace winners: how they did it, Business Week, 22 June, pp. 154-60. [18] Groot, H., de and R. Goudriaan (1991). De productiviteit van de overheid: over prestaties, personeel en uitgaven in de publieke sector. Academic Service, Schoonhoven, The Netherlands. [19] Hazlett, S.A. and Hill, F. (2003). E-government: the realities of using IT to transform the public sector. Managing Service Quality, Vol. 13, No. 6, pp. 445-452. [20] Janssen, M.F.W.H.A. (2001). Designing Electronic Intermediaries. Doctoral Dissertation, Delft University of Technology. [21] Janssen, Marijn & Davidse, Anouk (2004). Evaluation of a Performance-based Accountability System. The 4th European Conference on E-government (ECEG), Dublin Castle, Dublin, Ireland, 17-18 June 2004 [22] Jensen, M. and Meckling, W. (1976). Theory of the Firm: Managerial behavior, agency costs, and capital structure, Journal of Financial Economics, 5: 305-360. [23] Layne, KJL & Lee, J. (2001) "Developing fully functional E-government : A four stage model", Government Information Quarterly, Vol 18, No. 2, pp 122-136. [24] Lee, J.K. (2003). A model for monitoring public sector web site strategy. Internet Research. Electronic networking application and policy. Vol. 13, no. 4, pp259-266. [25] Malone, T.W. & Crowston, K. (1994). The Interdisciplinary Study of Coordination. ACM Computing Surveys, vol. 26, no. 2, pp. 87-119. [26] Merwe, R. van der, and Bekker,J. (2003). A framework and methodology for evaluating e-commerce web sites. Internet Research: electronic Networking Applications and Policy. Vol. 13, No.5, pp. 330-341. [27] Moon, M.J. (2002). The Evolution of E-Government Among Municipalities; Rhetoric or reality? Public Administration Review. Vol. 62, no. 4, pp. 424-433. [28] Parasuraman, A., Zeithaml, V.A. and Berry, L.L. (1988). SERVQUAL: a multiple item scale for measuring consumer perceptions of service quality, Journal of Retailing, Vol. 64, pp. 12-40. [29] Peters, Rob and Wilson, Frank (2003). Natural Language access to regional information sources: the PortofRotterdam case: 4th International Workshop on Image Analysis for Multimedia Interactive Services, WIAMIS 2003. 489 [30] Provincial council (2003) Innovatiebrief kenniseconomie Zuid-Holland "Kennismaken met Kenniszaken, http://www.zuid-holland.nl/images/126_107822.pdf , page 14. [31] Rohleder, S.J. et al. (2004). eGovernment Leadership: High Performance, Maximum Value. Fifth Annual Accenture eGovernment Study. Accenture Government Executive Studies, http://www.accenture.com/xdoc/en/industries/government/go ve_egov_value.pdf [32] Simeon, R. (1999), ``Evaluating domestic and international Web site strategies'', Internet Research, Vol. 9 No. 4, pp. 297-308. [33] Quinn, R.E. and Rohrbaugh, J.W. (1983). A Spatial Model of Effectiveness criteria: Towards a competing values approach to organizational effectiveness. Management Science 29: 363-377. [34] Van Engers, T.M., 2004, Legal Engineering: A Knowledge Engineering Approach To Improving Legal Quality, in eGovernment and eDemocracy: Progress and Challenges, Padget, J., Neira, R., De Len, J.L., Editors, Instituto Politchnico Nacional Centro de Investigacion en Computacin, ISBN 970-36-0152-9, p189-206. [35] Williamson, O.E. (1975). Market and Hierarchies, Analysis and Antitrust Implications. A study in the economics of internal organization. Macmillan, New York. [36] Yin, R.K. (1989). Case Study Research: Design and methods. Sage publications, Newbury Park, California.

E-government;law;benchmark;interoperability;evaluation;measurement;architectures;e-government;business process;public administration

137

Minimizing Average Flow Time on Related Machines

We give the first on-line poly-logarithmic competitve algorithm for minimizing average flow time with preemption on related machines, i.e., when machines can have different speeds. This also yields the first poly-logarithmic polynomial time approximation algorithm for this problem. More specifically, we give an O(log P log S)-competitive algorithm, where P is the ratio of the biggest and the smallest processing time of a job, and S is the ratio of the highest and the smallest speed of a machine. Our algorithm also has the nice property that it is non-migratory. The scheduling algorithm is based on the concept of making jobs wait for a long enough time before scheduling them on slow machines.

INTRODUCTION We consider the problem of scheduling jobs that arrive over time in multiprocessor environments. This is a fundamental scheduling problem and has many applications, e.g., servicing requests in web servers. The goal of a scheduling Work done as part of the "Approximation Algorithms" partner group of MPI-Informatik, Germany. Supported by an IBM faculty development award and a travel grant from the Max Plank Society. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. STOC'06, May 2123, 2006, Seattle, Washington, USA. Copyright 2006 ACM 1-59593-134-1/06/0005 ... $ 5.00. algorithm is to process jobs on the machines so that some measure of performance is optimized. Perhaps the most natural measure is the average flow time of the jobs. Flow time of a job is defined as the difference of its completion time and its release time, i.e., the time it spends in the system. This problem has received considerable attention in the recent past [1, 2, 10]. All of these works make the assumption that the machines are identical, i.e., they have the same speed. But it is very natural to expect that in a heterogenous processing environment different machines will have different processing power, and hence different speeds. In this paper , we consider the problem of scheduling jobs on machines with different speeds, which is also referred to as related machines in the scheduling literature. We allow for jobs to be preempted. Indeed, the problem turns out to be intractable if we do not allow preemption. Kellerer et. al. [9] showed that the problem of minimizing average flow time without preemption has no online algorithm with o(n) competitive ratio even on a single machine. They also showed that it is hard to get a polynomial time O(n 1/2 )-approximation algorithm for this problem. So preemption is a standard, and much needed, assumption when minimizing flow time. In the standard notation of Graham et. al. [7], we consider the problem Q|r j , pmtn| P j F j . We give the first poly-logarithmic competitive algorithm for minimizing average flow time on related machines. More specifically, we give an O(log 2 P log S)-competitive algorithm, where P is the ratio of the biggest and the smallest processing time of a job, and S is the ratio of the highest and the smallest speed of a machine . This is also the first polynomial time poly-logarithmic approximation algorithm for this problem. Despite its similarity to the special case when machines are identical, this problem is more difficult since we also have to worry about the processing times of jobs. Our algorithm is also non-migratory , i.e., it processes a job on only one machine. This is a desirable feature in many applications because moving jobs across machines may have many overheads. Related Work The problem of minimizing average flow time (with preemption) on identical parallel machines has received much attention in the past few years. Leonardi and Raz [10] showed that the Shortest Remaining Processing Time (SRPT) algorithm has a competitive ratio of O(log(min( n m , P ))), where n is the number of jobs, and m is the number of machines. A matching lower bound on the competitive ratio of any on-line (randomized) algorithm for this problem was also shown by the same authors. Awerbuch et. al. [2] gave a non-migratory version of SRPT with competitive ratio of O(log(min(n, P ))). Chekuri et. al.[5] 730 gave a non-migratory algorithm with competitive ratio of O(log(min( n m , P ))). One of the merits of their algorithm was a much simpler analysis of the competitive ratio. Instead of preferring jobs according to their remaining processing times, their algorithm divides jobs into classes when they arrive. A job goes to class k if its processing time is between 2 k-1 and 2 k . The scheduling algorithm now prefers jobs of smaller class irrespective of the remaining processing time. We also use this notion of classes of jobs in our algorithm. Azar and Avrahami[1] gave a non-migratory algorithm with immediate dispatch, i.e., a job is sent to a machine as soon as it arrives. Their algorithm tries to balance the load assigned to each machine for each class of jobs. Their algorithm also has the competitive ratio of O(log(min( n m , P ))). It is also interesting to note that these are also the best known results in the off-line setting of this problem. Kalyanasundaram and Pruhs[8] introduced the resource augmentation model where the algorithm is allowed extra resources when compared with the optimal off-line algorithm. These extra resources can be either extra machines or extra speed. For minimizing average flow time on identical parallel machines, Phillips et. al. [11] showed that we can get an optimal algorithm if we are given twice the speed as compared to the optimal algorithm. In the case of single machine scheduling, Bechheti et. al.[4] showed that we can get O(1/)-competitive algorithms if we are give (1 + ) times more resources. Bansal and Pruhs[3] extended this result to a variety of natural scheduling algorithms and to L p norms of flow times of jobs as well. In case of identical parallel machines, Chekuri et. al.[6] gave simple scheduling algorithms which are O(1/ 3 )-competitive with (1 + ) resource augmentation. Our Techniques A natural algorithm to try here would be SRPT. We feel that it will be very difficult to analyze this algorithm in case of related machines. Further SRPT is migratory. Non-migratory versions of SRPT can be shown to have bad competitive ratios. To illustrate the ideas involved , consider the following example. There is one fast machine, and plenty of slow machines. Suppose many jobs arrive at the same time. If we distribute these jobs to all the available machines, then their processing times will be very high. So at each time step, we need to be selective about which machines we shall use for processing. Ideas based on distributing jobs in proportion to speeds of machines as used by Azar and Avrahami[1] can also be shown to have problems. Our idea of selecting machines is the following. A job is assigned to a machine only if it has waited for time which is proportional to its processing time on this machine. The intuition should be clear if a job is going to a slow machine, then it can afford to wait longer before it is sent to the machine. Hopefully in this waiting period, a faster machine might become free in which case we shall be able to process it in smaller time. We also use the notion of class of jobs as introduced by Chekuri et. al.[5] which allows machines to have a preference ordering over jobs. We feel that this idea of delaying jobs even if a machine is available is new. As mentioned earlier, the first challenge is to bound the processing time of our algorithm. In fact a bulk of our paper is about this. The basic idea used is the following if a job is sent to a very slow machine, then it must have waited long. But then most of this time, our algorithm would have kept the fast machines busy. Since we are keeping fast machines busy, the optimum also can not do much better. But converting this idea into a proof requires several technical steps. The second step is of course to bound the flow time of the jobs. It is easy to see that the total flow time of the jobs in a schedule is same as the sum over all time t of the number of waiting jobs at time t in the schedule. So it would be enough if we show that for any time t, the number of jobs which are waiting in our schedule is close to that in the optimal schedule. Chekuri et. al. [5] argue this in the following manner. Consider a time t. They show that there is a time t &lt; t such that the number of waiting jobs of a certain class k or less in both the optimal and their schedule is about the same (this is not completely accurate, but captures the main idea). Further they show that t is such that all machines are busy processing jobs of class k or less during (t , t). So it follows that the number of waiting jobs of this class or less at time t are about the same in both the schedules. We can not use this idea because we would never be able to keep all machines busy (some machines can be very slow). So we have to define a sequence of time steps like t for each time and make clever use of geometric scaling to show that the flow time is bounded. PRELIMINARIES We consider the on-line problem of minimizing total flow time for related machines. Each job j has a processing requirement of p j and a release date of r j . There are m machines , numbered from 1 to m. The machines can have different speeds, and the processing time of a job j on a machine is p j divided by the speed of the machine. The slowness of machine is the reciprocal of its speed. It will be easier to deal with slowness, and so we shall use slowness instead of speed in the foregoing discussion. Let s i denote the slowness of machine i. So the time taken by job j to process on machine i is p j s i . Assume that the machines have been numbered so that s 1 s 2 s m . We shall assume without loss of generality that processing time, release dates, and slowness are integers. We shall use the term volume of a set of jobs for denoting their processing time on a unit speed machine. Let A be a scheduling algorithm. The completion time of a job j in A is denoted by C A j . The flow time of j in A is defined as F A j = C A j - r j . Our goal is to find an on-line scheduling algorithm which minimizes the total flow time of jobs. Let O denote the optimal off-line scheduling algorithm. We now develop some notations. Let P denote the ratio of the largest to the smallest processing times of the jobs, and S be the ratio of the largest to the smallest slowness of the machines. For ease of notation, we assume that the smallest processing requirement of any job is 1, and the smallest slowness of a machine is 1. Let and be suitable chosen large enough constants. We divide the jobs and the machines into classes. A job j is said to be in class k if p j [ k-1 , k ) and a machine i is said to be in class l if s i [ l-1 , l ). Note that there are O(log P ) classes for jobs and O(log S) classes for machines. Given a schedule A, we say that a job j is active at time t in A if r j t but j has not finished processing by time t in A. SCHEDULING ALGORITHM We now describe the scheduling algorithm. The under-731 lying idea of the algorithm is the following -- if we send a job j to machine i, we make sure that it waits for at least p j s i units of time (which is its processing time on machine i). Intuitively, the extra waiting time can be charged to its processing time. Of course, we still need to make sure that the processing time does not blow up in this process. The algorithm maintains a central pool of jobs. When a new job gets released, it first goes to the central pool and waits there to get assigned to a machine. Let W (t) denote the set of jobs in the central pool at time t. Our algorithm will assign each job to a unique machine -- if a job j gets assigned to a machine i, then j will get processed by machine i only. Let A i (t) be the set of active jobs at time t which have been assigned to machine i. We shall maintain the invariant that A i (t) contains at most one job of each class. So |A i (t)| log P . We say that a job j W (t) of class k is mature for a machine i of class l at time t, if it has waited for at least k l time in the central pool, i.e., t - r j k l . For any time t, we define a total order on the jobs in W (t) as follows j j if either (i) class(j) &lt; class(j ), or (ii) class(j) = class(j ) and r j &gt; r j (in case class(j) = class(j ) and r j = r j we can order them arbitrarily). Now we describe the actual details of the algorithm. Initially , at time 0, A i (0) is empty for each machine. At each time t, the algorithm considers machines in the order 1, 2, . . . , m (recall that the machines have been arranged in the ascending order of their slowness). Let us describe the algorithm when it considers machine i. Let M i (t) be the jobs in W (t) which are mature for machine i at time t. Let j M i (t) be the smallest job according to the total order . If class(j) &lt; class(j ) for all jobs j A i (t), then we assign j to machine i (i.e., we delete j from W (t) and add it to A i (t)). Once the algorithm has considered all the machines at time t, each machine i processes the job of smallest class in A i (t) at time t. This completes the description of our algorithm. It is also easy to see that the machines need to perform the above steps for only a polynomial number of time steps t (i.e., when a job finishes or matures for a class of machines). We remark that both the conditions in the definition of are necessary. The condition (i) is clear because it prefers smaller jobs. Condition (ii) makes sure that we make old jobs wait longer so that they can mature for slower machines. It is easy to construct examples where if we do not obey condition (ii) then slow machines will never get used and so we will incur very high flow time. ANALYSIS We now show that the flow time incurred by our algorithm is within poly-logarithmic factor of that of the optimal algorithm . The outline of the proof is as follows. We first argue that the total processing time incurred by our algorithm is not too large. Once we have shown this, we can charge the waiting time of all the jobs in A i (t) for all machines i and time t to the total processing time. After this, we show that the total waiting time of the jobs in the central pool is also bounded by poly-logarithmic factor times the optimum's flow time. Let A denote our algorithm. For a job j, define the dispatch time d A j of j in A as the time at which it is assigned to a machine. For a job j and class l of machines, let t M (j, l) denote the time at which j matures for machines of class l, i.e., t M (j, l) = r j + k l , where k is the class of job j. Let F A denote the total flow time of our algorithm. For a job j let P A j denote the time it takes to process job j in A (i.e., the processing time of j in A). Similarly, for a set of jobs J define P A J as the total processing time incurred by A on these jobs. Let P A denote the sum P j P A j , i.e., the total processing time incurred by A. Define F O , P O j , P O J and P O similarly. Let m l denote the number of machines of class l. 4.1 Bounding the Processing Time We shall now compare P A with F O . For each value of class k of jobs and class l of machines, let J(k, l) be the jobs of class k which get processed by A on machines of class l. For each value of k and l, we shall bound the processing time incurred by A on J(k, l). So fix a class k of jobs and class l of machines. The idea of the proof is as follows. We shall divide the time line into intervals I 1 = (t 1 b , t 1 e ), I 2 = (t 2 b , t 2 e ), . . . so that each interval I q satisfies the following property A is almost busy in I q processing jobs of class at most k on machines of class l - 1 or less. Further these jobs have release time at least t q b . We shall denote these jobs by H q . Now, if O schedules jobs in J(k, l) on machines of class l or more, we have nothing to prove since O would incur at least as much processing time as we do for these jobs. If O schedules some jobs in J(k, l) on machines of class l - 1 or less during the interval I q , then one of these two cases must happen (i) some jobs in H q need to be processed on machines of class l or more, or (ii) some jobs in H q get processed after time t q e . We shall show that both of these cases are good for us and we can charge the processing times of jobs in J(k, l) to the flow time of jobs in H q in O. Let us formalize these ideas (see Figure 1). The starting points of the intervals I 1 , I 2 , . . . will be in decreasing order, i.e., t 1 b &gt; t 2 b &gt; (so we shall work backwards in time while defining these intervals). Suppose we have defined intervals I 1 , . . . , I q-1 so far. Let J q (k, l) denote the set of jobs in J(k, l) which get released before interval I q-1 begins, i.e., before t q-1 b (J 1 (k, l) is defined as J(k, l)). Now we define I q . Let j q 0 J q (k, l) be the job with the highest dispatch time. Let r q 0 denote the release time of j q 0 . Let k q 0 denote the class of job j q 0 . Let d q 0 denote the dispatch time of j q 0 . The right end-point t q e of I q is defined as d q 0 . Consider the jobs in J(k, l) which get dispatched during (r q 0 , d q 0 ). Let j q 1 be such a job with the earliest release date. Define r q 1 , k q 1 , d q 1 similarly. Let H q 0 (l ), l &lt; l, be the set of jobs of class at most k q 0 which are dispatched to a machine of class l during the time interval (t M (j q 0 , l ), d q 0 ). Note that the phrase "class at most k q 0 " in the previous sentence is redundant because we cannot dispatch a job of class greater than k q 0 during (t M (j q 0 , l ), d q 0 ) on machines of class l (otherwise we should have dispatched j q 0 earlier). Let H q 0 denote l-1 l =1 H q 0 (l ). Define H q 1 (l ), H q 1 similarly. If all jobs in H q 1 H q 0 get released after r q 1 , we stop the process here. Otherwise find a job in H q 1 H q 0 which has the earliest release date, let j q 2 denote this job. As before define r q 2 as the release date of j q 2 , and k q 2 as the class of j q 2 . Again define H q 2 (l ) as the set of jobs (of class at most k q 2 ) which get dispatched on machines of class l during (t M (j q 2 , l ), d q 2 ). Define H q 2 analogously. So now assume we have defined j q 0 , j q 1 , . . . , j q i and H q 0 , H q 1 , . . . , H q i , i 2. If all jobs in H q i are 732 j q 3 j q+1 0 I q j q 2 I q-1 r q 0 d q 0 = t q e d q 2 r q 1 d q 3 r q 2 r q 3 j q 0 l l l - 1 l - 2 1 j q 1 Figure 1: An illustration of the notation used released after r q i , then we stop the process. Otherwise, define j q i+1 as the job in H q i with the earliest release date. Define H q i+1 in a similar manner (see Figure 1). We remark that the first two steps of this process differ from the subsequent steps. This we will see later is requierd to ensure that the intervals do not overlap too much. The following simple observation shows that this process will stop. Claim 4.1. For i 2, k q i &lt; k q i-1 . Proof. Consider the job j q i H q i-1 (l ). A prefers j q i over j q i-1 . If class(j q i ) = class(j q i-1 ), release time of j q i must be at least that of j q i-1 . But this is not the case. Thus, class of j q i must be less than that of j q i-1 . Suppose this process stops after u q steps. Define the beginning of I q , i.e., t q b as r q u q . This completes our description of I q . Let H q denote i H q i . We are now ready to show that interval I q is sufficiently long, and that for a large fraction of this time, all machines of class less than l are processing jobs of class less than k which are released in this interval itself. This would be valuable in later arguing that O could not have sched-uled jobs of J(k, l) on the lower class machines and thereby incurred small processing time. Lemma 4.2. Length of the interval I q is at least k l . Proof. Indeed, job j q 0 must wait for at least k l amount of time before being dispatched to a machine of class l. So t q e - t q s d q 0 - r q 0 k l . Lemma 4.3. H q consists of jobs of class at most k and all jobs in H q are released after t q b . Proof. The first statement is clear from the definition of H q . Let us look at H q i . As argued in proof of Claim 4.1 all jobs of class k q i in H q i are released after r q i . If all jobs of class less than k q i in H q i are released after r q i , then we are done. Otherwise, all such jobs are released after r q i+1 (after r q 2 if i = 0). This completes the proof of the lemma. Lemma 4.4. A machine of class l &lt; l processes jobs of H q for at least |I q | - 6 k l amount of time during I q . Proof. Let us fix a machine i of class l &lt; l. Let us consider the interval (r q p , d q p ). Job j q p matures for i at time t M (j q p , l ). So machine i must be busy during (t M (j q p , l ), d q p ), otherwise we could have dispatched j q p earlier. We now want to argue that machine i is mainly busy during this interval processing jobs from H q . Let j be a job which is processed by i during (t M (j q p , l ), d q p ). We have already noted that j can be of class at most k q p . If j gets dispatched after t M (j q p , l ), then by definition it belongs to H q . If j gets dispatched before t M (j q p , l ), it must belong to A i (t ). But A i (t ) can contain at most one job of each class. So the total processing time taken by jobs of A i (t ) during (t , d p ) is at most l (+ 2 + + k q p ) 3/2 k q p l . So during (r q p , d q p ), i processes jobs from H q except perhaps for a period of length (t M (j q p , l ) - r q p ) + 3/2 k q p l = 5/2 k q p l . Since p (r q p , d p ) covers I q , the amount of time i does not process jobs from H q is at most 5/2 l ( k + k + + 1), which proves the lemma. We would like to charge the processing time for jobs in J(k, l) in our solution to the flow time of jobs in H q in O. But to do this we require that the sets H q be disjoint. We next prove that this is almost true. Lemma 4.5. For any q, I q and I q+2 are disjoint. Hence H q and H q+2 are also disjoint. Proof. Recall that J q+1 (k, l) is the set of jobs in J(k, l) which get released before t q b . However some of these jobs may get dispatched after t q b and hence I q and I q+1 can intersect . Consider some job j J q+1 (k, l) which is dispatched during I q . Now, observe that j q+1 0 is released before t q b (by definition of J q+1 (k, l)) and dispatched after j. So, j gets dispatched in (r q 0 , d q 0 ). This means that release date of j q+1 1 must be before release date of j. But t q+1 b r q+1 1 and so, j is released after t q+1 b . But then j / J q+2 (k, l). So all jobs in J q+2 (k, l) get dispatched before I q begins, which implies that I q and I q+2 are disjoint. Consider an interval I q . Let D q (k, l) denote the jobs in J(k, l) which get released after t q b but dispatched before t q e . It is easy to see that D q (k, l) is a superset of J q (k, l) J q+1 (k, l). So q D q (k, l) covers all of J(k, l). Now we would like to charge the processing time of jobs in D q (k, l) to the flow time incurred by O on the jobs in H q . Before we do this, let us first show that O incurs significant 733 amount of flow time processing the jobs in H q . This is proved in the technical theorem below whose proof we defer to the appendix. Theorem 4.6. Consider a time interval I = (t b , t e ) of length T . Suppose there exists a set of jobs J I such that every job j J I is of class at most k and is released after t b . Further A dispatches all the jobs in J I during I and only on machines of class less than l. Further, each machine i of class l &lt; l satisfies the following condition : machine i processes jobs from J I for at least T - 6 k l amount of time. Assuming T k l , the flow time F O J I incurred by O on the jobs in J I is at least `P A J I . Substituting t b = t q b , t e = t q e , I = I q , T = |I q |, J I = H q in the statement of Theorem 4.6 and using Lemmas 4.2, 4.3, 4.4, we get P A H q O(F O H q ). (1) We are now ready to prove the main theorem of this section . Theorem 4.7. The processing time incurred by A on D q (k, l), namely P A D q (k,l) , is O(F O H q + F O D q (k,l) ). Proof. Let V denote the volume of D q (k, l). By Lemma 4.4, machines of class l do not process jobs from H q for at most 6 k l units of time during I q . This period translates to a job-volume of at most 6 k . If V is sufficiently small then it can be charged to the processing time (or equivalently the flow time in O) of the jobs in H q . Lemma 4.8. If V c k (m 0 + . . . + m l-1 ) where c is a constant, then P A D q (k,l) is O(F O H q ). Proof. The processing time incurred by A on D q (k, l) is at most l V = O( k l+1 (m 0 + . . .+m l-1 )). Now, Lemmas 4.2 and 4.4 imply that machines of class l &lt; l process jobs from H q for at least k l /2 amount of time. So P A H q is at least k l (m 0 + . . . + m l-1 )/2. Using equation (1), we get the result. So from now on we shall assume that V c k+1 (m 0 + . . . + m l-1 ) for a sufficiently large constant c. We deal with easy cases first : Lemma 4.9. In each of the following cases P A D q (k,l) is O(F O D q (k,l) + F O H q ) : (i) At least V /2 volume of D q (k, l) is processed on machines of class at least l by O. (ii) At least V /4 volume of D q (k, l) is finished by O after time t q e k l /2. (iii) At least V /8 volume of H q is processed by O on machines of class l or more. Proof. Note that P A D q (k,l) is at most V l . If (i) happens, O pays at least V/2 l-1 amount of processing time for D q (k, l) and so we are done. Suppose case (ii) occurs. All jobs in D q (k, l) get dispatched by time t q e . So they must get released before t q e k l . So at least 1/4 volume of these jobs wait for at least k l /2 amount of time in O. This means that F O D q (k,l) is at least V /8 l , because each job has size at most k . This again implies the lemma. Case (iii) is similar to case (i). So we can assume that none of the cases in Lemma 4.9 occur. Now, consider a time t between t q e k l /2 and t q e . Let us look at machines of class 1 to l - 1. O finishes at least V /4 volume of D q (k, l) on these machines before time t. Further, at most V /8 volume of H q goes out from these machines to machines of higher class. The volume that corresponds to time slots in which these machines do not process jobs of H q in A is at most V/16 (for a sufficiently large constant c). So, at least V /16 amount of volume of H q must be waiting in O at time t. So the total flow time incurred by O on H q is at least V /(16 k ) k l /2 which again is (V l ). This proves the theorem. Combining the above theorem with Lemma 4.5, we get Corollary 4.10. P A J(k,l) is O(F O ), and so P A is O(log S log P F O ). 4.2 Bounding the flow time We now show that the average flow time incurred by our algorithm is within poly-logarithmic factor of that incurred by the optimal solution. We shall say that a job j is waiting at time t in A if it is in the central pool at time t in A and has been released before time t. Let V A k (t) denote the volume of jobs of class at most k which are waiting at time t in A. Define V O k (t) as the remaining volume of jobs of class at most k which are active at time t in O. Note the slight difference in the definitions for A and O -- in case of A, we are counting only those jobs which are in the central pool, while in O we are counting all active jobs. Our goal is to show that for all values of k, P t V A k (t) is bounded from above by P t V O k (t) plus small additive factors, i.e., O( k (P O + P A )). Once we have this, the desired result will follow from standard calculations. Before we go to the details of the analysis, let us first show how to prove such a fact when all machines are of the same speed, say 1. The argument for this case follows directly from [5], but we describe it to develop some intuition for the more general case. Fix a time t and class k of jobs. Suppose all machines are processing jobs of class at most k at time t in our schedule. Let t be the first time before t at which there is at least one machine in A which is not processing jobs of class k or less (if there is no such time, set t as 0). It follows that at time t there are no jobs of class k or less which are mature for these machines (since all machines are identical, a job becomes mature for all the machines at the same time). So these jobs must have been released after t k . During (t k , t ), the optimal schedule can process at most m k volume of jobs of class at most k. So it follows that V A k (t ) - V O k (t ) is at most m k . Since our schedule processes jobs of class at most k during (t , t), we get V A k (t) - V O k (t) x A (t) k , where x A (t) denotes the number of busy machines at timer t in our schedule (x A (t) is same as m because all machines are busy at time t). The other case when a machine may be processing jobs of class more than k or remain idle at time t can be shown similarly to yield the same expression. Adding this for all values of t, we get P t V A k (t) - P t V O k (t) is at most k P A , which is what we wanted to show. This argument does not extend to the case when machines can have different speeds. There might be a very slow machine which is not processing jobs of class k or less (it may be idle), but then the jobs which are waiting could have been released much earlier and so we cannot argue that the 734 volume of remaining jobs of class at most k at time t in the two schedules are close. This complicates our proof and instead of defining one time t as above, we need to define a sequence of times. Before we describe this process, we need more notations. These are summarized in the table below for ease of reference . Let j k (t) J k (t) be the job with the earliest release date. Let r k (t) denote the release date of job j k (t), and c k (t) denote the class of j k (t). Let l k (t) denote the largest l such that j k (t) has matured for machines of class l at time t. In other words, l k (t) is the highest value of l such that t M (j, l) t. Observe that all machines of class l k (t) or less must be busy processing jobs of class at most c k (t) at time t, otherwise our algorithm should have dispatched j by time t. Our proof will use the following lemma. Lemma 4.11. For any class k of jobs and time t, V A k (t) - V O k (t) 2 c k (t) m (l k (t)-1) + V A (c k (t)-1) (r k (t)) - V O (c k (t)-1) (r k (t)) + X ll k (t) P O l (r k (t), t) l-1 Proof. Let V A denote the volume of jobs of class at most k which are processed by A on machines of class l k (t) - 1 or less during (r k (t), t). Define U A as the volume of jobs of class at most k which are processed by A on machines of class l k (t) or more during (r k (t), t). Define V O and U O similarly. Clearly, V A k (t) - V O k (t) = V A k (r k (t)) - V O k (r k (t)) + (V O V A ) + (U O - U A ). Let us look at V O - V A first. Any machine of class l l k (t) - 1 is busy in A during (t M (j k (t), l ), t) processing jobs of class at most c k (t). The amount of volume O can process on such machines during (r k (t), t M (j k (t), l )) is at most m l ck(t) l l -1 , which is at most m l c k (t) . So we get V O - V A m (l k (t)-1) c k (t) . Let us now consider V A k (r k (t)) - V O k (r k (t)). Let us consider jobs of class c k (t) or more which are waiting at time r k (t) in A (so they were released before r k (t)). By our definition of j k (t), all such jobs must be processed by time t in A. If l l k (t) - 1, then such jobs can be done on machines of class l only during (t M (j k (t), l ), t). So again we can show that the total volume of such jobs is at most m (l k (t)-1) c k (t) + U A . Thus we get V A k (r k (t)) V O k (r k (t)) m (l k (t)-1) c k (t) +U A +V A (c k (t)-1) (r k (t))V O (c k (t)-1) (r k (t)), because V O (c k (t)-1) (r k (t)) V O k (r k (t)). Finally note that U O is at most P ll k (t) P O l (r k (t),t) l-1 . Combining everything, we get the result. The rest of the proof is really about unraveling the expression in the lemma above. To illustrate the ideas involved, let us try to prove the special case for jobs of class 1 only. The lemma above implies that V A 1 (t) - V O 1 (t) 2 m (l 1 (t)-1) +P ll 1 (t) P O l (r 1 (t),t) l-1 . We are really interested in P t (V A 1 (t) - V O 1 (t)). Now, the sum P t m (l 1 (t)-1) is not a problem because we know that at time t all machines of class l 1 (t) or less must be busy in A. So, x A (t) m (l 1 (t)-1) . So P t m (l 1 (t)-1) is at most P t x A (t), which is the total processing time of A. It is little tricky to bound the second term. We shall write P ll 1 (t) P O l (r 1 (t),t) l-1 as P ll 1 (t) P t t =r 1 (t) x O l (t ) l-1 . We can think of this as saying that at time time t , r 1 (t) t t, we are charging 1/ l-1 amount to each machine of class l which is busy at time t in O. Note that here l is at least l 1 (t). Now we consider P t P ll 1 (t) P t t =r 1 (t) x O l (t ) l-1 . For a fixed time t and a machine i of class l which is busy at time t in O, let us see for how many times t we charge to i. We charge to i at time t if t lies in the interval (r 1 (t), t), and l l 1 (t). Suppose this happens. We claim that t - t has to be at most l+1 . Indeed otherwise t - r 1 (t) l+1 and so j 1 (t) has matured for machines of class l + 1 as well. But then l &lt; l 1 (t). So the total amount of charge machine i gets at time t is at most l+1 1/ l-1 = O(1). Thus, the total sum turns out to be at most a constant times the total processing time of O. Let us now try to prove this for the general case. We build some notation first. Fix a time t and class k. We shall define a sequence of times t 0 , t 1 , and a sequence of jobs j 0 (t), j 1 (t), . . . associated with this sequence of times. c i (t) shall denote the class of the job j i (t). Let us see how we define this sequence. First of all t 0 = t, and j 0 (t) is the job j k (t) (as defined above). Recall the definition of j k (t) it is the job with the earliest release date among all jobs of class at most k which are waiting at time t in A. Note that c 0 (t), the class of this job, can be less than k. Now suppose we have defined t 0 , . . . , t i and j 0 (t), . . . , j i (t), i 0. t i+1 is the release date of j i (t). j i+1 (t) is defined as the job j c i (t)-1 (t i+1 ), i.e., the job with the earliest release date among all jobs of class less than c i (t) waiting at time t i+1 in A. We shall also define a sequence of classes of machines l 0 (t), l 1 (t), . . . in the following manner l i (t) is the highest class l of machines such that job j i (t) has matured for machines of class l at time t i . Figure 2 illustrates these definitions. The vertical line at time t i denotes l i (t) the height of this line is proportional to l i (t). We note the following simple fact. Claim 4.12. l i (t) c i (t) t i - t i+1 l i (t)+1 c i (t) . Proof. Indeed t i+1 is the release date of j i (t) and j i (t) matures for machines of class l i (t) at time t i , but not for machines of class l i+1 (t) at time t i . The statement in Lemma 4.11 can be unrolled iteratively to give the following inequality : V A k (t) - V O k (t) 2 ( c 0 (t) m &lt;l 0 (t) + c 1 (t) m &lt;l 1 (t) + ) + 0 @ X ll 0 (t) P O l (t 1 , t 0 ) l-1 + X ll 1 (t) P O l (t 2 , t 1 ) l-1 + 1 A (2) Let us try to see what the inequality means. First look at the term c i (t) m &lt;l i (t) . Consider a machine of class l &lt; l i (t). It is busy in A during (t M (j i (t), l), t i ). Now t i - t M (j i (t), l) = (t i - t i+1 ) - (t M (j i (t), l) - t i+1 ) l i (t) c i (t) c i (t) l c i (t) l , as l &lt; l i (t). Hence machine l is also busy in A during (t i c i (t) l , t i ), So the term c i (t) m &lt;l i (t) is essentially saying that we charge 1/ l amount to each machine of class l &lt; l i (t) during each time in (t i 735 x A (t), x O (t) Number of machines busy at time t in A, O. x A l (t), x O l (t) Number of machines of class l busy at time t in A, O. P A l (t 1 , t 2 ), P O l (t 1 , t 2 ) Total processing time incurred by machines of class l during (t 1 , t 2 ) in A, O. m l , m l , m &lt;l Number of machines of class l, at most l, less than l. m (l 1 ,l 2 ) Number of machines of class between (and including) l 1 and l 2 . J(k, t) Set of jobs of class at most k which are waiting at time t in A. Table 1: Table of definitions t t t t t 1 0 2 3 4 l l l l l 0 1 2 3 4 t 5 5 l t 6 r jj j r r r r r j j j j j 0 1 2 3 4 5 Figure 2: Illustrating the definitions of t 0 , . . . and c 0 (t), . . .. c i (t) l , t i ). These intervals are shown in figure 2 using light shade. So a machine i of class l which is busy in A during these lightly shaded regions is charged 1/ l units. Let us now look at the term P ll i (t) P O l (t i+1 ,t i ) l-1 . This means the following consider a machine of class l l i (t) which is busy in O at some time t during (t i+1 , t i ) then we charge it 1/ l-1 units. Figure 2 illustrates this fact we charge 1/ l-1 to all machines of class l l i (t) which are busy in O during the darkly shaded regions. Let us see how we can simplify the picture. We say that the index i is suffix maximum if l i (t) &gt; l i-1 (t), . . . , l 0 (t) (i = 0 is always suffix maximum). In Figure 2, t 0 , t 2 and t 5 are suffix maximum. Let the indices which are suffix maximum be i 0 = 0 &lt; i 1 &lt; i 2 &lt; . . .. The following lemma says that we can consider only suffix maximum indices in (2). We defer its proof to the appendix. Lemma 4.13. V A k (t) - V O k (t) 4 2 X i u c iu (t) m (l iu-1 (t),l iu (t)-1) + X i u X ll iu (t) P O l (t i u+1 , t i u ) l-1 , where i u varies over the suffix maximum indices (define l i -1 (t) as 1). Recall that m (l 1 ,l 2 ) denotes the number of machines of class between l 1 and l 2 . Figure 3 shows what Lemma 4.13 is saying. In the lightly shaded region, if there is a machine of class l which is busy at some time t in A, we charge it 1/ l units at time t . In the darkly shaded region if there is a machine i of class l which is busy at time t in O we charge it 1/ l-1 units. Now we try to bound the charges on the darkly shaded region for all time t. Let us fix a machine h of class l. Suppose h is busy in O at time t . We are charging 1/ l-1 amount to h at time t if the following condition holds : for all i such that t i t , l i (t) is at most l. Now we ask, if we fix t , for how many values of t do we charge h at time t ? The following claim shows that this can not be too large. Claim 4.14. Given a machine h of class l, we can charge it for the darkly shaded region at time t for at most 2 l+1 k values of t. Proof. Suppose we charge h at time t for some value of t. Fix this t. Clearly t t . Let i be the largest index such that t t i . So t - t t - t i+1 . Now consider any i i. Lemma 4.12 implies that t i - t i +1 l i (t)+1 i (t) l+1 c i (t) . Since c i (t) decrease as i increases, P i i =0 (t i - t i +1 ) 2 l+1 k . This implies the claim. So the total amount of charge to machine h at time t is at most 2 2 k . Thus we get the following fact : X t 0 @X i u X ll iu (t) P O l (t i u +1 , t i u ) l-1 1 A 2 2 k P O (3) Now we look at the charges on the lightly shaded region. Let h be a machine of class l which is busy in A at time t . As argued earlier, we charge 1/ l units to h at time t if the following condition holds there exists a suffix maximum index i u such that t lies in the interval (t i u c iu (t) l , t i u ). Further for all suffix maximum indices i &lt; i u , it must be the case that l i (t) &lt; l. Now we want to know for how many values of t do we charge h at time t . 736 t t t t t 1 0 2 3 4 l l l l l 0 1 2 3 4 t 5 5 l t 6 r jj j r r r r r j j j j j 0 1 2 3 4 5 Figure 3: Illustrating Lemma 4.13 Claim 4.15. Given a machine h of class l, we can charge it for the lightly shaded region at time t for at most 3 l+1 k values of t. Proof. Fix a time t such that while accounting for V A k (t) we charge h at time t . So there is a maximum index i u such that t i u - t c iu (t) l . Further, if i is an index less than i u , then l i (t) must be less than l. We can argue as in the proof of Claim 4.14 that t - t i u can be at most 2 l+1 k . So t - t 3 l+1 k . So we get X t X i u c iu (t) m (l iu-1 (t),l iu (t)-1) ! 3 k P A (4) Putting everything together, we see that Lemma 4.13, and equations (3) and (4) imply that there is a constant c such that X t V A k (t) X t V O k (t) + c k (P O + P A ). (5) The final result now follows from standard calculations. Theorem 4.16. F A is O(log S log 2 P F O ). Proof. We have already bounded the processing time of A. Once a job gets dispatched to a machine i, its waiting time can be charged to the processing done by i. Since at any time t, there are at most log P active jobs dispatched to a machine, the total waiting time of jobs after their dispatch time is at most O(log P P A ). So we just need to bound the time for which jobs are waiting in the central pool. Let n A k (t) be the number of jobs of class k waiting in the central pool at time t in our algorithm. Let n O k (t) be the number of jobs of class k which are active at time t in O (note the difference in the definitions of the two quantities). Since jobs waiting in the central pool in A have not been processed at all, it is easy to see that n A k (t) V A k (t) k . Further , V O k (t) k n O k (t) + + n O 1 (t). Combining these observations with equation (5), we get for all values of k, X t n A k (t) X t ,, n O k (t) + n O k-1 (t) + + n O 1 (t) k-1 +c (P O + P A ). We know that total flow time of a schedule is equal to the sum over all time t of the number of active jobs at time t in the schedule. So adding the equation above for all values of k and using Corollary 4.10 implies the theorem. ACKNOWLEDGEMENTS We would like to express our thanks to Gagan Goel, Vinayaka Pandit, Yogish Sabharwal and Raghavendra Udupa for useful discussions. REFERENCES [1] N. Avrahami and Y. Azar. Minimizing total flow time and total completion time with immediate dispatching. In Proc. 15th Symp. on Parallel Algorithms and Architectures (SPAA), pages 1118. ACM, 2003. [2] Baruch Awerbuch, Yossi Azar, Stefano Leonardi, and Oded Regev. Minimizing the flow time without migration. In ACM Symposium on Theory of Computing, pages 198205, 1999. [3] N. Bansal and K. Pruhs. Server scheduling in the l p norm: A rising tide lifts all boats. In ACM Symposium on Theory of Computing, pages 242250, 2003. [4] Luca Becchetti, Stefano Leonardi, Alberto Marchetti-Spaccamela, and Kirk R. Pruhs. Online weighted flow time and deadline scheduling. Lecture Notes in Computer Science, 2129:3647, 2001. [5] C. Chekuri, S. Khanna, and A. Zhu. Algorithms for weighted flow time. In ACM Symposium on Theory of Computing, pages 8493. ACM, 2001. [6] Chandra Chekuri, Ashish Goel, Sanjeev Khanna, and Amit Kumar. Multi-processor scheduling to minimize flow time with epsilon resource augmentation. In ACM Symposium on Theory of Computing, pages 363372, 2004. [7] R. L. Graham, E. L. Lawler, J. K. Lenstra, and A. H. G. Rinnooy Kan. Optimization and approximation in deterministic sequencing and scheduling : a survey. Ann. Discrete Math., 5:287326, 1979. [8] Bala Kalyanasundaram and Kirk Pruhs. Speed is as powerful as clairvoyance. In IEEE Symposium on 737 Foundations of Computer Science, pages 214221, 1995. [9] Hans Kellerer, Thomas Tautenhahn, and Gerhard J. Woeginger. Approximability and nonapproximability results for minimizing total flow time on a single machine. In ACM Symposium on Theory of Cmputing, pages 418426, 1996. [10] Stefano Leonardi and Danny Raz. Approximating total flow time on parallel machines. In ACM Symposium on Theory of Computing, pages 110119, 1997. [11] C. A. Phillips, C. Stein, E. Torng, and J. Wein. Optimal time-critical scheduling via resource augmentation. In ACM Symposium on Theory of Computing, pages 140149, 1997. Appendix Proof of Theorem 4.6. Let M denote the set of machines of class less than l. First observe that the processing time incurred by A on J I is at most twice of |M | T (the factor twice comes because there may be some jobs which are dispatched during I but finish later -- there can be at most one such job for a given class and a given machine). So we will be done if we can show that F O J I is ( |M | T ). Let V be the volume of jobs in J I which are done by O on machines of class l or more. If V |M |T l+1 , then we are done because then the processing time incurred by O on J I is at least V l . So we will assume in the rest of the proof that V |M |T l+1 . Let i be a machine of class less than l. We shall say that i is good is i processes jobs from J I for at least T /4 units of time during I in the optimal solution O. Otherwise we say that i is bad. Let G denote the set of good machines. If |G| |M | , then we are done again, because P O J I is at least |G| T/4. Let B denote the set of bad machines. So we can assume that the number of good machines is at most 1/ fraction of the number of machines of class less than l. Now consider a time t in the interval (t b + T /2, t e ). Claim 6.1. At time t, at least k |M | volume of jobs from J I is waiting in O. Proof. Let V 1 denote the volume of jobs from J I which is done by A during (t b , t). Let V 2 denote the volume of jobs from J I which is done by O on machines of class less than l during (t b , t). Recall that for a machine i, s i denotes the slowness of i. Since a machine i of class l &lt; l does not perform jobs from J I for at most 6 k l amount of time during I, we see that V 1 P iM t-t b -6 k ci s i , where c i denotes the class of i. Let us look at V 2 now. In O all bad machines do not process jobs from J I for at least 3T /4 units of time during I. So they do not process jobs from J I for at least T /4 units of time during (t b , t b +T /2). So V 2 P iM (t-t b ) s i -P iB T 4s i . This shows that V 1 - V 2 P iB T 4s i - P iM 6 k ci s i . For a bad machine i, T /4 - 6 k c i T/8, since c i l 1 (assuming is large enough). So, we can see that this difference is at least P iB T 8s i - P i / B 6 k . Since T l k , we see that this difference is at least T l-1 ( |B|/8 - 6|G|), which is at least T |M | 10 l-1 , because we have assumed that |B| is larger than |G| by a sufficiently high constant factor. Recall that V is the volume of jobs in J I which is done by O on machines of class l or more. Clearly, the volume of jobs from J I which is waiting at time t in O is at least V 1 - V 2 - V . But V is at most T |M | l+1 . Hence the volume waiting at time t is at least T |M | l . This proves the lemma. Since each job in J I is of size at most k , we see that at least ( |M |) jobs are waiting at time t. Summing over all values of t in the range (t b + T /2, t e ) implies the theorem. Proof of Lemma 4.13. Consider an i, i u+1 &lt; i i u . Then l i (t) l i u (t), otherwise we should have another suffix maximal index between i u+1 and i u . So P i u+1 -1 i=i u c i (t) m &lt;l i (t) m &lt;l iu (t) P i u -1 i=i u c i (t) 2 m &lt;l iu (t) c iu (t) . So we get P i c i (t) m &lt;l i (t) 2 P i u m &lt;l iu (t) c iu (t) . Now we consider the sum P i P ll i (t) P O l (t i+1 ,t i ) l-1 . Fix an i, i u+1 &lt; i i u . Using Claim 4.12, we see that P O l (t i+1 , t i ) m l l i (t)+1 c i (t) . So we get X ll i (t) P O l (t i+1 , t i ) l-1 X ll iu (t) P O l (t i+1 , t i ) l-1 + l iu (t)-1 X l=l i (t) m l l i (t)+1 c i (t) l-1 Now the second term on the right hand side above is at most 2 m &lt;l iu (t) c i (t) . So we get i u+1-1 X i=i u X ll i (t) P O l (t i+1 , t i ) l-1 X ll iu (t) P O l (t i u+1 , t i u ) l-1 + 2 2 c iu (t) m &lt;l iu (t) , because c i (t) scales down geometrically as i increases. Finally note that P i u c iu (t) m &lt;l iu (t) is at most twice of P i u c iu (t) m (l iu-1 (t),l iu (t)-1) , because c i (t) scale down geometrically. This proves the lemma (using (2)). 738

non-migratory algorithm;flow-time;average flow time;approximation algorithms;processing time;competitive ratio;related machines;poly-logarithmic factor;preemption;multiprocessor environment;scheduling;Scheduling

138

Modeling and Predicting Personal Information Dissemination Behavior

In this paper, we propose a new way to automatically model and predict human behavior of receiving and disseminating information by analyzing the contact and content of personal communications. A personal profile, called CommunityNet, is established for each individual based on a novel algorithm incorporating contact, content, and time information simultaneously. It can be used for personal social capital management. Clusters of CommunityNets provide a view of informal networks for organization management. Our new algorithm is developed based on the combination of dynamic algorithms in the social network field and the semantic content classification methods in the natural language processing and machine learning literatures. We tested CommunityNets on the Enron Email corpus and report experimental results including filtering, prediction, and recommendation capabilities. We show that the personal behavior and intention are somewhat predictable based on these models. For instance, &quot;to whom a person is going to send a specific email&quot; can be predicted by one's personal social network and content analysis. Experimental results show the prediction accuracy of the proposed adaptive algorithm is 58% better than the social network-based predictions, and is 75% better than an aggregated model based on Latent Dirichlet Allocation with social network enhancement. Two online demo systems we developed that allow interactive exploration of CommunityNet are also discussed.

INTRODUCTION Working in the information age, the most important is not what you know, but who you know [1]. A social network, the graph of relationships and interactions within a group of individuals, plays a fundamental role as a medium for the spread of information, ideas, and influence. At the organizational level, personal social networks are activated for recruitment, partnering, and information access. At the individual level, people exploit their networks to advance careers and gather information. Informal network within formal organizations is a major, but hard to acquire, factor affecting companies' performance. Krackhardt [2] showed that companies with strong informal networks perform five or six times better than those with weak networks, especially on the long-term performance. Friend and advice networks drive enterprise operations in a way that, if the real organization structure does not match the informal networks, then a company tends to fail [3]. Since Max Weber first studied modern bureaucracy structures in the 1920s, decades of related social scientific researches have been mainly relying on questionnaires and interviews to understand individuals' thoughts and behaviors for sensing informal networks. However, data collection is time consuming and seldom provides timely, continuous, and dynamic information. This is usually the biggest hurdle in social studies. Personal Social Network (PSN) could provide an organizing principle for advanced user interfaces that offer information management and communication services in a single integrated system. One of the most pronounced examples is the networking study by Nardi et al. [4], who coined the term intensional networks to describe personal social networks. They presented a visual model of user's PSN to organize personal communications in terms of a social network of contacts. From this perspective, many tools were built such as LinkedIn [5], Orkut [6], and Friendster [7]. However, all of them only provide tools for visually managing personal social networks. Users need to manually input, update, and manage these networks. This results in serious drawbacks. For instance, people may not be able to invest necessary efforts in creating rich information, or they may not keep the information up-to-date as their interests, responsibilities, and network change. They need a way to organize the relationship and remember who have the resources to help them. We coin the terminology of managing these goals as personal social capital management 1 . In this paper, we develop a user-centric modeling technology, which can dynamically describe and update a person's personal social network with context-dependent and temporal evolution information from personal communications. We refer to the model as a CommunityNet. Senders and receivers, time stamps, subject and content of emails contribute three key components content semantics, temporal information, and social relationship. We propose a novel Content-Time-Relation (CTR) algorithm to capture dynamic and context-dependent information in an unsupervised way. Based on the CommunityNet models, many questions can be addressed by inference, prediction and filtering. For instance, 1) Who are semantically related to each other? 2) Who will be involved in a special topic? Who are the important (central) people in this topic? 3) How does the information flow? and 4) If we want to publicize a message, whom should we inform? Figure 1 shows the procedure of our proposed scheme. First, topic detection and clustering is conducted on training emails in order to define topic-communities. Then, for each individual, CommunityNet is built based on the detected topics, the sender and receiver information, and the time stamps. Afterwards, these personal CommunityNets can be applied for inferring organizational informal networks and predicting personal behaviors to help users manage their social capitals. We incorporate the following innovative steps: 1) Incorporate content analysis into social network in an unsupervised way 2) Build a CommunityNet for each user to capture the context-dependent , temporal evolutionary personal social network based on email communication records 3) Analyze people's behaviors based on CommunityNet, including predicting people's information sending and receiving behaviors 4) Show the potential of using automatically acquired personal social network for organization and personal social capital management Input: Emails From: sally.beck@enron.com To: shona.wilson@enron.com Subject: Re: timing of submitting information to Risk Controls Good memo - let me know if you see results. ...... Topic Detection, Content Analysis Topics Meeting schedule Agreement California Energy Game Holiday celebration CommunityNet CommunityNet Modeling Applications Recommendation system Prediction, Filtering Input: Emails From: sally.beck@enron.com To: shona.wilson@enron.com Subject: Re: timing of submitting information to Risk Controls Good memo - let me know if you see results. ...... Topic Detection, Content Analysis Topics Meeting schedule Agreement California Energy Game Holiday celebration CommunityNet CommunityNet Modeling Applications Recommendation system Prediction, Filtering Figure 1. An Overview of CommunityNet We tested the CommunityNet model on the Enron email corpus comprising the communication records of 154 Enron employees dating from Jan. 1999 to Aug. 2002. The Enron email dataset was originally made available to public by the Federal Energy Regulatory Commission during the investigation [9]. It was later collected and prepared by Melinda Gervasio at SRI for the CALO (A Cognitive Assistant that Learns and Organizes) project. William Cohen from CMU has put up the dataset on the web for research purpose [9]. This version of the dataset contains around 517,432 emails within 150 folders. We clean the data and extract 154 users from those 150 folders with 166,653 unique messages from 1999 to 2002. In the experiments, we use 16,873 intra-organizational emails which connect these 154 people. The primary contributions of this paper are three-fold. First we develop an algorithm incorporating content-time-relation detection. Second, we generate an application model which describes personal dynamic community network. Third, we show how this model can be applied to organization and social capital management. To the best of our knowledge, this is among the first reported technologies on fusing research in the social network analysis field and the content analysis field for information management. We propose the CTR algorithm and the CommunityNet based on the Latent Dirichlet Allocation algorithm. In our experiments, we observed clear benefit of discovering knowledge based on multi-modality information rather than using only single type of data. The rest of the paper is organized as follows. In Section 2, we present an overview of related work. In Section 3, we present our model. We discuss how to use CommunityNet to analyze communities and individuals in section 4 and 5, respectively. In Section 6, we show two demo systems for query, visualization and contact recommendation. Finally, conclusions and future work are addressed in Section 7. RELATED WORK To capture relationships between entities, social network has been a subject of study for more than 50 years. An early sign of the potential of social network was perhaps the classic paper by Milgram [10] estimating that on average, every person in the world is only six edges away from each other, if an edge between i and j means &quot;i knows j&quot;. Lately, introducing social network analysis into information mining is becoming an important research area. Schwartz and Wood [11] mined social relationships from email logs by using a set of heuristic graph algorithms. The Referral Web project [12] mined a social network from a wide variety of publicly-available online information, and used it to help individuals find experts who could answer their questions based on geographical proximity. Flake et al. [13] used graph algorithms to mine communities from the Web (defined as sets of sites that have more links to each other than to non-members). Tyler et al. [14] use a betweenness centrality algorithm for the automatic identification of communities of practice from email logs within an organization. The Google search engine [15] and Kleinberg's HITS algorithm of finding hubs and authorities on the Web [16] are also based on social network concepts. The success of these approaches, and the discovery of widespread network topologies with nontrivial properties, have led to a recent flurry of research on applying link analysis for information mining. A promising class of statistical models for expressing structural properties of social networks is the class of Exponential Random Graph Models (ERGMs) (or p* model) [17]. This statistical model can represent structural properties that define complicated dependence patterns that cannot be easily modeled by deterministic models. Let Y denote a random graph on a set of n nodes and let y denote a particular graph on those nodes. Then, the probability of Y equals to y is ( ) ( ) ( ) ( ) exp T s y P Y y c = = (1) 480 Industry/Government Track Paper where ( ) s y is a known vector of graph statistics (Density, Reciprocity, Transitivity, etc) on y, is a vector of coefficients to model the influence of each statistics for the whole graph, T means "transpose", ( ) c is a normalization term to satisfy ( ) 1 y P Y y = = . The parameters are estimated based on the observed graph obs y by maximum likelihood estimation. All the research discussed above has focused on using static properties in a network to represent the complex structure. However, social networks evolve over time. Evolution property has a great deal of influence; e.g., it affects the rate of information diffusion, the ability to acquire and use information, and the quality and accuracy of organizational decisions. Dynamics of social networks have attracted many researchers' attentions recently. Given a snapshot of a social network, [19] tries to infer which new interactions among its members are likely to occur in the near future. In [20], Kubica et al. are interested in tracking changes in large-scale data by periodically creating an agglomerative clustering and examining the evolution of clusters over time. Among the known dynamical social networks in literature, Snijder's dynamic actor-oriented social network [18] is one of the most successful algorithms. Changes in the network are modeled as the stochastic result of network effects (density, reciprocity, etc.). Evolution is modeled by continuous-time Markov chains, whose parameters are estimated by the Markov chain Monte Carlo procedures. In [21], Handcock et al. proposed a curved ERGM model and applied it to the new specifications of ERGMs This latest model uses nonlinear parameters to represent structural properties of networks. The above mentioned dynamic analyses show some success in analyzing longitudinal stream data. However, most of them are only based on pure network properties, without knowing what people are talking about and why they have close relationships. 2.2 Content Analysis In statistical Natural Language processing, one common way of modeling the contributions of different topics to a document is to treat each topic as a probability distribution over words, viewing a document as a probability distribution over words, and thus viewing a document as a probabilistic mixture over these topics. Given T topics, the probability of the ith word in a given document is formalized as: ( ) ( ) ( ) 1 | T i i i i j P w P w z j P z j = = = = (2) where i z is a latent variable indicating the topic from which the i th word was drawn and ( ) | i i P w z j = is the probability of the word i w under the jth topic. ( ) i P z j = gives the probability of choosing a word from topics j in the current document, which varies across different documents. Hofmann [22] introduced the aspect model Probabilistic Latent Semantic Analysis (PLSA), in which, topics are modeled as multinomial distributions over words, and documents are assumed to be generated by the activation of multiple topics. Blei et al. [23] proposed Latent Dirichlet Allocation (LDA) to address the problems of PLSA that parameterization was susceptible to overfitting and did not provide a straightforward way to infer testing documents. A distribution over topics is sampled from a Dirichlet distribution for each document. Each word is sampled from a multinomial distribution over words specific to the sampled topic. Following the notations in [24], in LDA, D documents containing T topics expressed over W unique words, we can represent ( | ) P w z with a set of T multinomial distributions over the W words, such that ( ) ( | ) w j P w z j = = , and P(z) with a set of D multinomial distribution over the T topics, such that for a word in document d, ( ) ( ) d j P z j = = . Recently, the Author-Topic (AT) model [25] extends LDA to include authorship information, trying to recognize which part of the document is contributed by which co-author. In a recent unpublished work, McCallum et al. [26] further extend the AT model to the Author-Recipient-Topic model by regarding the sender-receiver pair as an additional author variable for topic classification. Their goal is role discovery, which is similar to one of our goals as discussed in Sec. 4.1.2 without taking the temporal nature of emails into consideration. Using LDA, and are parameters that need to be estimated by using sophisticated approximation either with variational Bayes or expectation propagation. To solve this problem, Griffiths and Steyvers [24] extended LDA by considering the posterior distribution over the assignments of words to topics and showed how Gibbs sampling could be applied to build models. Specifically, ( ) ( ) ( ) ( ) ( ) , , , , | , i i i w d i j i j i i d i j i n n P z j z w n W n T + + = + + (3) where ( ) i n is a count that does not include the current assignment, ( ) w j n is the number of times word w has been assigned to topic j in the vector of assignments z, ( ) d j n is the number of times a word from document d has been assigned to topic j, ( ) j n is a sum of ( ) w j n , ( ) d n is a sum of ( ) d j n . Further, one can estimate ( ) w j , the probability of using word w in topic j, and ( ) d j , the probability of topic j in document d as follows: ( ) ( ) ( ) ^ w w j j j n n W + = + (4) ( ) ( ) ( ) ^ d d j j d n n T + = + (5) In [24], experiments show that topics can be recovered by their algorithm and show meaningful aspects of the structure and relationships between scientific papers. Contextual, relational, and temporal information are three key factors for current data mining and knowledge management models. However, there are few papers addressing these three components simultaneously. .In our recent paper, we built user models to explicitly describe a person's expertise by a relational and evolutionary graph representation called ExpertisetNet [27]. In this paper, we continue exploring this thread, and build a CommunityNet model which incorporates these three components together for data mining and knowledge management. COMMUNITYNET In this section, we first define terminologies. Then, we propose a Content-Time-Relation (CTR) algorithm to build the 481 Industry/Government Track Paper personal CommunityNet. We also specifically address the prediction of the user's behaviors as a classification problem and solve it based on the CommunityNet models. 3.1 Terminology Definition 1. Topic-Community: A topic community is a group of people who participate in one specific topic. Definition 2: Personal Topic-Community Network (PTCN): A personal topic-community network is a group of people directly connected to one person about a specific topic. Definition 3. Evolutionary Personal Social Network: An evolutionary personal social network illustrates how a personal social network changes over time. Definition 4. Evolutionary Personal Topic-Community Network: An evolutionary network illustrates how a person's personal topic-community network changes over time. Definition 5. Personal Social Network Information Flow: A personal social network information flow illustrates how the information flows over a person's personal social network to other people's personal social networks Definition 6: Personal Topic-Community Information Flow: A personal Topic-CommunityNet information flow illustrates how the information about one topic flows over a person's personal social network to other people's personal social networks. 3.2 Personal Social Network We build people's personal social networks by collecting their communication records. The nodes of a network represent whom this person contacts with. The weights of the links measure the probabilities of the emails he sends to the other people: A basic form of the probability that an user u sending email to a recipient r is: ( ) number of times sends emails to | total number of emails sent out by u r P r u u = (6) We build evolutionary personal social networks to explore the dynamics and the evolution. The ERGM in Eq. (1) can be used to replace Eq. (6) for probabilistic graph modeling. A big challenge of automatically building evolutionary personal social network is the evolutionary segmentation, which is to detect changes between personal social network cohesive sections. Here we apply the same algorithm as we proposed in [27]. For each personal social network in one time period t, we use the exponential random graph model [17] to estimate an underlying distribution to describe the social network. An ERGM is estimated from the data in each temporal sliding window. With these operations, we obtain a series of parameters which indicates the graph configurations. 3.3 Content-Time-Relation Algorithm We begin with email content, sender and receiver information, and time stamps, and use these sources of knowledge to create a joint probabilistic model. An observation is (u, r, d, w, t) corresponds to an event of a user u sending to receivers r an email d containing words w during a particular time period t. Conceptually, users choose latent topics z, which in turn generate receivers r, documents d, and their content words w during time period t. ( ) ( ) ( ) , | , , | , | , z P u r d t P u r z t P z d t = (7) where , u r is a sender-receiver pair during time period t . , u r can be replaced by any variable to indicate the user's behavior, as long as it is also assumed to be dependent on latent topics of emails. In order to model the PTCN, one challenge is how to detect latent topics dynamically and at the same time track the emails related to the old topics. This is a problem similar to topic detection tracking [28]. We propose an incremental LDA (ILDA) algorithm to solve it, in which the number of topics is dynamically updated based on the Bayesian model selection principle [24]. The procedures of the algorithm are illustrated as follows: Incremental Latent Dirichlet Allocation (ILDA) algorithm: Input: Email streams with timestamp t Output: ( ) , w j t , ( ) , d j t for different time period t Steps: 1) Apply LDA on a data set with currently observed emails in a time period t to generate latent topics j z and estimate ( ) ( ) 0 0 , | , w j j t P w z t = and ( ) ( ) 0 0 , | , d j j t P z d t = by equation (4) and (5). The number of topics is determined by the Bayesian model selection principle. 2) When new emails arrive during time period k, use Bayesian model selection principle to determine the number of topics and apply ( ) ( ) ( ) ( ) , 1 , | , , | , i i d i j i i k i k d i n P z j z w t P w z j t n T + = = + to estimate ( ) | , k P z d t , ( ) | , k P w z t , and ( ) | , k P z w t . 3) Repeat step 2) until no data arrive. Based on this ILDA algorithm, we propose a Content-Time-Relation (CTR) algorithm. It consists of two phases, the training phase and the testing phase. In the training phase, emails as well as the senders, receivers and time stamps are available. ( ) | , old P w z t and ( ) , | , old P u r z t are learnt from the observed data. In the testing phase, we apply ILDA to learn ( ) | , new P z d t . Based on ( ) , | , old P u r z t , which is learnt from the training phase, , u r can be inferred. Again, , u r represents a sender-receiver pair or any variable to indicate the user's behavior, as long as it is dependent on the latent topics of emails. Content-Time-Relation (CTR) algorithm: 1) Training phase Input: Old emails with content, sender and receiver information, and time stamps old t Output: ( ) ( ) ( ) | , , | , , and , | , old old old P w z t P z d t P u r z t Steps: a) Apply Gibbs Sampling on the data according to equation (3). b) Estimate ( ) ( ) , | , old w j old j t P w z t = and ( ) ( ) , | , old d j old j t P z d t = by equation (4), and (5). c) Estimate ( ) ( ) ( ) ( ) ( ) , | , , | , | , , | , | , old old old d old old d P u r z t P u r d t P d z t P u r d t P z d t = (8) 2) Testing phase Input: New emails with content and time stamps new t 482 Industry/Government Track Paper Output: ( ) ( ) ( ) , | , , | , , and | , new new new P u r d t P w z t P z d t Steps: a) Apply incremental LDA by Gibbs Sampling based on ( ) ( ) ( ) ( ) , , | , , , | i i d i j i i new i old d i n P z j z w t P w z j t n T + = = + to estimate ( ) | , j new P w z t , and ( ) | , new P z d t by equation (4) and (5). b) If the topics are within the training set, estimate ( ) ( ) ( ) ^ , | , , | , | , new old new z P u r d t P u r z t P z d t = , else if the sender and receivers are within the training set, estimate ( ) ^ , | , new P u r d t by topic-independent social network ( ) , | old P u r t . c) If there are new topics detected, update the model by incorporating the new topics. Inference, filtering, and prediction can be conducted based on this model. For the CTR algorithm, sender variable u or receiver variable r is fixed. For instance, if we are interested in ( ) | , , P r u d t , which is to answer a question of whom we should send the message d to during the time period t. The answer will be ( ) ( ) ( ) ( ) ( ) / ^ argmax | , , argmax | , , | , , | , | , , old old new t t t old new old new r t old t new old t new r z z z P r u d t P r u z t P z u d t P r u t P z u d t = + (9) where z / new old t t z represents the new topics emerging during the time period t. Another question is if we receive an email, who will be possibly the sender? ( ) ( ) ( ) ( ) ( ) / ^ argmax | , , argmax | , , | , , | , | , , old old old t t t old new old new u t old t new old t new u z z z P u r d t P u r z t P z r d t P u r t P z r d t = + (10) Eq. (9) and Eq. (10) integrate the PSN, content and temporal analysis. Social network models such as ERGM in Eq. (1) or the model in Sec. 3.2 can be applied to the ( ) , | , P u r d t terms. Figure 2 illustrates the CTR model and compares to the LDA, AT and ART models. In CTR, the observed variables not only include the words w in an email but also the sender u and the timestamp on each email d. 3.4 Predictive Algorithms For the sake of easier evaluation, we focus on prediction schemes in details. Specifically, we address the problem of predicting receivers and senders of emails as a classification problem, in which we train classifiers to predict the senders or receivers and other behavior patterns given the observed people's communication records. The trained classifier represents a function in the form of: : ( , ) f Comm t i t Y (11) where ( ) , Comm t i t is the observed communication record during the interval from time t-i to t, Y is a set of receivers or senders or other user behavior patterns to be discriminated, and the value of ( ) ( ) , f Comm t i t is the classifier prediction regarding which user behavior patterns gave rise to the observed communication records. The classifier is trained by providing the history of the communication records with known user behaviors. 3.4.1 Using Personal Social Network Model We aggregate all the communication records in the history of a given user, and build his/her personal social network. We choose those people with the highest communication frequency with this person as the prediction result. 3.4.2 Using LDA combined with PSN Model We use the LDA model and combine it with PSN to do the prediction, which is referred as LDA-PSN in the paper. Latent topics are detected by applying original LDA on the training set and LDA is used for inference in testing data without incorporating new topics when time passes by. The possible senders and receivers when new emails arrive, ( ) , | , new P u r d t is estimated as ( ) ( ) ( ) ^ , | , , | , | , new old new z P u r d t P u r z t P z d t = . People are ranked by this probability as the prediction results. 3.4.3 Using CTR Model People tend to send emails to different group of people under different topics during different time periods. This is the assumption we made for our predictive model based on CTR. LDA AT ART u u LDA AT ART u u LDA AT ART u u : observations A N D T u z w r CTR S Tm t : observations A N D T u z w r CTR S Tm t Figure 2. The graphical model for the CTR model comparing to LDA, AT and ART models, where u: sender, t: time, r: receivers, w: words, z: latent topics, S: social network, D: number of emails, N: number of words in one email, T: number of topics, Tm: size of the time sliding window, A: number of authors, , and are the parameters we want to estimate with the hyperparameters , , 483 Industry/Government Track Paper ( ) , | , new P u r d t is estimated by applying the CTR model discussed in section 3.3. The prediction results are people with highest scores calculated by equation (9) and (10). 3.4.4 Using an Adaptive CTR Model Both the personal social network and the CTR model ignore a key piece of information from communication records -- the dynamical nature of emails. Both personal social network and Topic-Community dynamically change and evolve. Only based on the training data which are collected in history will not get the optimal performance for the prediction task. Adaptive prediction by updating the model with newest user behavior information is necessary. We apply several strategies for the adaptive prediction. The first strategy is aggregative updating the model by adding new user behavior information including the senders and receivers into the model. Then the model becomes: ( ) ( ) ( ) ( ) ( ) 1 / ^ , | , , | , | , , | | , i t t i old K i k old k i old t i k z z P u r d t P u r z t P z d t P u r t P z d t = = + (12) where K is the number of old topics. Here, we always use the data from old t , including 0 t to 1 i t to predict the user behavior during i t . In the second strategy, we assume the correlation between current data and the previous data decays over time. The more recent data are more important. Thus, a sliding window of size n is used to choose the data for building the prediction model, in which the prediction is only dependent on the recent data, with the influence of old data ignored. Here in equation (12), old t consists of i n t to 1 i t . 3.5 CommunityNet Model We then build a CommunityNet model based on the CTR algorithm. The CommunityNet model, which refers to the personal Topic-Community Network, draws upon the strengths of the topic model and the social network as well as the dynamic model, using a topic-based representation to model the content of the document, the interests of the users, the correlation of the users and the receivers and all these relationship changing over time. For prediction, CommunityNet incorporates the adaptive CTR model as described in Section 3.4.4. COMMUNITY ANALYSIS The first part of our analysis focuses on identifying clusters of topics, and the senders and receivers who participated in those topics. First, we analyze the topics detected from the Enron Corpus. Then, we study the topic-community patterns. 4.1 Topic Analysis In the experiment, we applied Bayesian model selection [24] to choose the number of topics. In the Enron intra-organization emails, there are 26,178 word-terms involved after we apply stop-words removal and stemming, We computed ( ) | P w T for T values of 30, 50, 70, 100, 110, 150 topics and chose T = 100 with the maximum value of ( ) ( ) log | P w T for the experiment. 4.1.1 Topic Distribution After topic clustering based on words, for each document, we have P(z|d), which indicates how likely each document belongs to each topic. By summing up this probability for all the documents, we get the topic distribution of how likely each topic occurs in this corpus. We define this summed likelihood as "Popularity" of the topic in the dataset. From this topic distribution, we can see that some topics are hot - people frequently communicate with each other about them, while some others are cold, with only few emails related to them. Table 1 illustrates the top 5 topics in Enron corpus. We can see that most of them are talking about regular issues in the company like meeting, deal, and document. Table 2 illustrates the bottom 5 topics in Enron corpus. Most of them are specific and sensitive topics, like "Stock" or "Market". People may feel less comfortable to talk about them broadly. Table 1. Hot Topics meeting deal Petroleum Texas document meeting plan conference balance presentation discussion deal desk book bill group explore Petroleum research dear photo Enron station Houston Texas Enron north America street letter draft attach comment review mark Table 2. Cold Topics Trade stock network Project Market trade London bank name Mexico conserve Stock earn company share price new network world user save secure system Court state India server project govern call market week trade description respond 4.1.2 Topic Trend Analysis To sense the trend of the topics over time, we calculate the topic popularity for year 2000 and 2001, and calculate the correlation coefficients of these two series. For some topics, the trends over years are similar. Figure 3(a) illustrates the trends for two topics which have largest correlation coefficients between two years. Topic 45, which is talking about a schedule issue, reaches a peak during June to September. For topic 19, it is talking about a meeting issue. The trend repeats year to year. Figure 3(b) illustrates the trend of Topic "California Power" over 2000 to 2001. We can see that it reaches a peak from the end of year 2000 to the beginning of year 2001. From the timeline of Enron [29], we found that "California Energy Crisis" occurred at exactly this time period. Among the key people related to this topic, Jeff Dasovich was an Enron government relations executive. His boss, James Steffes was Vice President of Government Affairs. Richard Schapiro was Vice President of Regulatory Affairs. Richard Sanders was Vice President and Assistant General Counsel. Steven Kean was Executive Vice President and Chief of Staff. Vincent Kaminski was a Ph.D. economist and Head of Research for Enron Corp. Mary Han was a lawyer at Enron's West Coast trading hub. From the timeline, we found all these people except Vince were very active in this event. We will further analyze their roles in Section 5. 484 Industry/Government Track Paper Topic Trend Comparison 0 0.005 0.01 0.015 0.02 0.025 0.03 Jan Mar May Jul Sep Nov Po p u l a r i t y Topic45(y2000) Topic45(y2001) Topic19(y2000) Topic19(y2001) (a) Trends of two yearly repeating events. Topic Analysis for Topic 61 0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 Jan-00 Apr-00 Jul-00 Oct-00 Jan-01 Apr-01 Jul-01 Oct-01 P opul ar i t y Keywords with ( ) | P w z power 0.089361 California 0.088160 electrical 0.087345 price 0.055940 energy 0.048817 generator 0.035345 market 0.033314 until 0.030681 Key people with ( ) | P u z Jeff_Dasovich 0.249863 James_Steffes 0.139212 Richard_Shapiro 0.096179 Mary_Hain 0.078131 Richard_Sanders 0.052866 Steven_Kean 0.044745 Vince_Kaminski 0.035953 (b) The trend of "California Power" and most related keywords and people. Figure 3. Topic trends 4.2 Predicting Community Patterns We assume that, people communicate with certain people only under certain few topics. People in the same community under a topic would share the information. Thus, if there is something new about one topic, people in that topic-community will most likely get the information and propagate it to others in the community. Finally, many people in the community will get the information. To evaluate our assumption and answer the question of who will be possibly involved in an observed email, we collect the ground truth about who are the senders and receivers for the emails and use the CTR algorithm to infer ( ) , | , j new P u r z t by ( ) , | , j old P u r z t . We partitioned the data into training set and testing set. We tried two strategies for this experiment. First is to randomly partition the data into a training set with 8465 messages and a testing set with 8408 messages. Prediction accuracy is calculated by comparing the inference results and the ground truth (i.e., receiver-sender pair of that email). We found that 96.8446% people stick in the old topics they are familiar with. The second strategy is to partition data by time: emails before 1/31/2000 as the training data (8011) and after that as the testing data (8862). We found 89.2757% of the people keep their old topics. Both results are quite promising. It is found that people really stick in old topics they are familiar with. INDIVIDUAL ANALYSIS In this section, we evaluate the performance of CommunityNet. First, we show how people's roles in an event can be inferred by CommunityNet. Then, we show the predicting capability of the proposed model in experiments. 5.1 Role Discovery People with specific roles at company hierarchy behave specifically on specific topics. Here we show it is possible to infer people's roles by using CommunityNet. In Section 4.1.2, we show there are some key people involved in "California Energy Crisis". In reality, Dasovich, Steffes, Schapiro, Sanders, and Kean, were in charge of government affairs. Their roles were to "solve the problem". Mary Hain was a lawyer during the worst of the crisis and attended meetings with key insiders. We calculated the correlation coefficients of the trends of these people and the overall trend of this topic. Jeff Dasovich got 0.7965, James Steffes got 0.6501, Mary Hain got 0.5994, Richard Shapiro got 0.5604, Steven Kean got 0.3585 (all among the 10 highest correlation scores among 154 people), and Richard Sanders got 0.2745 (ranked 19), while Vince Kaminski had correlation coefficient of -0.4617 (Figure 4). We can see that all the key people except Vince Kaminski have strong correlation with the overall trend of "California Energy Crisis". From their positions, we can see that all of them were sort of politicians while Vince Kaminski is a researcher. Thus, it is clear to see the difference of their roles in this topic. 0 0.1 0.2 0.3 0.4 0.5 Jan-00 May-00 Sep-00 Jan-01 May-01 Sep-01 Po p u l a r i t y Overall trend Jeff_Dasovich Vince_Kaminski Figure 4. Personal topic trend comparison on "California Power" 5.2 Predicting Receivers Here we want to address the problem of whether it is possible to infer who will possibly be the receivers by a person's own historic communication records and the content of the email-to -send. One possible application is to help people organize personal social capital. For instance, if a user has some information to send or a question to ask, CommunityNet can recommend the right persons to send the info or get the answer. We conduct experiments by partitioning the dataset into a training set with the emails from 1999 to 2000, and a testing set with the emails from 2001 to 2002. The testing set is further partitioned into sub-sets with emails from one month as a subset. With this, we have 15 testing sets. (We exclude the emails after March 2002 because the total number of emails after that is only 78.) One issue we want to mention is that the number of people from 1999 to 2000 is 138, while from 2001 to 2002 is 154. In this study, we test each email in the training set by using its content, 485 Industry/Government Track Paper sender, and time as prior information to predict the receiver, which is compared to the real receiver of that email. In Figure 5, we illustrate the prediction performance by comparing the CTR algorithm, PSN, and the aggregated LDA-PSN model. The result shows that CTR beats PSN by 10% on accuracy. The aggregated LDA-PSN model performs even worse than PSN, because of the inaccurate clustering results. The performance gain is 21%. Moreover, intuitively, personal contacts evolve over time. Models built at a specific time should have decreasing predicting capability over time. In this figure, we obtain strong evidence of this hypothesis by observing that the performance of these models monotonically decays. This also implies our models well match the practice. 0 0.2 0.4 0.6 0.8 1 Jan-01 Apr-01 Jul-01 Oct-01 Jan-02 A c cu r a cy by PSN by CTR by LDA-PSN (a) Accuracy based on the top 5 most likely people 0 0.2 0.4 0.6 0.8 1 Jan-01 Apr-01 Jul-01 Oct-01 Jan-02 A c cu r a cy by PSN by CTR by LDA-PSN (b) Accuracy based on the top 10 most likely people Figure 5. Prediction Accuracy comparisons. Accuracy is measured by testing whether the "real" receiver is among the prediction list of the top 5 or 10 most likely people 5.3 Inferring Senders We test whether it is possible to infer who will possibly be the senders given a person's CommunityNet and the content of the email. One possible application is to exclude spam emails or detect identification forgery. Figure 6 illustrates the prediction result, which also shows the prediction accuracy decays over time. 0 0.2 0.4 0.6 0.8 1 Jan-01 Mar-01 May-01 Jul-01 Sep-01 Nov-01 A ccu r a c y top5 top3 Figure 6. Predicting senders given receiver and content 5.4 Adaptive Prediction We observed the prediction performance decays over time from the results of 5.2 and 5.3, which reflects the changes of the nature of email streams. Here we apply adaptive prediction algorithms we mentioned in 3.4.3, in which we incrementally and adaptively estimate statistical parameters of the model by gradually forgetting out-of-state statistics. 0 0.2 0.4 0.6 0.8 1 Jan-01 Mar-01 May-01 Jul-01 Sep-01 Nov-01 Ac c u r a c y Adaptive CT R(T op 5) Adaptive CT R(T op 10) CT R(T op5) CT R(T op10) (a). Comparison between Adaptive CTR and CTR models 0 5 10 15 20 25 30 35 40 45 Jan-01 Mar-01 May-01 Jul-01 Sep-01 Nov-01 Adaptive CTR(aggregative) Adaptive CTR(6 months) CTR LDA-PSN PSN (b) Comparison of algorithms using Breese evaluation metrics Figure 7. Performance evaluation for adaptive prediction algorithm and overall comparison Figure 7 (a) illustrates the performance of the Adaptive CTR algorithm and compares it to the CTR algorithm. For the data far away from the training data, the improvement is more than 30%. And, if we compare it to the PSN and LDA-PSN algorithms, the performance gains are 58% and 75%, respectively. Evaluation by this accuracy metric tells us how related the top people ranked in the prediction results are. To understand the overall performance of the ranked prediction results, we apply the evaluation metric proposed by Breese [30], and illustrate the overall comparison in Figure 7(b). This metric is an aggregation of the accuracy measurements in various top-n retrievals in the ranked list. Among all predictive algorithms, adaptive CTR models perform best and PSN performs worst. In adaptive CTR models, estimating from recent data of six months beats aggregative updating the model from all the data from the history. COMMUNITYNET APPLICATIONS In this section, we show two application systems we built based on the CommunityNet. The first one is a visualization and query tool to demonstrate informal networks incorporation. The second one is a receiver recommendation tool which can be used in popular email systems. These demos can be accessed from http://nansen.ee.washington.edu/CommunityNet/. 6.1 Sensing Informal Networks 6.1.1 Personal Social Network Figure 8 illustrates the interface of a visualization and query system of CommunityNet. The distance of nodes represents the closeness (measured by the communication frequencies) of a person to the center person. Users can click on the node to link to 486 Industry/Government Track Paper the CommunityNet of another person. This system can show personal social networks, which includes all the people a user contacts with during a certain time period. For instance, Figure 8(a) illustrates the personal social network of Vice President John Arnold from January 1999 to December 2000. During this period, there were 22 people he sent emails to, regardless what they were talking about. An evolutionary personal social network is illustrated in Figure 8(b), in which we show people's personal social network changes over time. From Jan. 1999 to Dec. 2000, no new contact was added to John's PSN. However, people's relationship changed in 2000. A Personal Social Network Information Flow is illustrated in Figure 8(c), in which we show how the information flows through the network (here we illustrate the information in two levels.) (a) Personal Social Network of John Arnold (b1) Jan-`99 to Dec-`99 (b2) Jan-`00 to Jun-`00 (b3) Jul-`00 to Dec-`00 (b) Evolutionary Personal Social Network (c) Personal Social Network Information Flow with two-level personal social network of John Arnold Figure 8. Personal social networks of John Arnold 6.1.2 Personal Topic-Community Network Personal topic-community network can show whom this user will contact with under a certain topic. On retrieval, keywords are required for inferring the related topics. Figure 9 illustrates several personal topic-community networks for John Arnold. First, we type in "Christmas" as the keyword. CommunityNet infers it as "holiday celebration" and shows the four people John contacted with about this topic. About "Stock", we find John talked with five people on "Stock Market" and "Company Share" from Jan. 1999 to Dec. 2000. Personal Topic-Community network can be depicted by the system, too. Figure 9. Personal Topic-Community Networks when we type in "Christmas" and "Stock" 6.2 Personal Social Capital Management Receiver Recommendation Demo When a user has some questions, he/she may want to know whom to ask how to find an expert and who may tell him/her more details because of their close relationships. In our second demo, we show a CommunityNet application which addresses this problem. This tool can be incorporated with general email systems to help users organize their personal social capitals. First, after a user login a webmail system, he can type in content and/or subject then click on the &quot;Show Content Topic-Community&quot;. This tool shall recommend appropriate people to send this email to, based on the learned personal social network or personal topic-community . The distances of nodes represent the closeness of the people to the user. Users can click on the node to select an appropriate person to send email to. If the center node is clicked, then a sphere grows to represent his ties to a group of experts. Click on "Mail To", then the people in the sphere will be included in the sender list. In the examples in Figure 10, we log in as Jeff Dasovich. He can ask his closest friends whenever he has questions or wants to disseminate information. If he wants to inform or get informed on 487 Industry/Government Track Paper "Government" related topics, the system will suggest him to send emails to Steffes, Allen, Hain, or Scott. The topics are inferred by matching the terms from the Subject as well as the content of the email. He can also type in "Can you tell me the current stock price?" as the email content. This system will detect "Stock Market" as the most relevant topic. Based on Dasovich's CommunityNet, it shows three possible contacts. He then chooses appropriate contact(s). (a) Receiver recommendation for "Government" (b) Receiver recommendation for "Can you tell me the current stock price?" Figure 10.Receiver recommendation demo system CONCLUSIONS AND FUTURE WORK In this paper, we propose a new way to automatically model and predict human behavior of receiving and disseminating information. We establish personal CommunityNet profiles based on a novel Content-Time-Relation algorithm, which incorporates contact, content, and time information simultaneously from personal communication. CommunityNet can model and predict the community behavior as well as personal behavior. Many interesting results are explored, such as finding the most important employees in events, predicting senders and receivers of emails, etc. Our experiments show that this multi-modality algorithm performs better than both the social network-based predictions and the content-based predictions. Ongoing work includes studying the response time of each individual to emails from different people to further analyze user's behavior, and also incorporating nonparametric Bayesian methods such as hierarchical LDA with contact and time information. ACKNOWLEDGMENTS We would like to thank D. Blei, T. Griffiths, Yi Wu and anonymous reviewers for valuable discussions and comments. This work was supported by funds from NEC Labs America. REFERENCES [1] B. A. Nardi, S. Whittaker, and H. Schwarz. "It's not what you know, it's who you know: work in the information age," First Mon., 5, 2000. [2] D. Krackhardt, "Panel on Informal Networks within Formal Organizations," XXV Intl. Social Network Conf., Feb. 2005. [3] D. Krackhardt and M. Kilduff, &quot;Structure, culture and Simmelian ties in entrepreneurial firms,&quot; Social Networks, Vol. 24, 2002. [4] B. Nardi, S. Whittaker, E. Isaacs, M. Creech, J. Johnson, and J. Hainsworth, "ContactMap: Integrating Communication and Information Through Visualizing Personal Social Networks," Com. of the Association for Computing Machinery. April, 2002. [5] https://www.linkedin.com/home?trk=logo. [6] https://www.orkut.com/Login.aspx. [7] http://www.friendster.com/. [8] N. Lin, "Social Capital," Cambridge Univ. Press, 2001. [9] W. Cohen. http://www-2.cs.cmu.edu/~enron/. [10] S. Milgram. "The Small World Problem," Psychology Today, pp 60-67, May 1967. [11] M. Schwartz and D. Wood, "Discovering Shared Interests Among People Using Graph Analysis", Comm. ACM, v. 36, Aug. 1993. [12] H. Kautz, B. Selman, and M. Shah. "Referral Web: Combining social networks and collaborative filtering," Comm. ACM, March 1997. [13] G. W. Flake, S. Lawrence, C. Lee Giles, and F. M. Coetzee. "Self-organization and identification of Web communities," IEEE Computer, 35(3):6670, March 2002. [14] J. Tyler, D. Wilkinson, and B. A. Huberman. "Email as spectroscopy: Automated Discovery of Community Structure Within Organizations," Intl. Conf. on Communities and Technologies., 2003. [15] L. Page, S. Brin, R. Motwani and T. Winograd. "The PageRank Citation Ranking: Bringing Order to the Web," Stanford Digital Libraries Working Paper, 1998. [16] J. Kleinberg. "Authoritative sources in a hyperlinked environment," In Proc. 9th ACM-SIAM Symposium on Discrete Algorithms, 1998. [17] S. Wasserman, and P. E. Pattison, "Logit models and logistic regression for social networks: I. An introduction to Markov graphs and p*", Psychometrika, 61: 401 425, 1996. [18] T. A.B. Snijders. "Models for Longitudinal Network Data," Chapter 11 in Models and methods in social network analysis, New York: Cambridge University Press, 2004. [19] D. L.-Nowell and J. Kleinberg, "The Link Prediction Problem for Social Networks," In Proceedings of the 12th Intl. Conf. on Information and Knowledge Management, 2003. [20] J. Kubica, A. Moore, J. Schneider, and Y. Yang. "Stochastic Link and Group Detection," In Proceedings of the 2002 AAAI Conference. Edmonton, Alberta, 798-804, 2002. [21] M. Handcock and D. Hunter, "Curved Exponential Family Models for Networks," XXV Intl. Social Network Conf., Feb. 2005. [22] T. Hofmann, "Probabilistic Latent Semantic Analysis," Proc. of the Conf. on Uncertainty in Artificial Intelligence, 1999. [23] D. Blei, A. Ng, and M. Jordan, "Latent Dirichlet allocation," Journal of Machine Learning Research, 3:993-1022, January 2003. [24] T. Griffiths and M. Steyvers, "Finding Scientific Topics," Proc. of the National Academy of Sciences, 5228-5235, 2004. [25] M. R.-Zvi, T. Griffiths, M. Steyvers and P. Smyth, "The Author-Topic Model for Authors and Documents", Proc. of the Conference on Uncertainty in Artificial Intelligence volume 21, 2004. [26] A. McCallum, A. Corrada-Emmanuel, and X. Wang, "The Author-Recipient-Topic Model for Topic and Role Discovery in Social Networks: Experiments with Enron and Academic Email," Technical Report UM-CS-2004-096, 2004. [27] X. Song, B. L. Tseng, C.-Y. Lin, and M.-T. Sun, &quot;ExpertiseNet: Relational and Evolutionary Expert Modeling,&quot; 10th Intl. Conf. on User Modeling, Edinburgh, UK, July 24-30, 2005. [28] J. Allan, R. Papka, and V. Lavrenko. "On-line New Event Detection and Tracking," Proc. of 21st ACM SIGIR, pp.37-45, August 1998. [29] http://en.wikipedia.org/wiki/Timeline_of_the_Enron_scandal. [30] J. Breese, D. Heckerman, and C. Kadie. "Empirical analysis of predictive algorithms for collaborative filtering," Conf. on Uncertainty in Artificial Intelligence, Madison,WI, July 1998. 488 Industry/Government Track Paper

user behavior modeling;information dissemination;personal information management

139

Modeling behavioral design patterns of concurrent objects

Object-oriented software development practices are being rapidly adopted within increasingly complex systems, including reactive, real-time and concurrent system applications. While data modeling is performed very well under current object-oriented development practices, behavioral modeling necessary to capture critical information in real-time, reactive, and concurrent systems is often lacking. Addressing this deficiency, we offer an approach for modeling and analyzing concurrent object-oriented software designs through the use of behavioral design patterns, allowing us to map stereotyped UML objects to colored Petri net (CPN) representations in the form of reusable templates. The resulting CPNs are then used to model and analyze behavioral properties of the software architecture, applying the results of the analysis to the original software design.

Introduction Object-oriented software development practices are being rapidly adopted within increasingly complex systems, including reactive, real-time and concurrent system applications. In practice, though, object-oriented software design techniques are still predominantly focused on the creation of static class models. Dynamic architectural models capturing the overall behavioral properties of the software system are often constructed using ad hoc techniques with little consideration given to the resulting performance or reliability implications until the project reaches implementation. Efforts to analyze behavioral issues of these architectures occur through opportunistic rather than systematic approaches and are inherently cumbersome, unreliable, and unrepeatable. One means of improving the behavioral modeling capabilities of object-oriented architecture designs is to integrate formalisms with the object-oriented specifications. Using this technique, object-oriented design artifacts are captured in a format such as the Unified Modeling Language (UML) [1], which is intuitive to the software architect. The native object-oriented design is then augmented by integrating an underlying formal representation capable of providing the necessary analytical tools. The particular method used in this research [2] is to integrate colored Petri nets (CPNs) [3] with object-oriented architecture designs captured in terms of UML communication diagrams. Specifically, this paper will present a method to systematically translate a UML software architecture design into an underlying CPN model using a set of pre-defined CPN templates based on a set of object behavioral roles. These behavioral roles are based on the object structuring criteria found in the COMET method [4], but are not dependent on any given method and are applicable across application domains. This paper will also demonstrate some of the analytical benefits provided by constructing a CPN representation of the UML software architecture. After a survey of related research, Section 2 descries the concept of behavioral design pattern templates for modeling concurrent objects. Section 3 discusses how we construct an overall CPN model of the concurrent software architecture by interconnecting the individual behavioral design pattern templates. Section 4 describes the validation of the approach. 1.1 Related Research There are many existing works dealing with the use of Petri nets for describing software behavior. As they relate to this paper, the existing works can be broadly categorized into the modeling of software code and the modeling of software designs. In this research, the focus is on improving reliability of object-oriented software designs rather than delaying detection to the software code. In terms of object-oriented design, the related Petri net research can be categorized as new development methodologies [5-8]; object-oriented extensions to Petri nets [9-12]; and the integration of Petri nets with existing object-oriented methodologies [13-20]. Since one of the goals of this research effort is to provide a method that requires no additional tools or language constructs beyond those currently available for the UML and CPN definitions, this approach [2,21-25] falls into the last category of integrating Petri nets with existing methodologies. The main features that distinguish this approach from other related works are a focus on the concurrent software Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. ICSE'06, May 2028, 2006, Shanghai, China. Copyright 2006 ACM 1-59593-085-X/06/0005...$5.00. 202 architecture design and the use of consistent, reusable CPN templates to model the behavior of concurrent objects and their interactions. This paper also extends our more recent works [25] by specifically focusing on the behavioral design patterns of individual concurrent objects and applying these patterns to construct an underlying representation of the concurrent software design architecture. Modeling Behavioral Design Patterns To model concurrent object behavioral design patterns with CPNs, our approach starts with a concurrent software architecture model captured in UML. For the construction of this architecture model, we identify a set of behavioral design patterns used to categorize the objects along with a set of specification requirements necessary to correctly model the concurrent behavior with the underlying CPN model. Each of the identified behavioral design patterns then has a corresponding template, represented as a CPN segment, which is paired with the UML object and is instantiated to capture specific behavioral characteristics based on the object specifications. The following sections describe the object architecture definition along with the concept of behavioral pattern templates for modeling concurrent objects. Section 3 will then discuss how we construct an overall CPN model of the concurrent object architecture by connecting the individual behavioral pattern templates. 2.1 Concurrent Object Modeling Our approach uses a UML communication diagram to capture the concurrent software architecture. Depending on the desired level of modeling, this architecture model can be constructed for an entire software system or for one or more individual subsystems. This communication diagram contains a collection of concurrent (active) and passive objects along with the message communication that occurs between the objects. Using our approach, objects within the concurrent software architecture are organized using the notion of components and connectors. Under this paradigm, concurrent objects are treated as components that can be connected through passive message communication objects and entity objects. In keeping with the COMET object structuring criteria, each object is assigned a UML stereotype to indicate its behavioral design pattern. Objects are broadly divided into application objects, which perform the work, and connector objects, which provide the means of communicating between application objects. For application objects, we use six stereotyped behavioral design patterns as illustrated in Figure 1: interface, entity, coordinator, state-dependent, timer, and algorithm. Additionally, connector objects can take the roles of: queue, buffer, or buffer-with-response , corresponding to asynchronous, synchronous, and return messages. These patterns are not intended to be an exhaustive list, but rather are intended to represent sufficient variety to model concurrent systems across a wide range of domains while also allowing these patterns to be extended as necessary for future applications. The identification of stereotyped behavioral roles allows us to select a specific CPN template to model each object (further described in Section 3.2). These behavioral stereotypes are generic across applications, so we also capture specific application information using the following tagged values: Execution Type. Each object must be declared as either passive or concurrent and for concurrent objects, further specified to be asynchronous or periodic. IO Mapping. Input-output message pairings must be specified for each object Communication Type. Indicate whether message communication occurs through asynchronous or synchronous means. Activation Time. The period of activation must be specified for each periodic concurrent object. Processing Time. Estimated processing times for completing an execution cycle should be assigned to each object if timing is to be accounted for in performance analysis. Operation Type. Indicate whether operations on entity objects perform "reader" or "writer" functionality. Statechart. For each state-dependent object, a UML statechart is used to capture the state behavior for that object. A detailed discussion of how the statechart is translated into the CPN model is provided in Pettit and Gomaa [24]. Figure 1. Stereotype Hierarchy for Application Objects 2.2 Defining Behavioral Pattern Templates The basis for our approach to modeling concurrent object behavior lies in the notion of a behavioral design pattern (BDP) template, which represents concurrent objects according to their role along with associated message communication constructs. For each BDP template, we employ a self-contained CPN segment that, through its places, transitions, and tokens, models a given stereotyped behavioral pattern. Each template is generic in the sense that it provides us with the basic behavioral pattern and component connections for the stereotyped object but does not contain any application-specific information. The connections provided by each template are consistent across the set of templates and allow concurrent objects to be connected to passive objects (entities or message communication) in any order. We provide a BDP template for each object type identified in the previous section. Since each of these templates captures a generic behavioral design pattern, when a template is assigned to a specific object, we then augment that template with the information captured in the tagged values for the object. For the resulting CPN representation, this affects the color properties of the tokens (e.g. to represent specific messages) and the rules for processing tokens (e.g. to account for periodic processing or special algorithms). The following sections describe a subset of our behavioral templates for both concurrent object components and their connectors. application interface entity control algorithm coordinator timer state dependent application interface entity control algorithm coordinator timer state dependent application interface entity control algorithm coordinator timer state dependent application interface entity control algorithm coordinator timer state dependent 203 2.2.1 Asynchronous Interface Object Template Consider the case of an asynchronous, input-only interface object. The template for this behavioral design pattern is given in Figure 2. This template represents a concurrent object, that is, an object that executes its own thread of control concurrently with other objects in the software system. While this template models relatively simple behavior (wait for input; process input; wait for next input), it features characteristics found throughout the concurrent object templates. First, to model the thread of control within a concurrent object, a control token (CTRL) is assigned to each concurrent object. For this template, a control token is initially present in the Ready place. Thus, this template is initialized in a state whereby it is ready to receive an input at the ProcessInput transition. As an input arrives (and given that the control token is in the Ready place), ProcessInput is allowed to fire, simulating the processing of the external input and the behavior of the asynchronous input interface object. ProcessInput consumes both a token representing the external input as well as the control token representing the executable thread of control. An output arc from ProcessInput uses a function, processInput (Input_event) to generate the appropriate token representing an internal message passed to another object within the system. The exact behavior of the processInput function (as with any arc-inscription functions throughout the templates) is determined from the object specification when a template is instantiated for a specific object. Finally, to complete the behavioral pattern for this template, the control token is passed to the MessageSent place and eventually back to the Ready place, enabling the template to process the next input. 2.2.2 Periodic Algorithm Object Template The asynchronous interface template addressed asynchronous behavior for a concurrent object, where the object is activated on demand by the receipt of a message or an external stimulus (as in the case of the interface example). For periodic behavior, where an object is activated on a regular periodic interval, consider the template for a concurrent periodic algorithm object given in Figure 3. Algorithm objects are internal concurrent objects that encapsulate algorithms, which may be activated or deactivated on demand. In the case of the periodic algorithm object, once the algorithm is enabled, it awakens on its activation period, performs the desired algorithmic task, and then returns to a sleep state until the next activation period. Looking at the periodic algorithm template from Figure 3, you should notice that, like the previous concurrent object template, there is Ready place with a control token that indicates when the object is ready to start its next processing cycle and models the thread of execution. This is common across all concurrent object templates. To model the ability for an algorithm object to be enabled or disabled, the input interface to this template occurs through the Enable_Alg and Disable_Alg transitions. (Note that we maintain the use of transitions as the interface points for all concurrent objects.) Thus, in addition to the control token being present on the Ready place, an Enable token must also be present on the Alg_Enabled place in order for the Perform_Alg transition to be enabled and subsequently fired. The actual behavior performed by the algorithm is captured by decomposing the Perform_Alg transition. The resulting decomposition uses one or more place-transition paths to model the behavior performed within the algorithm. The information necessary to derive the CPN algorithm model may be contained in the UML class specification for the algorithm object or, for more complex algorithms, may be captured in supporting UML artifacts such as the activity diagram. Multiple algorithms may be encapsulated within the same algorithm object. In these cases, the enable/disable transitions, enabled place, and processing transition are repeated for each encapsulated algorithm. However, there will only ever be one control token and ready place in a single concurrent object as our approach does not allow for multi-threaded concurrent objects. Finally, to capture the periodic nature of this template, a Sleep place along with Wakeup and Timeout transitions have been added to the basic asynchronous object template. This place-transition pair will be common to all periodic templates. In this case, the periodic algorithm starts in the Sleep place rather than Ready. After the desired sleep time (indicating the activation period of the object) has elapsed, the Wakeup transition is enabled and, when fired, removes the CTRL token from the Sleep place and places it in the Ready place. This now enables the template to perform any enabled algorithms. If one or more algorithms are enabled, the template proceeds in the same manner as the previous asynchronous algorithm template. However, if no algorithms are enabled when the template wakes up, the Timeout transition will fire and return the Control token to the Sleep place and wait for the next period of activation. 2.2.3 Entity Object Template In contrast to concurrent objects, passive objects do not execute their own thread of control and must rely on operation calls from a concurrent object. Using our approach, the entity objects from Figure 1 are passive objects. The purpose of an entity object is to store persistent data. Entities provide operations to access and manipulate the data stored within the object. These operations provide the interface to the entity object. To account for the possibility of multiple concurrent objects accessing a single entity object, our approach stipulates that each operation be tagged as having "read" or "write" access and for the object to be tagged with "mutually exclusive" or "multiple-reader/single -writer" rules for access control. This allows us to apply the appropriate template with the desired mutual exclusion protection for the encapsulated object attributes. The behavioral design pattern template representing an entity object with mutually exclusive access is shown in Figure 4. In this template, attributes are modeled with a CPN place containing tokens representing the attribute values. The underlying functionality of each operation is captured in an "idmOperation" transition that can be further decomposed as necessary to implement more complex functions. When instantiated for a specific entity object, the "idm" tag is replaced with a specific identifier for each operation. Finally, the interface to each operation is provided by a pair of CPN places one place for the operation call and another for the return. Collectively, these places form the interface to the entity object. As opposed to concurrent objects, all passive objects and message connectors will use CPN places for their interface, allowing concurrent objects to be connected through their transition interfaces. Thus, for performing an operation call, a 204 concurrent object places its control token and any necessary parameter tokens on the calling place and then waits for the control token to be returned along with any additional operation results at the call return place. Recall that entity objects do not have their own thread of control, thus they become part of the calling object's thread of control for the duration of the operation call. 2.2.4 Message Communication Templates Finally, in addition to application object templates, our method also provides templates for connector objects representing message communication. These connectors may represent asynchronous or synchronous message communication between two concurrent objects. Figure 2. Asynchronous Input-Only Interface Object: (a) UML (2.0); (b) CPN Template Figure 3. Periodic Algorithm Template: (a) UML; (b) CPN Representation {Execution = async; IO = input Process Time = &lt;process time&gt; } asyncInput Interface &lt;&lt;interface&gt;&gt; external InputSource &lt;&lt;external input device&gt;&gt; inputEvent asyncMsg To internal connector object (a) (b) {Execution = async; IO = input Process Time = &lt;process time&gt; } asyncInput Interface &lt;&lt;interface&gt;&gt; external InputSource &lt;&lt;external input device&gt;&gt; inputEvent asyncMsg To internal connector object (a) (b) {Execution = periodic; Activation Time = &lt;sleep time&gt; Process Time = &lt;process time&gt; } periodic Algorithm Object &lt;&lt;algorithm&gt;&gt; enable (a) (b) periodic Algorithm Object {Execution = periodic; Activation Time = &lt;sleep time&gt; Process Time = &lt;process time&gt; } periodic Algorithm Object &lt;&lt;algorithm&gt;&gt; enable (a) (b) periodic Algorithm Object 205 Figure 4. Passive Entity Template: (a) UML; (b) CPN Representation Consider the message buffer template shown in Figure 5. Notice that, as with passive entity objects, the interface to connector objects always occurs through a place rather than a transition, thus allowing concurrent object interfaces to be linked with connector interfaces while still enforcing the Petri net connection rules of only allowing arcs to occur between transitions and places. The message buffer template models synchronous message communication between two concurrent objects. Thus, only one message may be passed through the buffer at a time and both the producer (sender) and consumer (receiver) are blocked until the message communication has completed. The behavior of synchronous message communication is modeled through this template by first having the producer wait until the buffer is free as indicated by the presence of a "free" token in the buffer. The producer then places a message token in the buffer and removes the free token, indicating that the buffer is in use. Conversely, the consumer waits for a message token to appear in the buffer. After retrieving the message token, the consumer sets the buffer once again to free and places a token in the "Return" place, indicating to the producer that the communication has completed. Asynchronous message connector templates continue to employ places for their interfaces. However, asynchronous message communication, which involves the potential for queuing of messages, is more involved than the simple synchronous message buffer and must therefore add a transition to handle this behavior. The corresponding template is shown in Figure 6. With asynchronous communication the sender is not blocked awaiting acknowledgement that the sent message has been received and a message queue is allowed to form for the object receiving the asynchronous messages. In this template, the ManageQueue transition is decomposed into a subnet that implements the FIFO placement and retrieval of messages in the queue [26]. To send an asynchronous message, a concurrent object places a message token on the Enqueue place. The subnet under ManageQueue would then add this message token to the tail of the queue. Another concurrent object receiving the asynchronous message would wait for a message token to be available in the Dequeue place (representing the head of the queue). It would then remove the message token from Dequeue and signal DequeueComplete in a similar manner to the operation calls previously described for entities. This signals the queue that a message token has been removed from the head of the queue and that the remaining messages need to be advanced. Figure 5. Synchronous Message Buffer Connector Template: (a) UML; (b) CPN Representation (a) (b) anActiveObject anotherActive Object anEntityObject read() write() &lt;&lt;entity&gt;&gt; {Access Control = mutually-exclusive} (a) (b) anActiveObject anotherActive Object anEntityObject read() write() &lt;&lt;entity&gt;&gt; {Access Control = mutually-exclusive} (a) (b) producer consumer data (a) (b) producer consumer data 206 Figure 6. Asynchronous Message Queue Template Constructing CPN Models from UML Up to this point, we have just discussed individual CPN templates being used to model behavioral design patterns of concurrent objects, passive entity objects, and message communication mechanisms. This section presents our method for constructing a CPN model of the concurrent software architecture by applying and interconnecting these templates. The basic construction process consists of the following steps: 1. Construct a concurrent software architecture model using a UML communication diagram to show all concurrent and passive objects participating in the (sub) system to be analyzed along with their message communication. 2. Begin constructing the CPN model by first developing a context-level CPN model showing the system as a single CPN substitution (hierarchically structured) transition and the external interfaces as CPN places. Using a series of hierarchically structured transitions allows us to work with the CPN representation at varying levels of abstraction, from a completely black-box view, a concurrent software architecture view (in the next step), or within an individual object as desired for the level of analysis being applied to the model. 3. Decompose the system transition of the CPN context-level model, populating an architecture-level model with the appropriate CPN templates representing the objects from the concurrent software architecture. 4. Elaborate each instance of CPN template to account for the specific behavioral properties of the object it models. 5. Connect the templates, forming a connected graph between concurrent object templates and passive entity objects or message communication mechanisms. To illustrate the application of this approach, consider a partial example from the well-known Cruise Control System [4]. This example was chosen for this paper as it requires little explanation for the UML model and allows us to focus on the use of behavioral design pattern templates and the CPN representations. Figure 7 provides a partial communication diagram of the Cruise Control System concurrent software architecture. To begin, focus on the input events being provided by the Cruise Control Lever. (We will return to the brake and engine inputs later in this section.) Cruise control lever events enter the system via a concurrent interface object that sends an asynchronous message to the state dependent control object to process the requests based on rules defined in a corresponding statechart. Based on the state of the CruiseControl object, commands are given to a concurrent periodic algorithm object enabling it to compare speed values from two passive entity objects and determine the correct throttle values, which are then passed on to the periodic output interface, ThrottleInterface. Given this concurrent software architecture, the second step in our process would construct the context-level CPN model shown in Figure 8. At this level, we see the system as a black-box represented as a single transition, "CruiseControlSystem". External input and output interfaces for the cruise control lever, brake, and engine devices are represented as places. The purpose of this context-level CPN model is to provide a central starting point for our modeling and analysis. By structuring the CPN model in this way, we can analyze the system as a black box, dealing only with external stimuli and observed results (corresponding to the tokens stored in these places) or we can use hierarchical decomposition to gain access to the individual object behavioral design pattern templates (and their detailed CPN implementation) by systematically decomposing the hierarchically structured transitions (indicated with the HS tag). In the third step, the CruiseControlSystem transition from the context-level model is decomposed into an architecture-level model populated with the appropriate CPN behavioral design pattern template for each of the cruise control objects. Given the architecture design from Figure 7 (and continuing to ignore AutoSensors for the moment), we would need to instantiate two interface templates, two entity templates, one state dependent control template, and one algorithm template. We would also need to use queue and buffer templates for the asynchronous and synchronous message communication respectively. Once the appropriate templates have been assigned to each object, the fourth step in the process is to elaborate each template to model a specific object. To illustrate, consider CruiseControlLeverInterface. This object is an asynchronous input-only interface that accepts events from the cruise control lever device and, based on the input event, generates the appropriate messages for the cruise control request queue. Applying the asynchronous input interface template from Figure 2, we arrive at the elaborated CPN segment for CruiseControlLeverInterface shown in Figure 9. To elaborate the template for the CruiseControlLeverInterface, the place and transition names from the basic template have been appended with the object ID (1) for the specific object. The control token for this model has also been set to the specific control token for the CruiseControlLeverInterface object (CTRL1) and the time region for the PostProcessing_1 transition has been set to "@+100" to reflect the Process Time tagged value. The CruiseControlLeverInterface CPN representation is then connected to the software architecture by establishing an input arc from the CruiseControlLeverDevice place, representing the external input from the device, and an output arc to the Enqueue place, modeling the asynchronous message communication identified in the UML software architecture. Token types (colors) are then specifically created to represent the incoming event and outgoing messages. Finally, the processInput1() function is elaborated to generate the appropriate asynchronous message based on an incoming lever event. This elaboration process is similar for all templates. (a) (b) anActiveObject anotherActive Object data (a) (b) anActiveObject anotherActive Object data 207 Figure 7. Partial Concurrent Software Architecture for Cruise Control Figure 8. CPN Context-Level Model for Cruise Control Figure 9. Asynchronous Input-Only Interface Template Applied to CruiseControlLeverInterface Once all templates have been elaborated, our fifth and final step connects the templates to form a connected graph of the concurrent software architecture. The entire CPN architecture model for cruise control is too large for inclusion in this paper. However, Figure 10 illustrates the component connections between the CruiseControlLeverInterface and the CruiseControl templates using an asynchronous message queue connector. As can be seen from this figure, the two concurrent object templates communicate via the queue connector by establishing arcs between the interface transitions of the concurrent objects and the interface places of the queue connector. This component connection method applies to the entire software architecture using our approach of allowing concurrent objects to be connected to either passive entity objects or to a message communication connector. To further illustrate the component-based approach used for constructing these CPN let us now consider expanding the model to include input from the brake and engine devices. In addition to the cruise control lever inputs, Figure 7 also shows brake and engine status messages arriving from the respective devices. These status messages are handled by the AutoSensors periodic interface object and are passed to CruiseControl via an asynchronous message through the same cruise control request queue already being used by CruiseControlLeverInterface. Using our component-based modeling approach, the AutoSensors object can be added to our CPN model by simply instantiating a CPN representation of the periodic input interface behavioral design pattern template using the specified characteristics for AutoSensors and then connecting it to the existing queue template. The resulting CPN model is given in Figure 11. The addition of AutoSensors also illustrates another capability of the interface template. Whereas the cruise control lever is an asynchronous device, providing interrupts to CruiseControlLeverInterface, the brake and engine devices are passive devices that must be polled for their status. In Figure 11, every time AutoSensors is activated, it retrieves the status token from the brake and engine device places. After checking the status, the token is immediately returned to the device places, modeling persistence of device status information that can be polled as necessary. The remainder of the AutoSensors template should be familiar, being constructed of the standard Ready and ProcessInput place-transition pair for interface object templates (Section 2.2.1) and the Sleep and Wakeup place-transition pair included for periodic objects (Section 2.2.2). As demonstrated in this section, the primary benefits of our component-based modeling approach are that connections can easily be added or modified as the architecture evolves or to provide rapid "what-if" modeling and analysis. select(), clear() entity :DesiredSpeed entity :CurrentSpeed ccCommand throttleValue throttleOutput to throttle read() read() cruiseControlLeverInput cruiseControlRequest {Execution = async; IO = input Process Time = 100ms } {Execution = async; Process Time = 200ms } {Execution = periodic; Activation Time = 100ms Process Time = 50ms } {Execution = periodic; IO = output Process Time = 20ms Activation Time = 100ms } cruiseControlRequest brakeStatus engineStatus {Execution = periodic; IO = input Activation Time = 100ms Process Time = 20ms } CruiseControl LeverInterface &lt;&lt;interface&gt;&gt; CruiseControl &lt;&lt;state dependent&gt;&gt; Speed Adjustment &lt;&lt;algorithm&gt;&gt; Throttle Interface &lt;&lt;interface&gt;&gt; AutoSensors &lt;&lt;interface&gt;&gt; CruiseControl LeverDevice EngineDevice BrakeDevice &lt;&lt;external input device&gt;&gt; &lt;&lt;external input device&gt;&gt; &lt;&lt;external input device&gt;&gt; select(), clear() entity :DesiredSpeed entity :CurrentSpeed ccCommand throttleValue throttleOutput to throttle read() read() cruiseControlLeverInput cruiseControlRequest {Execution = async; IO = input Process Time = 100ms } {Execution = async; Process Time = 200ms } {Execution = periodic; Activation Time = 100ms Process Time = 50ms } {Execution = periodic; IO = output Process Time = 20ms Activation Time = 100ms } cruiseControlRequest brakeStatus engineStatus {Execution = periodic; IO = input Activation Time = 100ms Process Time = 20ms } CruiseControl LeverInterface &lt;&lt;interface&gt;&gt; CruiseControl &lt;&lt;state dependent&gt;&gt; Speed Adjustment &lt;&lt;algorithm&gt;&gt; Throttle Interface &lt;&lt;interface&gt;&gt; AutoSensors &lt;&lt;interface&gt;&gt; CruiseControl LeverDevice EngineDevice BrakeDevice &lt;&lt;external input device&gt;&gt; &lt;&lt;external input device&gt;&gt; &lt;&lt;external input device&gt;&gt; 208 Figure 10 Connecting CruiseControlLeverInterface and CruiseControl via Asynchronous Communication Figure 11. Addition of AutoSensors to the CPN Architecture Furthermore, by maintaining the integrity between a CPN template and the object it represents, modeling and analysis results can readily be applied to the original UML software architecture model. Thus, while from a pure CPN perspective, our CPNs could be further optimized, we feel that it is of greater benefit to maintain a component-based architecture that closely represents the structure of our original UML design artifacts. Validation The validation of our approach was in three parts. First, there was the issue of whether our behavioral stereotypes and corresponding templates could be applied across domains and projects. This was demonstrated by successfully applying our process to two case studies, the cruise control system (a portion of which was shown in the previous sections) and the signal generator system [2]. Secondly, we performed validation to determine if the resulting CPN models provided a correct model of the concurrent software architecture. This was necessary to validate that our approach would result in an accurate representation of the original architecture and was by far the most tedious part of validation, as it required manual inspection and unit testing of each object and its corresponding CPN template representation for the two case studies. Finally, after determining that our template approach satisfied the modeling requirements for both case studies, we then sought to demonstrate the analytical capabilities gained from using CPNs to model concurrent software architectures. The behavioral analysis addresses both the functional behavior of the concurrent architecture as well as its performance, as described next. The detailed analytical results for both case studies are provided in [2]. 4.1 Validating Functional Behavior For functional analysis, the simulation capabilities of the DesignCPN tool are used to execute the model over a set of test cases. These test cases may be black-box tests in which we are only monitoring the context-level model in terms of input events and output results or they may be white-box tests in which we analyze one or more individual object representations. In our approach, black box test cases were derived from use cases while white box test cases were derived from object interactions, object specifications, and statecharts. In each of these cases, the 209 appropriate inputs for each test case were provided by placing tokens on the CPN places representing the external actors in the context model. The CPN model was then executed in the simulator and observed at the desired points to determine if the correct output was generated or if the correct logical paths were chosen. Again, consider the cruise control system. Figure 12 illustrates a black-box simulation in which the driver has selected "Accelerate" from the cruise control lever (with the engine being on and the brake being released). Figure 12(a) shows the state of the system before the simulation run and Figure 12(b) illustrates the results of accelerating, namely a value being sent to the throttle. This form of simulation may be applied to as low or as high of a level of abstraction as desired in order to gain visibility into the desired behavior of the architecture. For example, one could choose to simply conduct black box testing by placing input tokens on actor places, executing the simulation, and then observing the resulting token values on output actor places. Alternatively, if a more detailed investigation is desired, the engineer may navigate the CPN hierarchical construction and observe such characteristics as the behavior of state changes within a state dependent object's CPN representation. A detailed analysis of this state-dependent behavior is provided in [24]. Figure 12. Example Cruise Control Black-Box Simulation 4.2 Validating Performance In addition to simulation capabilities, the DesignCPN [27] tool used in this effort also has a very powerful performance tool [28] that can be employed to analyze performance aspects of the concurrent software architecture. This tool can be used to analyze such things as queue backlogs, system throughput, and end-to-end timing characteristics. As an example of the latter, we conducted a test to monitor the cruise control system response times to commands being input from the cruise control lever. To conduct this analysis, commands were issued to the cruise control system while the system was in a simulated state of operation with a speed of 60 miles per hour (100 kph). The performance tool was used to monitor changes in the throttle output and compare the time at an observed output change to the time the original command was issued. The results from this analysis are shown in Figure 13. From this figure, we can see that all cruise control commands complete in less than one second (1000ms) and most complete in less than 500ms. Detailed performance requirements were not provided for our cruise control case study. However, if this cruise control system was an actual production system, an engineer could compare the analysis results against documented performance requirements to determine if the system in fact satisfies the necessary performance criteria. By being able to conduct this form of analysis from the concurrent software design, an engineer can both improve the reliability of the software architecture at the design level and correct problems prior to implementation. Figure 13. Cruise Control End-to-End Timing Analysis Conclusions and Future Research The long-term goal of this research effort is to provide an automated means of translating a UML concurrent software architecture design into an underlying CPN representation that can then be used to conduct behavioral analysis with results communicated in terms of the original UML model. To date, we have developed a method for systematically translating a UML software architecture into a CPN representation. This method employs reusable templates that model the behavior of a set of objects according to their stereotyped behavioral roles. Each template provides a consistent interface that allows templates to be interconnected as components of a larger system, thus creating the overall CPN representation. The resulting CPN model enables the analysis of both the functional and performance behavior of the concurrently executing objects. As the CPN representation mirrors the structure of the concurrent software architecture, the results can be readily applied to the original UML model. Future research in this area will need to investigate approaches to facilitate the automated translation from a UML model into a CPN model that can be read by a tool such as DesignCPN. Additional research also needs to be conducted to investigate the scalability of this approach to larger systems, including distributed applications and providing behavioral templates for the COMET distributed components [4]. Finally, the use of state space analysis should be investigated further. Most of the analysis conducted with this research effort has focused on the use of simulations for functional analysis and on the performance tool for performance analysis. State space analysis Cruise Control End-To-End Timing Performance 0 200 400 600 800 1000 1200 0 5000 10000 15000 20000 25000 30000 Elapsed Time (ms) C o m m a nd C o m p l e t i on Ti m e ( m s ) (a) (b) 1`"BrakeOff" 1`"Accel" 1`"Engine On" 1 1 1 1 1 1 1`"BrakeOff" 1`50 1`"Engine On" 210 could also be used to further refine deadlock detection as well as to analyze system-wide state changes. References [1] J. Rumbaugh, I. Jacobson, and G. Booch, The Unified Modeling Language Reference Manual. 2 nd Edition. Addison-Wesley, 2005. [2] R. G. Pettit, Analyzing Dynamic Behavior of Concurrent Object-Oriented Software Designs, Ph.D., School of IT&E, George Mason University, 2003. [3] K. Jensen, Coloured Petri Nets: Basic Concepts, Analysis Methods, and Practical Use, vol. I-III. Berlin, Germany: Springer-Verlag, 1997. [4] H. Gomaa, Designing Concurrent, Distributed, and Real-Time Applications with UML, Addison-Wesley, 2000. [5] M. Baldassari, G. Bruno, and A. Castella, "PROTOB: an Object-Oriented CASE Tool for Modeling and Prototyping Distributed Systems," Software-Practice & Experience, v.21, pp. 823-44, 1991. [6] B. Mikolajczak and C. A. Sefranek, "Integrating Object Oriented Design with Concurrency Using Petri Nets," IEEE International Conference on Systems, Man and Cybernetics, Piscataway, NJ, USA, 2001. [7] R. Aihua, "An Integrated Development Environment for Concurrent Software Developing Based on Object Oriented Petri Nets," Fourth International Conference/Exhibition on High Performance Computing in the Asia-Pacific Region., Los Alamitos, CA, USA, 2000. [8] X. He and Y. Ding, "Object Orientation in Hierarchical Predicate Transition Nets," Concurrent Object-Oriented Programming and Petri Nets. Advances in Petri Nets, Berlin: Springer-Verlag, 2001, pp. 196-215. [9] O. Biberstein, D. Buchs, and N. Guelfi, "Object-Oriented Nets with Algebraic Specifications: The CO-OPN/2 Formalism," Concurrent Object-Oriented Programming and Petri Nets. Advances in Petri Nets, Berlin: Springer-Verlag , 2001, pp. 73-130. [10] S. Chachkov and D. Buchs, "From Formal Specifications to Ready-to-Use Software Components: The Concurrent Object Oriented Petri Net Approach," Second International Conference on Application of Concurrency to System Design, Los Alamitos, CA, USA, 2001. [11] A. Camurri, P. Franchi, and M. Vitale, "Extending High-Level Petri Nets for Object-Oriented Design," IEEE International Conference on Systems, Man and Cybernetics, New York, NY, USA, 1992. [12] J. E. Hong and D. H. Bae, "Software Modeling and Analysis Using a Hierarchical Object-Oriented Petri Net," Information Sciences, v.130, pp. 133-64, 2000. [13] D. Azzopardi and D. J. Holding, "Petri Nets and OMT for Modeling and Analysis of DEDS," Control Engineering Practices, v.5, pp. 1407-1415, 1997. [14] C. Lakos, "Object Oriented Modeling With Object Petri Nets," Concurrent Object-Oriented Programming and Petri Nets. Advances in Petri Nets, Berlin: Springer-Verlag, 2001, pp. 1-37. [15] C. Maier and D. Moldt, "Object Coloured Petri Nets- A Formal Technique for Object Oriented Modelling," Concurrent Object-Oriented Programming and Petri Nets. Advances in Petri Nets, Berlin: Springer-Verlag, 2001, pp. 406-27. [16] J. A. Saldhana, S. M. Shatz, and H. Zhaoxia, "Formalization of Object Behavior and Interactions from UML Models," International Journal of Software Engineering & Knowledge Engineering, v.11, pp. 643-73, 2001. [17] L. Baresi and M. Pezze, "On Formalizing UML with High-Level Petri Nets," Concurrent Object-Oriented Programming and Petri Nets. Advances in Petri Nets, Berlin: Springer-Verlag, 2001, pp. 276-304. [18] K. M. Hansen, "Towards a Coloured Petri Net Profile for the Unified Modeling" Centre for Object Technology, Aarhus, Denmark, Technical Report COT/2-52-V0.1 (DRAFT), 2001. [19] J. B. Jrgensen, "Coloured Petri Nets in UML-Based Software Development - Designing Middleware for Pervasive Healthcare," CPN '02, Aarhus, Denmark, 2002. [20] B. Bordbar, L. Giacomini, and D. J. Holding, "UML and Petri Nets for Design and Analysis of Distributed Systems," International Conference on Control Applications, Anchorage, Alaska, USA, 2000. [21] R. G. Pettit and H. Gomaa, "Integrating Petri Nets with Design Methods for Concurrent and Real-Time Systems," Real Time Applications Workshop, Montreal, Canada, 1996. [22] R. G. Pettit, "Modeling Object-Oriented Behavior Using Petri Nets," OOPSLA Workshop on Behavioral Specification, 1999. [23] R. G. Pettit and H. Gomaa, "Validation of Dynamic Behavior in UML Using Colored Petri Nets," UML 2000, York, England, 2000. [24] R. G. Pettit and H. Gomaa, "Modeling State-Dependent Objects Using Colored Petri Nets," CPN 01 Workshop on Modeling of Objects, Components, and Agents, Aarhus, Denmark, 2001. [25] R.G. Pettit and H. Gomaa, "Modeling Behavioral Patterns of Concurrent Software Architectures Using Petri Nets." Working IEEE/IFIP Conference on Software Architectures, Oslo, Norway, 2004. [26] R. David and H. Alla, "Petri Nets for Modeling of Dynamic Systems: A Survey." Automatica v.30(2). Pp. 175-202. 1994. [27] K. Jensen, "DesignCPN," 4.0 ed. Aarhus, Denmark: University of Aarhus, 1999. [28] B. Lindstrom and L. Wells, "Design/CPN Performance Tool Manual," University of Aarhus, Aarhus, Denmark September 1999. 211

Software Architecture;Behavioral Design Patterns;Colored Petri Nets;COMET

14

A New Approach to Intranet Search Based on Information Extraction

This paper is concerned with `intranet search'. By intranet search, we mean searching for information on an intranet within an organization. We have found that search needs on an intranet can be categorized into types, through an analysis of survey results and an analysis of search log data. The types include searching for definitions, persons, experts, and homepages. Traditional information retrieval only focuses on search of relevant documents, but not on search of special types of information. We propose a new approach to intranet search in which we search for information in each of the special types, in addition to the traditional relevance search. Information extraction technologies can play key roles in such kind of `search by type' approach, because we must first extract from the documents the necessary information in each type. We have developed an intranet search system called `Information Desk'. In the system, we try to address the most important types of search first - finding term definitions, homepages of groups or topics, employees' personal information and experts on topics. For each type of search, we use information extraction technologies to extract, fuse, and summarize information in advance. The system is in operation on the intranet of Microsoft and receives accesses from about 500 employees per month. Feedbacks from users and system logs show that users consider the approach useful and the system can really help people to find information. This paper describes the architecture, features, component technologies, and evaluation results of the system.

INTRODUCTION Internet search has made significant progress in recent years. In contrast, intranet search does not seem to be so successful. The IDC white paper entitled "The high cost of not finding information" [13] reports that information workers spend from 15% to 35% of their work time on searching for information and 40% of information workers complain that they cannot find the information they need to do their jobs on their company intranets. Many commercial systems [35, 36, 37, 38, 39] have been developed for intranet search. However, most of them view intranet search as a problem of conventional relevance search. In relevance search, when a user types a query, the system returns a list of ranked documents with the most relevant documents on the top. Relevance search can only serve average needs well. It cannot, however, help users to find information in a specific type, e.g., definitions of a term and experts on a topic. The characteristic of intranet search does not seem to be sufficiently leveraged in the commercial systems. In this paper, we try to address intranet search in a novel approach. We assume that the needs of information access on intranets can be categorized into searches for information in different types. An analysis on search log data on the intranet of Microsoft and an analysis on the results of a survey conducted at Microsoft have verified the correctness of the assumption. Our proposal then is to take a strategy of `divide-and-conquer'. We first figure out the most important types of search, e.g., definition search, expert search. For each type, we employ information extraction technologies to extract, fuse, and summarize search results in advance. Finally, we combine all the types of searches together, including the traditional relevance Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. CIKM'05, October 31-November 5, 2005, Bremen, Germany. Copyright 2005 ACM 1-59593-140-6/05/0010...$5.00. 460 search, in a unified system. In this paper, we refer to the approach as `search by type'. Search by type can also be viewed as a simplified version of Question Answering, adapted to intranet. The advantage of the new search approach lies in that it can help people find the types of information which relevance search cannot easily find. The approach is particularly reasonable on intranets, because in such space users are information workers and search needs are business oriented. We have developed a system based on the approach, which is called `Information Desk'. Information Desk can help users to find term definitions, homepages of groups or topics, employees' personal information and experts on topics, on their company intranets. The system has been put into practical use since November 24 th , 2004. Each month, about 500 Microsoft employees make access to the system. Both the results of an analysis on a survey and the results of an analysis on system log show that the features of definition search and homepage search are really helpful. The results also show that search by type is necessary at enterprise. RELATED WORK The needs on search on intranets are huge. It is estimated that intranets at enterprises have tens or even hundreds of times larger data collections (both structured and unstructured) than internet. As explained above, however, many users are not satisfied with the current intranet search systems. How to help people access information on intranet is a big challenge in information retrieval. Much effort has been made recently on solutions both in industry and in academia. Many commercial systems [35, 36, 37, 38, 39] dedicated to intranet search have been developed. Most of the systems view intranet search as a problem of conventional relevance search. In the research community, ground designs, fundamental approaches, and evaluation methodologies on intranet search have been proposed. Hawking et al [17] made ten suggestions on how to conduct high quality intranet search. Fagin et al [12] made a comparison between internet search and intranet search. Recently, Hawking [16] conducted a survey on previous work and made an analysis on the intranet search problem. Seven open problems on intranet search were raised in their paper. Chen et al [3] developed a system named `Cha-Cha', which can organize intranet search results in a novel way such that the underlying structure of the intranet is reflected. Fagin et al [12] proposed a new ranking method for intranet search, which combine various ranking heuristics. Mattox et al [25] and Craswell et al [7] addressed the issue of expert finding on a company intranet. They developed methods that can automatically identify experts in an area using documents on the intranet. Stenmark [30] proposed a method for analyzing and evaluating intranet search tools. 2.2 Question Answering Question Answering (QA) particularly that in TREC (http://trec.nist.gov/) is an application in which users type questions in natural language and the system returns short and usually single answers to the questions. When the answer is a personal name, a time expression, or a place name, the QA task is called `Factoid QA'. Many QA systems have been developed, [2, 4, 18, 20, 22, 27]. Factoid QA usually consists of the following steps: question type identification, question expansion, passage retrieval, answer ranking, and answer creation. TREC also has a task of `Definitional QA'. In the task, "what is &lt;term&gt;" and "who is &lt;person&gt;" questions are answered in a single combined text [1, 11, 15, 33, 34]. A typical system consists of question type identification, document retrieval, key sentence matching, kernel fact finding, kernel fact ranking, and answer generation. OUR APPROACH TO INTRANET SEARCH Search is nothing but collecting information based on users' information access requests. If we can correctly gather information on the basis of users' requests, then the problem is solved. Current intranet search is not designed along this direction. Relevance search can help create a list of ranked documents that serve only average needs well. The limitation of this approach is clear. That is, it cannot help users to find information of a specific type, e.g., definitions of a term. On the other hand, Question Answering (QA) is an ideal form for information access. When a user inputs a natural language question or a query (a combination of keywords) as a description of his search need, it is ideal to have the machine `understand' the input and return only the necessary information based on the request. However, there are still lots of research work to do before putting QA into practical uses. In short term, we need consider adopting a different approach. One question arises here: can we take a hybrid approach? Specifically, on one hand, we adopt the traditional approach for search, and on the other hand, we realize some of the most frequently asked types of search with QA. Finally, we integrate them in a single system. For the QA part, we can employ information extraction technologies to extract, fuse, and summarize the results in advance. This is exactly the proposal we make to intranet search. Can we categorize users' search needs easily? We have found that we can create a hierarchy of search needs for intranet search, as will be explained in section 4. On intranets, users are information workers and their motivations for conducting search are business oriented. We think, therefore, that our approach may be relatively easily realized on intranets first. (There is no reason why we cannot apply the same approach to the internet, however.) To verify the correctness of the proposal, we have developed a system and made it available internally at Microsoft. The system called Information Desk is in operation on the intranet of Microsoft and receives accesses from about 500 employees per month. At Information Desk, we try to solve the most important types of search first - find term definitions, homepages of groups or topics, experts on topics, and employees' personal information. We are 461 also trying to increase the number of search types, and integrate them with the conventional relevance search. We will explain the working of Information Desk in section 5. ANALYSIS OF SEARCH NEEDS In this section, we describe our analyses on intranet search needs using search query logs and survey results. 4.1 Categorization of Search Needs In order to understand the underlying needs of search queries, we would need to ask the users about their search intentions. Obviously, this is not feasible. We conducted an analysis by using query log data. Here query log data means the records on queries typed by users, and documents clicked by the users after sending the queries. Our work was inspirited by those of Rose and Levinson [28]. In their work, they categorized the search needs of users on internet by analyzing search query logs. We tried to understand users' search needs on intranet by identifying and organizing a manageable number of categories of the needs. The categories encompass the majority of actual requests users may have when conducting search on an intranet. We used a sample of queries from the search engine of the intranet of Microsoft. First, we brainstormed a number of categories, based on our own experiences and previous work. Then, we modified the categories, including adding, deleting, and merging categories, by assigning queries to the categories. Given a query, we used the following information to deduce the underlying search need: the query itself the documents returned by the search engine the documents clicked on by the user For example, if a user typed a keyword of `.net' and clicked a homepage of .net, then we judged that the user was looking for a homepage of .net. As we repeated the process, we gradually reached the conclusion that search needs on intranet can be categorized as a hierarchical structure shown in Figure 1. In fact, the top level of the hierarchy resembles that in the taxonomy proposed by Rose and Levinson for internet [28]. However, the second level differs. On intranet, users' search needs are less diverse than those on internet, because the users are information workers and their motivations of conducting search are business oriented. There is a special need called `tell me about' here. It is similar to the traditional relevance search. Many search needs are by nature difficult to be categorized, for example, "I want to find documents related to both .net and SQL Server". We can put them into the category. We think that the search needs are not Microsoft specific; one can image that similar needs exist in other companies as well. Informational When (time) Where (place) Why (reason) What is (definition) Who knows about (expert) Who is (person) How to (manual) Tell me about (relevance) Navigational Person Product Technology Services Group Transactional Figure 1. Categories of search needs 4.2 Analysis on Search Needs by Query Log We have randomly selected 200 unique queries and tried to assign the queries to the categories of search needs described above. Table 1 shows the distribution. We have also picked up the top 350 frequently submitted queries and assigned them to the categories. Table 2 shows the distribution. (There is no result for `why', `what is', and `who knows about', because it is nearly impossible to guess users' search intensions by only looking at query logs.) For random queries, informational needs are dominating. For high frequency queries, navigational needs are dominating. The most important types for random queries are relevance search, personal information search, and manual search. The most important types for high frequency queries are home page search and relevance search. 4.3 Analysis on Search Needs by Survey We can use query log data to analyze users' search needs, as described above. However, there are two shortcomings in the approach. First, sometimes it is difficult to guess the search intensions of users by only looking at query logs. This is especially true for the categories of `why' and `what'. Usually it is hard to distinguish them from `relevance search'. Second, query log data cannot reveal users' potential search needs. For example, many employees report that they have needs of searching for experts on specific topics. However, it is difficult to find expert searches from query log at a conventional search engine, because users understand that such search is not supported and they do not conduct the search. To alleviate the negative effect, we have conducted another analysis through a survey. Although a survey also has limitation (i.e., it only asks people to answer pre-defined questions and thus can be biased), it can help to understand the problem from a different perspective. 462 Table 1. Distribution of search needs for random queries Category of Search Needs Percentage When 0.02 Where 0.02 Why NA What is NA Who knows about NA Who is 0.23 How to 0.105 Tell me about 0.46 Informational total 0.835 Groups 0.03 Persons 0.005 Products 0.02 Technologies 0.02 Services 0.06 Navigational total 0.135 Transactional 0.025 Other 0.005 Table 2. Distribution of search needs for high frequency queries Category of Search Needs Relative Prevalence When 0.0057 Where 0.0143 Why NA What is NA Who knows about NA Who is 0.0314 How to 0.0429 Tell me about 0.2143 Informational total 0.3086 Groups 0.0571 Persons 0.0057 Products 0.26 Technologies 0.0829 Services 0.2371 Navigational total 0.6428 Transactional 0.0086 Other 0.04 I have experiences of conducting search at Microsoft intranet to look for the web sites (or homepages) of (multiple choice) technologies 74 % products 74 % services 68 % projects 68 % groups 60 % persons 42 % none of the above 11 % I have experiences of conducting search at Microsoft intranet in which the needs can be translated into questions like? (multiple choice) `what is' - e.g., &quot;what is blaster&quot; 77 % `how to' - &quot;how to submit expense report&quot; 54 % `where' - e.g., &quot;where is the company store&quot; 51 % `who knows about' - e.g., &quot;who knows about data mining&quot; 51 % `who is' - e.g., &quot;who is Rick Rashid&quot; 45 % `when' - e.g., &quot;when is TechFest'05 &quot; 42 % `why' - e.g., &quot;why do Windows NT device drivers contain trusted code&quot; 28 % none of the above 14 % I have experiences of conducting search at Microsoft intranet in order to (multiple choice) download a software, a document, or a picture. E.g., &quot;getting MSN logo&quot; 71 % make use of a service. E.g., &quot;getting a serial number of Windows&quot; 53 % none of the above 18 % Figure 2. Survey results on search needs In the survey, we have asked questions regarding to search needs at enterprise. 35 Microsoft employees have taken part in the survey. Figure 2 shows the questions and the corresponding results. We see from the answers that definition search, manual search, expert finding, personal information search, and time schedule search are requested by the users. Homepage finding on technologies and products are important as well. Search for a download site is also a common request. 463 Who is Definition of Longhorn Where is homepage of Who knows about Longhorn is the codename for the next release of the Windows operating system, planned for release in FY 2005. Longhorn will further Microsoft's long term vision for ... http://url1 Longhorn is a platform that enables incredible user experiences that are unlike anything possible with OS releases to date . This session describes our approach and philosophy that... http://url2 Longhorn is the platform in which significant improvements in the overall manageability of the system by providing the necessary infrastructure to enable standardized configuration/change management, structured eventing and monitoring, and a unified software distribution mechanism will be made. In order to achieve this management with each Longhorn... http://url3 Longhorn is the evolution of the .NET Framework on the client and the biggest investment that Microsoft has made in the Windows client development platform in years. Longhorn is the platform for smart , connected... http://url4 Longhorn is the platform for smart, connected applications , combining the best features of the Web, such as ease of deployment and rich content with the power of the Win32 development platform, enabling developers to build a new breed of applications that take real advantage of the connectivity , storage, and graphical capabilities of the modern personal computer . http://url5 What is Longhorn Go What is Who is Where is homepage of Who knows about What is Who is Homepages of Office Office Go What is Who is Where is homepage of Who knows about Who knows about Office Portal Site This is the internal site for Office http://url1 Where is homepage of Office Site (external) Microsoft.com site offering information on the various Office products. Links include FAQs , downloads, support, and more. http:/url2 Office New Office Site http://url3 Office Office http://url4 Where is homepage of What is Who is People Associated with Data mining Who knows about Jamie MacLennan DEVELOPMENT LEAD US-SQL Data Warehouse +1 (425) XXXXXXX XXXXXX Associated documents(4): is author of document entitled Data Mining Tutorial http ://url1 is author of document entitled Solving Business Problems Using Data Mining http:// url2 Jim Gray DISTINGUISHED ENGINEER US-WAT MSR San Francisco +XXXXXXXXXXX Associated documents(2): Data Mining Go What is Who is Where is homepage of Who knows about is author of document entitled Mainlining Data Mining http ://url3 is author of document entitled Data Mining the SDSS SkyServer Database http:// url4 Where is homepage of Who knows about What is Who is Bill Gates CHRMN & CHIEF SFTWR ARCHITECT US-Executive-Chairman +1 (425) XXXXXXX XXXXXX Documents of Bill Gates(118) My advice to students: Education counts http://url1 Evento NET Reviewers Seattle 7/8 Novembro http://url2 A Vision for Life Long Learning Year 2020 http://url3 Bill Gates answers most frequently asked questions . http://url4 &gt;&gt;more Top 10 terms appearing in documents of Bill Gates Term 1 (984.4443) Term 2 (816.4247) Term 3 (595.0771) Term 4 (578.5604) Term 5 (565.7299) Term 6 (435.5366) Term 7 (412.4467) Term 8 (385.446) Term 9 (346.5993) Term 10 (345.3285) Bill Gates Go What is Who is Where is homepage of Who knows about Figure 3: Information Desk system INFORMATION DESK Currently Information Desk provides four types of search. The four types are: 1. `what is' search of definitions and acronyms. Given a term, it returns a list of definitions of the term. Given an acronym, it returns a list of possible expansions of the acronym. 2. `who is' search of employees' personal information. Given the name of a person, it returns his/her profile information, authored documents and associated key terms. 3. `where is homepage of' search of homepages. Given the name of a group, a product, or a technology, it returns a list of its related home pages. 4. `who knows about' search of experts. Given a term on a technology or a product, it returns a list of persons who might be experts on the technology or the product. Crawler & Extractor Web Server Information Desk MS Web term definition acronym what is person document key term who is term person document who knows about term homepage Where is homepage of Figure 4. Workflow of Information Desk There are check boxes on the UI, and each represents one search type. In search, users can designate search types by checking the corresponding boxes and then submit queries. By default, all the boxes are checked. For example, when users type `longhorn' with the `what is' box checked, they get a list of definitions of `Longhorn' (the first snapshot in figure 3). Users can also search for homepages (team web sites) related to `Office', using the `where is homepage' feature (the second snapshot in figure 3). Users can search for experts on, for example, `data mining' by asking `who knows about data mining' (the third snapshot in figure 3). Users can also get a list of documents that are automatically identified as being authored by `Bill Gates', for example, with the `who is' feature (the last snapshot in figure 3). The top ten key terms found in his documents are also given. Links to the original documents, from which the information has been extracted, are also available on the search result UIs. 5.2 Technologies 5.2.1 Architecture Information Desk makes use of information extraction technologies to support the search by type feaatures. The technologies include automatically extracting document metadata and domain specific knowledge from a web site using information extraction technologies. The domain specific knowledge includes definition, acronym, and expert. The document metadata includes title, author, key term, homepage. Documents are in the form of Word, PowerPoint, or HTML. Information Desk stores all the data in Microsoft SQL Server and provides search using web 464 $ 9 ! &quot; ! 4 9 % ) ! .$ , ! ! &quot; T 2 T ( T T T 2 T T U T T 3 T , M N ! 4 2 -V% KV % L *)7+ K L -V% $ ! K L ! : ! K L &gt; &gt; 7 ) ! ' &&' ' &lt;&lt;1 % *7 7' 77 7& F7 F9 ))+ 2 # ! &quot; ! ! ! -V% * )F+ $ M N KM NL M N M N &gt; 52 &quot; A &quot; 2 B KA&quot;2BL6 B B C $ ! ! ! $# &quot;$ % ( ; ''' ; ''' B B ' F;& ' 7F; ! D $ & K B B L ! 2 B P % *F)+ 3 ! D , B ' &lt;1& ' &lt;== 8 4 @ # ;%%; @ # J ! ! A7 ( ) 8 4 @ # ;%%; @ # J ! ! A7 ( ) 2 8$ , &quot; 465 recall for title extraction from PowerPoint are 0.907 and 0.951 respectively. Metadata extraction has been intensively studied. For instance, Han et al [14] proposed a method for metadata extraction from research papers. They considered the problem as that of classification based on SVM. They mainly used linguistic information as features. To the best of our knowledge, no previous work has been done on metadata extraction from general documents. We report our title extraction work in details in [19]. The feature of `who is' can help find documents authored by a person, but existing in different team web sites. Information extraction (specifically metadata extraction) makes the aggregation of information possible. 5.2.4 `Who knows about' The basic idea for the feature is that if a person has authored many documents on an issue (term), then it is very likely that he/she is an expert on the issue, or if the person's name co-occurs in many times with the issue, then it is likely that he/she is an expert on the issue. As described above, we can extract titles, authors, and key terms from all the documents. In this way, we know how many times each person is associated with each topic in the extracted titles and in the extracted key terms. We also go through all the documents and see how many times each person's name co-occurs with each topic in text segments within a pre-determined window size. In search, we use the three types of information: topic in title, topic in key term, and topic in text segment to rank persons, five persons for each type. We rank persons with a heuristic method and return the list of ranked persons. A person who has several documents with titles containing the topic will be ranked higher than a person whose name co-occurs with the topic in many documents. It appears that the results of the feature largely depend on the size of document collection we crawl. Users' feedbacks on the results show that sometimes the results are very accurate, however, sometimes they are not (due to the lack of information). Craswell et al. developed a system called `P@NOPTIC', which can automatically find experts using documents on an intranet [7]. The system took documents as plain texts and did not utilize metadata of documents as we do at Information Desk. 5.2.5 `Where is homepage of' We identify homepages (team web sites) using several rules. Most of the homepages at the intranet of Microsoft are created by SharePoint, a product of Microsoft. From SharePoint, we can obtain a property of each page called `ContentClass'. It tells exactly whether a web page corresponds to a homepage or a team site. So we know it is a homepage (obviously, this does not apply in general). Next we use several patterns to pull out titles from the homepages. The precision of home page identification is nearly 100%. In search, we rank the discovered home pages related to a query term using the URL lengths of the home pages. A home page with a shorter URL will be ranked higher. TREC has a task called `home/named page finding' [8, 9], which is to find home pages talking about a topic. Many methods have been developed for pursuing the task [5, 6, 26, 29]. Since we can identify homepages by using special properties on our domain, we do not consider employing a similar method. EVALUATION Usually it is hard to conduct evaluation on a practical system. We evaluated the usefulness of Information Desk by conducting a survey and by recording system logs. We have found from analysis results that the `what is' and `where is homepage of' features are very useful. The `who is' feature works well, but the `who knows about' feature still needs improvements. 6.1 Survey Result Analysis The survey described in section 4.3 also includes feedbacks on Information Desk. Figure 6 shows a question on the usefulness of the features and a summary on the answers. We see that the features `where is homepage of' and `what is' are regarded useful by the responders in the survey. Figure 7 shows a question on new features and a summary on the answers. We see that the users want to use the features of `how to', `when', `where' and `why' in the future. This also justifies the correctness of our claim on intranet search made in section 4. Figure 8 shows a question on purposes of use and a digest on the results. About 50% of the responders really want to use Information Desk to search for information. There is also an open-ended question asking people to make comments freely. Figure 9 gives some typical answers from the responders. The first and second answers are very positive, while the third and fourth point out the necessity of increasing the coverage of the system. Which feature of Information Desk has helped you in finding information? `where is homepage of' - finding homepages 54 % `what is' - finding definitions/acronyms 25 % `who is' - finding information about people 18 % `who knows about' - finding experts 3 % Figure 6. Users' evaluation on Information Desk What kind of new feature do you want to use at Information Desk? (multiple choice) `how to' - e.g., &quot;how to activate Windows&quot; 57 % `when' - e.g., &quot;when is Yukon RTM&quot; 57 % `where' - e.g., &quot;where can I find an ATM&quot; 39 % `why' - e.g., &quot;why doesn't my printer work&quot; 28 % others 9 % Figure 7. New features expected by users 466 I visited Information Desk today to conduct testing on Information Desk 54 % search for information related to my work 46 % Figure 8. Motivation of using Information Desk Please provide any additional comments, thanks! This is a terrific tool! Including `how to' and `when' capabilities will put this in the `can't live without it' category. Extremely successful searching so far! Very nice product with great potential. I would like to see more `Microsoftese' definitions. There is a lot of cultural/tribal knowledge here that is not explained anywhere. Typing in my team our website doesn't come up in the results, is there any way we can provide content for the search tool e.g., out group sharepoint URL? ... Figure 9. Typical user comments to Information Desk 6.2 System Log Analysis We have made log during the running of Information Desk. The log includes user IP addresses, queries and clicked documents (recall that links to the original documents, from which information has been extraction, are given in search). The log data was collected from 1,303 unique users during the period from November 26 th , 2004 to February 22 nd , 2005. The users were Microsoft employees. In the log, there are 9,076 query submission records. The records include 4,384 unique query terms. About 40% of the queries are related to the `what is' feature, 29% related to `where is homepage of', 30% related to `who knows about' and 22% related to `who is'. A query can be related to more than one feature. In the log, there are 2,316 clicks on documents after query submissions. The numbers of clicks for the `what is', `where is homepage of', `who knows about', and `who is' features are 694, 1041, 200 and 372, respectively. Note that for `what is', `where is home page of', and `who knows about' we conduct ranking on retrieved information. The top ranked results are considered to be the best. If a user has clicked a top ranked document, then it means that he is interested in the document, and thus it is very likely he has found the information he looks for. Thus a system which has higher average rank of clicks is better than the other that does not. We used average rank of clicked documents to evaluate the performances of the features. The average ranks of clicks for `what is', `where is homepage of' and `who knows about' are 2.4, 1.4 and 4.7 respectively. The results indicate that for the first two features, users usually can find information they look for on the top three answers. Thus it seems safe to say that the system have achieved practically acceptable performances for the two features. As for `who is', ranking of a person's documents does not seem to be necessary and the performance should be evaluated in a different way. (For example, precision and recall of metadata extraction as we have already reported in section 5). CONCLUSION In this paper, we have investigated the problem of intranet search using information extraction. Through an analysis of survey results and an analysis of search log data, we have found that search needs on intranet can be categorized into a hierarchy. Based on the finding, we propose a new approach to intranet search in which we conduct search for each special type of information. We have developed a system called `Information Desk', based on the idea. In Information Desk, we provide search on four types of information - finding term definitions, homepages of groups or topics, employees' personal information and experts on topics. Information Desk has been deployed to the intranet of Microsoft and has received accesses from about 500 employees per month. Feedbacks from users show that the proposed approach is effective and the system can really help employees to find information. For each type of search, information extraction technologies have been used to extract, fuse, and summarize information in advance. High performance component technologies for the mining have been developed. As future work, we plan to increase the number of search types and combine them with conventional relevance search. ACKNOWLEDGMENTS We thank Jin Jiang, Ming Zhou, Avi Shmueli, Kyle Peltonen, Drew DeBruyne, Lauri Ellis, Mark Swenson, and Mark Davies for their supports to the project. REFERENCES [1] S. Blair-Goldensohn, K.R. McKeown, A.H. Schlaikjer. A Hybrid Approach for QA Track Definitional Questions. In Proc. of Twelfth Annual Text Retrieval Conference (TREC12 ), NIST, Nov., 2003. [2] E. Brill, S. Dumais, and M. Banko, An Analysis of the AskMSR Question-Answering System, EMNLP 2002 [3] M. Chen, A. Hearst, A. Marti, J. Hong, and J. Lin, Cha-Cha: A System for Organizing Intranet Results. Proceedings of the 2nd USENIX Symposium on Internet Technologies and Systems. Boulder, CO. Oct. 1999. [4] C. L. A. Clarke, G. V. Cormack, T. R. Lynam, C. M. Li, and G. L. McLearn, Web Reinforced Question Answering (MultiText Experiments for TREC 2001). TREC 2001 [5] N. Craswell, D. Hawking, and S.E. Robertson. Effective site finding using link anchor information. In Proc. of the 24th annual international ACM SIGIR conference on research and development in information retrieval, pages 250--257, 2001. [6] N. Craswell, D. Hawking, and T. Upstill. TREC12 Web and Interactive Tracks at CSIRO. In TREC12 Proceedings, 2004. [7] N. Craswell, D. Hawking, A. M. Vercoustre, and P. Wilkins. P@noptic expert: Searching for experts not just for documents. Poster Proceedings of AusWeb'01, 467 2001b./urlausweb.scu.edu.au/aw01/papers/edited/vercoustre/ paper.htm. [8] N. Craswell, D. Hawking, R. Wilkinson, and M. Wu. Overview of the TREC-2003 Web Track. In NIST Special Publication: 500-255, The Twelfth Text REtrieval Conference (TREC 2003), Gaithersburg, MD, 2003. [9] N. Craswell, D. Hawking, R. Wilkinson, and M. Wu. Task Descriptions: Web Track 2003. In TREC12 Proceedings, 2004. [10] H. Cui, M-Y. Kan, and T-S. Chua. Unsupervised Learning of Soft Patterns for Definitional Question Answering, Proceedings of the Thirteenth World Wide Web conference (WWW 2004), New York, May 17-22, 2004. [11] A. Echihabi, U.Hermjakob, E. Hovy, D. Marcu, E. Melz, D. Ravichandran. Multiple-Engine Question Answering in TextMap. In Proc. of Twelfth Annual Text Retrieval Conference (TREC-12), NIST, Nov., 2003. [12] R. Fagin, R. Kumar, K. S. McCurley, J. Novak, D. Sivakumar, J. A. Tomlin, and D. P. Williamson. Searching the workplace web. Proc. 12th World Wide Web Conference, Budapest, 2003. [13] S. Feldman and C. Sherman. The high cost of not finding information. Technical Report #29127, IDC, April 2003. [14] H. Han, C. L. Giles, E. Manavoglu, H. Zha, Z. Zhang, and E. A. Fox. Automatic Document Metadata Extraction using Support Vector Machines. In Proceedings of the third ACM/IEEE-CS joint conference on Digital libraries, 2003 [15] S. Harabagiu, D. Moldovan, C. Clark, M. Bowden, J. Williams, J. Bensley. Answer Mining by Combining Extraction Techniques with Abductive Reasoning. In Proc. of Twelfth Annual Text Retrieval Conference (TREC-12), NIST, Nov., 2003. [16] D. Hawking. Challenges in Intranet search. Proceedings of the fifteenth conference on Australasian database. Dunedin, New Zealand, 2004. [17] D. Hawking, N. Craswell, F. Crimmins, and T. Upstill. Intranet search: What works and what doesn't. Proceedings of the Infonortics Search Engines Meeting, San Francisco, April 2002. [18] E. Hovy, L. Gerber, U. Hermjakob, M. Junk, and C. Y. Lin. Question Answering in Webclopedia. TREC 2000 [19] Y. Hu, H. Li, Y. Cao, D. Meyerzon, and Q. Zheng. Automatic Extraction of Titles from General Documents using Machine Learning. To appear at Proc. of Joint Conference on Digital Libraries (JCDL), 2005. Denver, Colorado, USA. 2005. [20] A. Ittycheriah and S. Roukos, IBM's Statistical Question Answering System-TREC 11. TREC 2002 [21] J. Klavans and S. Muresan. DEFINDER: Rule-Based Methods for the Extraction of Medical Terminology and their Associated Definitions from On-line Text. In Proceedings of AMIA Symposium 2000. [22] C. C. T. Kwok, O. Etzioni, and D. S. Weld, Scaling question answering to the Web. WWW-2001: 150-161 [23] Y. Li, H Zaragoza, R Herbrich, J Shawe-Taylor, and J. S. Kandola. The Perceptron Algorithm with Uneven Margins. in Proceedings of ICML'02. [24] B. Liu, C. W. Chin, and H. T. Ng. Mining Topic-Specific Concepts and Definitions on the Web. In Proceedings of the twelfth international World Wide Web conference (WWW-2003 ), 20-24 May 2003, Budapest, HUNGARY. [25] D. Mattox, M. Maybury and D. Morey. Enterprise Expert and Knowledge Discovery. Proceedings of the HCI International '99 (the 8th International Conference on Human-Computer Interaction) on Human-Computer Interaction: Communication, Cooperation, and Application Design-Volume 2 - Volume 2. 1999. [26] P. Ogilvie and J. Callan. Combining Structural Information and the Use of Priors in Mixed Named-Page and Homepage Finding. In TREC12 Proceedings, 2004. [27] D. R. Radev, W. Fan, H. Qi, H. Wu, and A. Grewal. Probabilistic question answering on the web. WWW 2002: 408-419 [28] D. E. Rose and D. Levinson. Understanding user goals in web search. Proceedings of the 13th international World Wide Web conference on Alternate track papers & posters, 2004 New York, USA. [29] J. Savoy, Y. Rasolofo, and L. Perret, L. Report on the TREC-2003 Experiment: Genomic and Web Searches. In TREC12 Proceedings, 2004. [30] D. Stenmark. A Methodology for Intranet Search Engine Evaluations. Proceedings of IRIS22, Department of CS/IS, University of Jyvskyl, Finland, August 1999. [31] V. N. Vapnik. The Nature of Statistical Learning Theory. Springer, 1995. [32] J. Xu, Y. Cao, H. Li, and M. Zhao. Ranking Definitions with Supervised Learning Methods. In Proc. of 14 th International World Wide Web Conference (WWW05), Industrial and Practical Experience Track, Chiba, Japan, pp.811-819, 2005. [33] J. Xu, A. Licuanan, R. Weischedel. TREC 2003 QA at BBN: Answering Definitional Questions. In Proc. of 12 th Annual Text Retrieval Conference (TREC-12), NIST, Nov., 2003. [34] H. Yang, H. Cui, M. Maslennikov, L. Qiu, M-Y. Kan, and TS . Chua, QUALIFIER in TREC-12 QA Main Task. TREC 2003: 480-488 [35] Intellectual capital management products. Verity, http://www.verity.com/ [36] IDOL server. Autonomy, http://www.autonomy.com/content/home/ [37] Fast data search. Fast Search & Transfer, http://www.fastsearch.com/ [38] Atomz intranet search. Atomz, http://www.atomz.com/ [39] Google Search Appliance. Google, http://www.google.com/enterprise/ 468

Search Needs;metadata extraction;features;architecture;Experimentation;definition search;INFORMATION DESK;information extraction;expert finding;Algorithms;intranet search;Human Factors;information retrieval;component technologies;Intranet search;types of information

140

Modeling Node Compromise Spread in Wireless Sensor Networks Using Epidemic Theory

Motivated by recent surfacing viruses that can spread over the air interfaces, in this paper, we investigate the potential disastrous threat of node compromise spreading in wireless sensor networks. Originating from a single infected node, we assume such a compromise can propagate to other sensor nodes via communication and preestablished mutual trust. We focus on the possible epidemic breakout of such propagations where the whole network may fall victim to the attack. Based on epidemic theory, we model and analyze this spreading process and identify key factors determining potential outbreaks. In particular, we perform our study on random graphs precisely constructed according to the parameters of the network, such as distance, key sharing constrained communication and node recovery, thereby reflecting the true characteristics therein. The analytical results provide deep insights in designing potential defense strategies against this threat. Furthermore , through extensive simulations, we validate our model and perform investigations on the system dynamics. Index Terms-- Sensor Networks, Epidemiology, Random Key Predistribution, Random Graph.

Introduction As wireless sensor networks are unfolding their vast potential in a plethora of application environments [1], security still remains one of the most critical challenges yet to be fully addressed. In particular, a vital problem in the highly distributed and resource constrained environment is node compromise, where a sensor node can be completely captured and manipulated by the adversary. While extensive work has focused on designing schemes that can either defend and delay node capture or timely identify and revoke compromised nodes themselves [5], little attention has been paid to the node compromise process itself. Inspired by recently emerged viruses that can spread over air interfaces, we identify in this paper the threat of epidemic spreading of node compromises in large scale wireless sensor networks and present a model that captures the unique characteristic of wireless sensor networks in conjunction with pairwise key schemes. In particular, we identify the key factors determining the potential epidemic outbreaks that in turn can be employed to devise corresponding defense strategies. A. Motivation Due to its scarce resources and hence low defense capabilities , node compromises can be expected to be common phenomena for wireless sensor networks in unattended and hostile environments. While extensive research efforts, including those from ourselves [15], have been engineered toward designing resilient network security mechanisms [12], [13], the compromise itself and in particular the propagation of node compromise (possible epidemics) have attracted little attention. While node compromise, thanks to physical capture and succeeding analysis, is naturally constrained by the adver-sary's capability, software originated compromises can be much more damaging. Specifically, the recently surfaced virus Cabir 1 that can spread over the air interface has unveiled a disastrous threat for wireless sensor networks. Inescapably, viruses targeting wireless sensor networks will emerge. Consequently, node compromise by way of virus spreading (over the air interface) can effortlessly devastate the entire network in a short period of time. With recent advancements on sensor design empowering nodes such as MICA2 motes with over-the-air programmability, the network becomes vulnerable to the above described attack. Even worse, the inherent dense, large scale nature of sensor networks undoubtedly further facilitates the virus propagation. While virus spreading over the internet has been widely studied, and notably by means of epidemic theory [2], [3], the distance and pairwise key restricted communication pattern in wireless sensor networks uniquely distinguish the phenomena from those on the Internet. B. Our Contribution In this paper, we investigate the spreading process of node compromise in large scale wireless sensor networks. Starting from a single point of failure, we assume that the adversary can effectively compromise neighboring nodes through wireless communication and thus can threat the whole network without engaging in full scale physical attacks. In particular, due to security schemes employed by the sensor networks, we assume that communication can only be performed when neighboring nodes can establish mutual trust by authenticating a common key. Therefore, node compromise is not only determined by the deployment of sensor nodes which in turn affects node density, but also determined by the pairwise key scheme employed therein. By incorporating these factors of the networks, we propose an epidemiological model to investigate the probability of a breakout (compromise of the whole network ) and if not, the sizes of the affected components (compromised clusters of nodes). Furthermore, we analyze the effect of node recovery in an active infection scenario and obtain critical values for these parameters that result in an outbreak. Through extensive simulations, we show that our analytical results can closely capture the effects in a wide range of network setups. The remainder of the paper is organized as follows. In Section II we present the preliminaries, including the threat model, random key pre-distribution, and epidemic theory. In Section III, we study the compromise propagation without node recovery and with node recovery, and detail our analytical results. We perform experimental study in Section IV. Related work is presented in Section V and we conclude in Section VI. Preliminaries In this section, we present our threat model and briefly overview pairwise key distribution in wireless sensor networks and epidemic theory. A. Threat model We assume that a compromised node, by directly communicating with a susceptible node, can spread the infection and conduce to the compromise of the susceptible node. Communication among sensor nodes is not only constrained by their distances, but also shall be secured and thus determined by the probability of pairwise key sharing. Therefore, the spreading of node compromise is dependent on the network deployment strategy and the pairwise key scheme employed therein. We assume that the "seed" compromise node could be originated by an adversary through physical capture and analysis of that node or by other similar means. The spread of node compromise in a wireless sensor network, particularly thanks to its dense nature, can lead to an epidemic effect where the whole network will get infected. We consider this epidemic effect as the key threat to the network and hence the investigation target of this paper. B. Pairwise Key Pre-distribution As the pairwise key scheme affects the communication and hence the propagation of the node compromise, we provide below, a brief overview of the key distribution schemes in wireless sensor networks. Due to the severe resource constraint of wireless sensor networks and limited networking bandwidth, proposed pairwise key schemes have commonly adopted the predistribution approach instead of online key management schemes with prohibitive resource consumption. The concept of pre-distribution was originated from [11], where the authors propose to assign a number of keys, termed key ring randomly drawn from a key pool. If two neighboring nodes share a common key on their key rings, a shared pairwise key exists and a secure communication can be established. Pre-distribution schemes that rely on bivariate polynomials is discussed in [13]. In this scheme, each sensor node is pre-distributed a set of polynomials. Two sensor nodes with the same polynomials can respectively derive the same key. Regardless of the specific key distribution scheme, a common parameter capturing the performance is the probability that two neighbors can directly establish a secure communication. We denote this probability by q. As it shall be revealed later, q plays an important role in the spreading of node compromise, because direct communication, as explained in the threat model, can result in propagation of malicious code. C. Node Recovery In the event that a node is compromised, its secrets will be revealed to the attacker. The network may attempt to recover the particular node. Recovery might be realized in several possible ways. For example, the keys of the nodes might be revoked and the node may be given a fresh set of secret keys. In this context, key revocation, which refers to the task of securely removing keys that are known to be compromised, has been investigated as part of the key management schemes, for example in [5]. Moreover, recovery can also be achieved by simply removing the compromised node from the network, for example by announcing a blacklist, or simply reload the node's programs. More sophisticated methods may include immunizing a node with an appropriate antivirus patch that might render the node immune from the same virus attack. Regardless, in our analysis, we will study virus spreading under the two cases respectively depending on whether a node can be recovered or not. D. Epidemic Theory Originally, epidemic theory concerns about contagious diseases spreading in the human society. The key feature of epidemiology [2], [7] is the measurement of infection outcomes in relation to a population at risk. The population at risk basically comprises of the set of people who possess a susceptibility factor with respect to the infection. This factor is dependent on several parameters including exposure, spreading rate, previous frequency of occurrence etc., which define the potential of the disease causing the infection. Example models characterizing the infection spreading process include the Susceptible Infected Susceptible (SIS) Model, Susceptible Infected Recovered (SIR) Model etc. In the former, a susceptible individual acquires infection and then after an infectious period, (i.e., the time the infection persists), the individual becomes susceptible again. On the other hand, in the latter, the individual recovers and becomes immune to further infections. Of particular interest is the phase transition of the spreading process that is dependent on an epidemic threshold : if the epidemic parameter is above the threshold, the infection will spread out and become persistent; on the contrary, if the parameter is below the threshold, the virus will die out. Epidemic theory indeed has been borrowed to the networking field to investigate virus spreading. In this paper, we will mainly rely on a random graph model to characterize the unique connectivity of the sensor network and perform the epidemic study [8], [10]. III Modelling and Analysis of Compromise Propagation In this section, we analyze the propagation of node compromise originating from a single node that has been affected. Our focus is to study the outbreak point of the epidemic effect where the whole network will fall victim to the compromise procedure. Our key method is to characterize the sensor network , including its key distribution, by mathematically formulating it as a random graph whose key parameters are precisely determined by those of the sensor network. Therefore, the investigation of epidemic phenomena can be performed on the random graph instead. Following this approach, we observe the epidemic process under two scenarios: without node recovery and with node recovery, depending on whether infected nodes will be recovered by external measures like key revocation, immunization, etc. A. Network Model as Random Graph Assume that sensor nodes are uniformly deployed in a disc area with radius R. Let = N R 2 denote the node density of the network where N is the total number of the nodes. For a sensor node with communication range r, the probability that l nodes are within its communication range is given by p(l) = nl p l (1 - p) n-l (1) where p is defined by p = r 2 R 2 = r 2 N . (2) Thus p is the probability of a link existing at the physical level, i.e., whether the two nodes fall within their respective communication ranges. We further assume that the probability that two neighboring nodes sharing at least one key in the random predistribution pairwise key is q. Notice that q is determined by the specific pairwise key scheme employed. For a particular node having l neighboring nodes, the probability that there are k nodes, k l, sharing at least one key with it is given by p(k|l) = lk q k (1 - q) l-k (3) Therefore, the probability of having k neighboring nodes sharing at least one key is p(k) = l=k p(l)p(k|l) (4) = l=k n l p l (1 - p) n-l l k q k (1 - q) l-k (5) Thus, based on both physical proximity and the probability of key sharing between neighbors, we get a degree distribution p(k). Notice that this degree distribution can be employed to generate a random graph G. Since G possesses the same property in terms of secure communication pattern as the sensor network of concern, we will next perform the analysis on G instead. B. Compromise Spread Without Node Recovery Given the random graph construction, we now analyze the case of compromise spread when no node recovery is performed. In other words, a compromised sensor node will remain infectious indefinitely. Let G 0 (x) be the generating function of the degree distribution of a randomly chosen vertex in G and is defined by G 0 (x) = k=0 p(k)x k (6) Moreover, with G 1 (x) given by G 1 (x) = 1 G 0 (1)G 0 (x) (7) and with denoting the infection probability of a node being infected by communicating with a compromised node, then following the analysis presented in [8], the average size of the outbreak is derived as s = 1 + G 0 (1) 1 - G 1 (1). (8) Infection probability essentially captures the spreading capability of the virus that could compromise the network: the larger it is, the stronger the virus is. We assume that its value can be obtained by means of measurement or analysis. Given the above result, we can see that the outbreak point for the network is = 1/G 1 (1) which marks the onset of an epidemic. For &gt; 1/G 1 (1) we have an epidemic in the form of a giant component in the random network and the size S of the epidemic, where S denotes the expected fraction of the network that will be compromised if an outbreak happens, is given by S = 1 - G 0 (u). Here u is the root of the self-consistency relation u = G 1 (u). Intuitively, the above conclusion reveals that if 1/G 1 (1), the component of compromised nodes is finite 3 Proceedings of the 2006 International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM'06) 0-7695-2593-8/06 $20.00 2006 IEEE 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 10 20 30 40 50 60 70 80 90 100 Infection Probability( ) Size of Cluster Compromised q = 0.01 q = 0.02 q = 0.04 q = 0.1 q = Key Sharing Prob (a) Non-epidemic cluster size vs. infection probability ( ) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Infection Probability ( ) Fraction of Network Compromised q = 0.01 q = 0.02 q = 0.04 q = 0.1 q = Key Sharing Prob (b) Epidemic size vs. infection probability ( ) Fig. 1. Size of compromised node clusters: (a) depicts the average size of infected clusters when there is no epidemic and (b) shows the epidemic size as the fraction of the entire network. The point where non-zero value appears indicates the transition from non-epidemic to epidemic in size regardless of the size of the network and each node's probability of being compromised is zero for large networks. On the contrary, if &gt; 1/G 1 (1), there always exists a finite probability for a node to be compromised. Fig. 1 depicts this effect for a network with N = 1000 nodes with different key sharing probabilities q. The underlying physical topology, determined by the communication range and node density, has an average edge probability of p = 0.25. Given the physical deployment, we vary the probability of direct pairwise key sharing ( q) and study the point of outbreak. As we can see in Fig. 1, while undoubtedly increasing q can facilitate communication in the network, the network also becomes more vulnerable to virus spreading. Specifically, when q = 0.01, network wide breakout is only possible when a compromised node has an infection probability ( ) larger than 0.4 to infect a neighbor. We note that in this case, we have an average node degree of 2.5. On the contrary, this probability only needs to be around 0.05 when q = 0.1 which subsequently makes the node degree 25. Fig. 1(b) illustrates the fraction of the network that is ultimately infected as the infection probability is increased beyond the critical point of the onset of outbreak. For instance, we observe that when q = 0.1, the whole network is compromised with a value of less than 0.2. On the contrary, with q = 0.01, 80% of the network could be compromised with only a high value of = 0.8. In summary, Fig. 1 clearly indicates the tradeoff between key sharing probability among sensor nodes and the vulnerability of the network to compromise. C. Compromise Spread With Node Recovery In this case, we assume that the network has the capability to recover some of the compromised nodes by either immunization or removal from the network. To capture this recovery effect, we assume that an infected node recovers or is removed from the network after an average duration of infectivity . In other words, a node in the sensor network remains infective for an average period after which it is immunized. During this infective period, the node transmits the epidemic to its neighbors with the infection rate , denoting the probability of infection per unit time. Evidently, the parameter is critical to the analysis as it measures how soon a compromised node recovers. Naturally, we will perform our analysis following the SIR model in epidemic theory [10], [8]. First, consider a pair of adjacent nodes where one is infected and the other is susceptible. If T denotes the compromise transmission probability, given the above definitions for and , we can say that the probability that the disease will not be transmitted from the infected to the susceptible is given by 1 - T = lim t0 (1 - t) /t = e . (9) Subsequently, we have the transmission probability T = 1 - e . In other words, the compromise propagation can be considered as a Poisson process, with average . The outcome of this process is the same as bond percolation and T is basically analogous to the bond occupation probability on the graph representing the key sharing network. Thus, the outbreak size would be precisely the size of the cluster of vertices that can be reached from the initial vertex (infected node) by traversing only occupied edges which are occupied with probability T . Notice that T explicitly captures node recovery in terms of the parameter . Replacing with T in Equation 8, and following similar steps, we get the size of the average cluster as s = 1 + T G 0 (1) 1 - T G 1 (1). (10) and the epidemic size is obtained by S = 1 - G 0 (u; T ). (11) where u is obtained by u = 1 - G 1 (u; T ), (12) 4 Proceedings of the 2006 International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM'06) 0-7695-2593-8/06 $20.00 2006 IEEE 0 50 100 150 0 50 100 150 Infectivity Duration Size of Cluster Compromised = 0.01 = 0.02 = 0.04 = 0.2 = Infection rate (a) Non-epidemic cluster size vs. infectivity duration 0 50 100 150 200 250 300 350 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Infectivity Duration Fraction of Network Compromised = 0.01 = 0.02 = 0.04 = 0.08 = 0.2 = infection rate (b) Epidemic size vs. infectivity duration Fig. 2. Size of compromised node clusters: (a) depicts the average size of infected clusters when there is no epidemic and (b) shows the epidemic size as the fraction of the entire network. The point where non-zero value appears indicates the transition from non-epidemic to epidemic and G 0 (u; T ) and G 1 (u; T ) are given respectively by G 0 (u; T ) = G 0 (1 + (u - 1)T ), (13) and G 1 (u; T ) = G 1 (1 + (u - 1)T ). (14) Fig. 2 summarizes this effect, depicting the epidemic outbreak against the average recovery time for the respective infection rates . The plots are for a sensor network with typical average degree of 10. In Fig. 2(a), we can identify the average duration that an infected node is allowed to remain infective before an epidemic outbreak occurs. We notice that, when the infection rate is 0.01, infected nodes have to be recovered/removed on the average in less than 100 time units in order to prevent an epidemic. As expected, this time is much lower when the infection rate is 0.2. Fig. 2(b) depicts the epidemic outbreak point for different infection rates in terms of the average duration of infectivity of a node. We remark that both the analytical and experimental results have significant implication for security scheme design in terms of revoking/immunizing compromised nodes in wireless sensor networks: it dictates the speed at which the network must react in order to contain/prevent the effect of network wide epidemic. Simulation We employ a discrete event-driven simulation to accu-rately simulate the propagation of the infection spreading process. In this section, we first outline our discrete-event driven simulation model for the gradual progress of the spread of node compromise. Then we use this model to capture the time dynamics of the spread of the compromise in the whole population. A. Simulation Setup In our simulation, we assume the number of sensor nodes in the network to be 1000. The sensor network is produced by uniformly distributing the sensors in a 12001200 unit 2 area. The communication range of each node is assumed to be 100 units. Our goal is to make the physical network fairly connected with an average node degree of around 20 to 25. We use the key sharing probability on top of this network to further reduce the average node degree of the final key sharing network to typical values of 3 and 10. We employ the random key pre-distribution scheme described in [11] to establish the pairwise key among sensor nodes. By tuning the parameters of the scheme, we can achieve any specific values for the probability of any two neighbors to share at least one key. Our simulation works in two phases. In the first phase, we form the network where each node identifies its set of neighbors and entries are made into a neighbor table. The average degree of the key sharing network is controlled by changing the value of the key sharing probability between neighbors. The entry for each node in the neighborhood table can indicate whether a node is susceptible, infected or recovered. We use typical values obtained for the average node degree of the network, namely, 3 and 10. In the second phase, we simulate actual virus propagation . Initially, at t = 0, the number of infected nodes, denoted by I(0) is set to be 1. At any time point t, the population is divided into the group of susceptible nodes, S(t), and the group of infected nodes, I(t). In the situation where we have nodes that are immunized and thus recovered, we denote that this set of recovered nodes by R(t). The sub-population dynamics is obtained by observing the population counts after fixed simulation intervals of 1 time unit. We assume that the time it takes for an infected node to infect its susceptible neighbor is negative exponentially distributed with a mean of 1 unit time. There are two simulation scenarios corresponding to our analysis. B. Simulation Results and Discussion 1) Simulation Results for No Recovery Case: The simulation results for the case without recovery are shown 5 Proceedings of the 2006 International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM'06) 0-7695-2593-8/06 $20.00 2006 IEEE 0 100 200 300 400 500 600 700 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Time(t) Fraction of compromised nodes Growth of infected nodes with time = 0.05 = 0.1 = 0.2 = 0.3 N = 1000 z = 5 (a) Average node degree = 5 0 100 200 300 400 500 600 700 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Time(t) Fraction of compromised nodes Growth of infected nodes with time = 0.05 = 0.1 = 0.2 = 0.3 N = 1000 Avg Degree z = 10 = Infection Prob (b) Average node degree = 10 Fig. 3. System dynamics without recovery in Fig. 3. We vary the value of the infection probability under different network connectivities and study the time dynamics of the infected population. We notice, as expected, that an increase in the average node degree from 5 to 10 has an impact on the rate of compromise of the network. For instance, the curve with the lowest value(0.05) has compromised the entire network by simulation time 700 when the average node degree is 10. However, with the node degree at 5, a value of 0.05 could compromise upto 70% of the network by that same simulation time. Thus, we find that in the no-recovery case the two key parameters affecting the network compromise rate are infection probability and the average node degree. 2) Simulation Results for Recovery Case: Fig. 4 and 5 show the simulation results for the three sub-populations (infected, immunized, and susceptible) in the situation where nodes do recover. In Fig. 4 we see the effects of the infectivity duration and infection rate on the dynamics of the epidemic. In Fig. 4(c), the highest point is reached very fast because of the high value of . Thereafter, its recovery also takes less time. However, in Fig. 4(a), is smaller but 0 is higher (i.e., 30), the infection rises slowly and also falls slowly because of the high recovery time. In comparison, Fig. 5 has better connectivity of average node degree of 5 which in turn increases the rate of infection significantly. Comparing Fig. 4(c) and Fig. 5(c), we observe that infection penetration is higher in the latter even in presence of a smaller value of . In Fig. 5(c), it shows that even with a low value of , the infection still rises to above 60%. Therefore, we observe that network connectivity has a high impact on the infection propagation and on the speed of reaching the maximal point of outbreak. However, thereafter during the recover phase, 0 affects aggressively the time it takes to recover the whole network. Related Work The mathematical modeling of epidemics is well documented [2], [7]. In fact, visualizing the population as a complex network of interacting individuals has resulted in the analysis of epidemics from a network or graph theoretic point of view [8], [9], [10]. Node compromise in sensor networks and the need for their security has also received immense attention [4]. A large portion of current research on security in sensor networks has been focused on protocols and schemes for securing the communication between nodes [12], [13]. Revocation of keys of compromised nodes has been studied in [14]. In [4], the authors demonstrate the ease with which a sensor node can be compromised and all its information extracted. Unfortunately, little work has been done on the defense strategies when the compromise of a single node could be used to compromise other nodes over the air. In this paper, we take the first step to model this potential disastrous propagation. In [6], the authors used an epidemic modeling technique for information dissemination in a MANET. However, they assumed homogeneous mixing which is not possible in a static sensor network as ours. In our work, we adopted some of the results presented in [8] where the author proposes a percolation theory based evaluation of the spread of an epidemic on graphs with given degree distributions. However, little has been shown there on the temporal dynamics of the epidemic spread and the authors only studied the final outcome of an infection spread. Conclusion In this paper, we investigate the potential threat for compromise propagation in wireless sensor networks. Based on epidemic theory, we model the process of compromise spreading from a single node to the whole network. In particular, we focus on the key network parameters that determine a potential epidemic outbreak in the network. Due to the unique distance and key sharing constrained communication pattern, we resort to a random graph model which is precisely generated according to the parameters of the real sensor network and perform the study on the graph. Furthermore, we introduce the effect of node recovery after compromise and adapt our model to accommodate this effect. Our results reveal key network parameters in defending and containing potential epidemics. In particular, 6 Proceedings of the 2006 International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM'06) 0-7695-2593-8/06 $20.00 2006 IEEE 0 100 200 300 400 500 600 700 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 Time(t) Fraction of Network Dynamics of the Infected, Susceptible and Revoked nodes Susceptible S(t) Infected I(t) Recovered R(t) Average Degree z = 3 Infection Prob = 0.6 (a) 0 = 30 0 100 200 300 400 500 600 700 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 Time(t) Fraction of Network Dynamics of the Infected, Susceptible and Revoked nodes S(t) I(t) R(t) z = 3 = 0.7 (b) 0 = 20 0 100 200 300 400 500 600 700 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 Time(t) Fraction of Network Dynamics of the Infected, Susceptible and Revoked nodes S(t) I(t) R(t) z = 3 = 0.9 (c) 0 = 10 Fig. 4. The dynamics of the population with recovery for average degree of 3 0 100 200 300 400 500 600 700 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 Time(t) Fraction of Network Dynamics of the Infected, Susceptible and Revoked nodes Susceptible S(t) Infected I(t) Recovered R(t) Average Degree z = 10 Infection Prob = 0.15 (a) 0 = 30 0 100 200 300 400 500 600 700 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 Time(t) Fraction of Network Dynamics of the Infected, Susceptible and Revoked nodes S(t) I(t) R(t) z = 10 = 0.17 (b) 0 = 20 0 100 200 300 400 500 600 700 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 Time(t) Fraction of Network Dynamics of the Infected, Susceptible and Revoked nodes S(t) I(t) R(t) z = 10 = 0.2 (c) 0 = 10 Fig. 5. The dynamics of the population with recovery for average degree of 10 the result provides benchmark time period for the network to recover a node in order to defend against the epidemic spreading. Our extensive simulation results validate our analyses and moreover, provide insights of the dynamics of the system in terms of temporal evolution. References [1] I Akyildiz, W. Su, Y Sankarasubramaniam, and E. Cayirci, "A Survey on sensor networks," IEEE Communications Magazine, vol. 40, no. 8, 2002. [2] R. M. Anderson and R. M. May, "Infectious Diseases of Human: Dynamics and Control" (Oxford Univ. Press, Oxford, 1991). [3] S. Staniford, V. Paxson, and N. Weaver. "How to Own the Internet in Your Spare Time". In 11th Usenix Security Symposium, San Francisco, August, 2002. [4] C. Hartung, J. Balasalle, and R. Han, "Node Compromise in Sensor Networks: The Need for Secure Systems", Technical Report CU-CS -990-05 (2005). [5] H. Chan, V. D. Gligor, A. Perrig, G. Muralidharan, "On the Distribution and Revocation of Cryptographic Keys in Sensor Networks", IEEE Transactions on Dependable and Secure Computing 2005. [6] A. Khelil, C. Becker, J. Tian, K. Rothermel, "An Epidemic Model for Information Diffusion in MANETs", MSWiM 2002, pages 54-60 . [7] N. T. J. Bailey, "The Mathematical Theory of Infectious Diseases and its Applications". Hafner Press, New York (1975) [8] M. E. J. Newman, "Spread of epidemic disease on networks", Phys. Rev. E, 66 (2002), art. no. 016128. [9] C. Moore and M. E. J. Newman, "Epidemics and percolation in small- world networks". Phys. Rev. E 61, 5678-5682 (2000) [10] P. Grassberger, "On the critical behavior of the general epidemic process and dynamic percolation", Math. Biosc. 63 (1983) 157. [11] L Eschenauer and V. D. Gligor. "A key-management scheme for distributed sensor networks", in Proc. of the 9th Computer Communication Security - CCS '02, pages 4147, Washington D.C., USA, November 2002. [12] H. Chan, A. Perrig, and D. Song, "Random key predistribution schemes for sensor networks", in Proc. of the IEEE Symposium on Research in Security and Privacy - SP '03, pages 197215, Washington D.C., USA, May 2003. [13] Donggang Liu and Peng Ning, "Establishing pairwise keys in distributed sensor networks", in Proc. of the 10th ACM Conference on Computer and Communications Security - CCS '03, pages 52 61, Washington D.C., USA, October 2003. [14] H. Chan; V.D. Gligor, A. Perrig, G. Muralidharan, "On the Distribution and Revocation of Cryptographic Keys in Sensor Networks", IEEE Transactions on Dependable and Secure Computing, Volume 2, Issue 3, July-Sept. 2005 [15] A. Chadha, Y. Liu. and S. Das, "Group key distribution via local collaboration in wireless sensor networks," in Proceedings of the IEEE SECON 2005, Santa Clara, CA, Sept. 2005. 7 Proceedings of the 2006 International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM'06) 0-7695-2593-8/06 $20.00 2006 IEEE

Random Key Predistribution;Sensor Networks;Random Graph;Epidemiology

141

Modelling Adversaries and Security Objectives for Routing Protocols in Wireless Sensor Networks

The literature is very broad considering routing protocols in wireless sensor networks (WSNs). However, security of these routing protocols has fallen beyond the scope so far. Routing is a fundamental functionality in wireless networks, thus hostile interventions aiming to disrupt and degrade the routing service have a serious impact on the overall operation of the entire network. In order to analyze the security of routing protocols in a precise and rigorous way, we propose a formal framework encompassing the definition of an adversary model as well as the "general" definition of secure routing in sensor networks. Both definitions take into account the feasible goals and capabilities of an adversary in sensor environments and the variety of sensor routing protocols. In spirit, our formal model is based on the simulation paradigm that is a successfully used technique to prove the security of various cryptographic protocols. However, we also highlight some differences between our model and other models that have been proposed for wired or wireless networks. Finally, we illustrate the practical usage of our model by presenting the formal description of a simple attack against an authenticated routing protocol, which is based on the well-known TinyOS routing.

INTRODUCTION Routing is a fundamental function in every network that is based on multi-hop communications, and wireless sensor networks are no exceptions. Consequently, a multitude of routing protocols have been proposed for sensor networks in the recent past. However, most of these protocols have not been designed with security requirements in mind. This means that they can badly fail in hostile environments . Paradoxically, research on wireless sensor networks have been mainly fuelled by their potential applications in military settings where the environment is hostile. The natural question that may arise is why then security of routing protocols for sensor networks has fallen beyond the scope of research so far. We believe that one important reason for this situation is that the design principles of secure routing protocols for wireless sensor networks are poorly understood today. First of all, there is no clear definition of what secure routing should mean in this context. Instead, the usual approach, exemplified in [10], is to list different types of possible attacks against routing in wireless sensor networks, and to define routing security implicitly as resistance to (some of) these attacks. However, there are several problems with this approach. For instance, a given protocol may resist a different set of attacks than another one. How to compare these protocols? Shall we call them both secure routing protocols ? Or on what grounds should we declare one protocol more secure than another? Another problem is that it is quite difficult to carry out a rigorous analysis when only a list of potential attack types are given. How can we be sure that all possible attacks of a given type has been considered in the analysis? It is not surprising that when having such a vague idea about what to achieve, one cannot develop the necessary design principles. It is possible to come up instead with some countermeasures, similar to the ones described in [10], which are potentially usefully to thwart some specific types of attacks, but it remains unclear how to put these ingredients together in order to obtain a secure and efficient routing protocol at the end. In order to remedy this situation, we propose to base the design of secure routing protocols for wireless sensor networks on a formal security model. While the benefit of formal models is not always clear (indeed, in some cases, they tend to be overly complicated compared to what they achieve), we have already demonstrated their advantages in the context of ad hoc network routing protocols. More specifically, we developed formal security models in [4, 1, 2], and we successfully used them to prove the security of some 49 ad hoc network routing protocols, and to find security holes in others. The idea here is to use the same approach in the context of wireless sensor networks. The rationale is that routing protocols in sensor networks are somewhat similar to those in ad hoc networks, hence they have similar pitfalls and they can be modeled in a similar way. Thus, in this paper, we present a formal model, in which security of routing is precisely defined, and which can serve as the basis for rigorous security analysis of routing protocols proposed for wireless sensor networks. Our model is based on the simulation paradigm, where security is defined in terms of indistinguishability between an ideal-world model of the system (where certain attacks are not possible by definition) and the real-world model of the system (where the adversary is not constrained, except that he must run in time polynomial). This is a standard approach for defining security, however, it must be adopted carefully to the specific environment of wireless sensor networks. Similar to [4], in this paper, we develop an adversary model that is different from the standard Dolev-Yao model, where the adversary can control all communications in the system. In wireless sensor networks, the adversary uses wireless devices to attack the systems, and it is more reasonable to assume that the adversary can interfere with communications only within its power range. In addition, we must also model the broadcast nature of radio communications. However, in addition to the model described in [4], here we take into account that there are some attacks which exploit the constraint energy supply of sensor nodes (e.g., the adversary decreases the network lifetime by diverting the traffic in order to overload, and thus, deplete some sensor nodes). Hence, we explicitly model the energy consumption caused by sending a message between each pair of nodes in the network. Another difference with respect to the model of [4] lies in the definition of the outputs of the ideal-world and the real-world models. It is tempting to consider the state stored in the routing tables of the nodes as the output, but an adversary can distort that state in unavoidable ways. This means that if we based our definition of security on the indistinguishability of the routing states in the ideal-world and in the real-world models, then no routing protocol would satisfy it. Hence, we define the output of the models as a suitable function of the routing state, which hides the unavoidable distortions in the states. This function may be different for different types of routing protocols, but the general approach of comparing the outputs of this function in the ideal-world and in the real-world models remain the same. For instance, this function could be the average length of the shortest pathes between the sensor nodes and the base station ; then, even if the routing tables of the nodes would not always be the same in the ideal-world and in the real-world models, the protocol would still be secure given that the difference between the distributions of the average length of the shortest pathes in the two models is negligibly small. The rest of the paper is organized as follows: In Section 2, we present the elements of our formal model, which includes the presentation of the adversary model adopted to wireless sensor networks, the description of the ideal-world and the real-world models, the general definition of the output of these models, as well as the definition of routing security. Then, in Section 3, we illustrate the usage of our model by representing in it a known insecurity of an authenticated version of the TinyOS routing protocol. Finally, in Section 4, we report on some related work, and in Section 5, we conclude the paper. We must note that the work described in this paper is a work in progress, and it should be considered as such. In particular, the reader will not find security proofs in this paper. There are two reasons for this: first, we are still developing the proof techniques, and second, we have not identified yet any routing protocols that would be secure in our model. THE MODEL OF WIRELESS SENSOR NETWORKS The adversary is represented by adversarial nodes in the network. An adversarial node can correspond to an ordinary sensor node, or a more resourced laptop-class device. In the former case, the adversary may deploy some corrupted sensor-class devices or may capture some honest sensor nodes. In the latter case, he has a laptop-class device with a powerful antenna and unconstrained energy supply. All of these adversarial nodes may be able to communicate in out-of-band channels (e.g., other frequency channel or direct wired connection), which may be used to create wormholes. In general, when capturing honest sensor nodes, the adversary may be able to compromise their cryptographic secrets (assuming that such secrets are used in the system). However, in this paper, we assume that the adversary cannot compromise cryptographic material. This is certainly a simplifying assumption, and we intend to relax it in our future work. The adversary attacking the routing protocol primarily intends to shorten the network lifetime, degrade the packet delivery ratio, increase his control over traffic, and increase network delay. Some of these goals are highly correlated; e.g., increasing hostile control over traffic may also cause the network delay to be increased. In order to achieve the aforementioned goals, the adversary is able to perform simple message manipulations: fab-ricated message injection, message deletion, message modification and re-ordering of message sequences. In the followings , we describe how the adversary can perform message deletion and injection in a wireless sensor network. Re-ordering of message sequences is straightforward using message deletion and insertion, thus, we do not elaborate it further. Basically, an adversarial node can affect the communication of two honest nodes in two cases: In the first case, an adversarial node relays messages between honest nodes which are not able to communicate directly with each other. In the second case, the honest nodes can also reach each other, and the adversarial node can also hear the nodes' communication , i.e., he can send and receive messages to/from both honest nodes. We further assume that communication range implies interference range, and vice-versa. In case of adversarial relaying of messages between the nodes, all of the message manipulations are quite straightforward . On the contrary, if the honest nodes can also communicate with each other, message manipulations must be performed in a very sophisticated way. The adversarial node can inject messages easily, but deletion and modification re-50 quire jamming capability. Message deletion may be achieved by employing various selective jamming techniques against either the sender node or the receiver node. Message modification is only feasible, if both the sender and the receiver nodes are within the communication range of the adversarial node. Here, we sketch two scenarios for message modification , which are illustrated on Figure 1. By these simple examples, we intend to point out the feasibility of message modification assuming even direct communication between the sender and the receiver node. Scenario 1: There are two honest nodes X and Y , and node X intends to send a message m to node Y . A 1 and A 2 are adversarial nodes, where A 2 is able to interfere with Y 's communication, but not with X's and A 1 's communication. Let A 1 be in the communication range of X and Y , whereas A 2 can only communicate with Y . When X transmits m to Y , node A 1 overhears m, meanwhile A 2 performs jamming to cause Y not to be able to receive m. In order to take this action, A 1 and A 2 are connected by an out-of-band channel, thus, A 1 can send a signal to A 2 when A 2 should start jamming Y 's communication. It is also feasible that A 2 performs constant jamming for a certain amount of time, afterwards, A 1 can send the modified message m to Y . Scenario 2: In this scenario, there is only one adversarial node denoted by A. We assume that transmitting a message from the routing sublayer consists of passing the message to the data-link layer, which, after processing the message, also passes it further to the physical layer. The data-link layer uses CRC in order to provide some protection against faults in noisy channels; a sender generally appends a frame check sequence to each frame (e.g., see [7]). The adversary can exploit this CRC mechanism to modify a message in the following way (illustrated on Figure 1). When X transmits message m to Y , node A also overhears m, in particular, he can see the frame(s) belonging to m. A intends to modify message m. Here, we must note that most messages originated from the routing sublayer are composed of only one frame per message in the data-link layer due to performance reasons, especially when they are used to discover routing topology. Upon reception of the frame corresponding to the message, the adversary can corrupt the frame check sequence by jamming once the data field of the frame has been received. This causes node Y to drop the frame (and the message), since Y detects that the last frame is incorrect , and waits for retransmission. At this point, if some acknowledgement mechanism is in use, A should send an acknowledgement to X so that it does not re-send the original frame. In addition, A retransmits message m in the name of X, where m is the modified message. The feasibility of jamming attacks is studied and demonstrated in [17]. Although, the authors conclude in that paper that the success of jamming attacks mainly depend on the distance of the honest nodes and the jammer node, various jamming techniques has been presented there that can severely interfere with the normal operation of the network. 2.2 Network model We assume that each honest device has exactly one antenna in the network. If the adversary uses several antennas we represent each of them by a distinct node. The network nodes are considered to be static, and we further assume that there is a single base station in the network. Let us denote the honest nodes in the network by v 0 , v 1 , . . . , v k , where v 0 denotes the base station. Similarly, v k+1 , . . . , v k+m represent the adversarial nodes. The set of all nodes is denoted by V . Furthermore, n denotes the number of all nodes in the network, i.e., n = |V | = k + m + 1. For each pair of nodes v i and v j , we define e v i ,v j to be the energy level needed to transmit a message from v i to v j , where v i , v j V . This values can be ordered in a matrix with size n n, called reachability matrix, and it is denoted by E. 1 In the rest, if we intend to emphasize the distinction between the honest and the adversarial nodes in the notation , we prefer to denote the adversarial nodes by v 1 , . . . , v m (where v = v k+ , 1 m). For the sake of simplicity, we also assume that at least energy e v i ,v j is needed for node v i to interfere with node v j 's packet reception. This means that if v i can reach v j , then v i can also interfere with all the communication of v j . Let us assume that each node uses a globally unique identifier in the network, and these identifiers are authenticated in some way (e.g., by symmetric keys). We denote the set of these identifiers by L, and there is a function L : V L {undef} that assigns an identifier to each node, where undef / L. According to our adversary model described in Subsection 2.1, we assume that the adversary has no (authenticated) identifier in the network, i.e., L(v j ) = undef for all 1 j m. We also introduce a cost function C : V R, which assigns a cost to each node (e.g., the remaining energy in the battery, or constant 1 to each node in order to represent hop-count). Configuration: A configuration conf is a quadruple (V, L, E, C) that consists of the set of nodes, the labelling function, the reachability matrix, and the cost function of nodes. 2.3 Security objective function Diverse sensor applications entail different requirements for routing protocols. For instance, remote surveillance applications may require minimal delay for messages, while sensor applications performing some statistical measurements favour routing protocols prolonging network lifetime. The diversity of routing protocols is caused by these conflicting requirements: e.g., shortest-path routing algorithms cannot maximize the network lifetime, since always choosing the same nodes to forward messages causes these nodes to run out of their energy supply sooner. Several sensor routing protocols use a trade-off to satisfy conflicting requirements [16, 11]. This small argument also points out that one cannot judge the utility of all routing protocols uniformly. Without a unified metric of utility we cannot refine our security objectives for routing protocols. By the above argument, a routing protocol that is secure against attacks aiming at decreasing network-lifetime cannot be secure against attacks aiming at increasing network delay. We model the negatively correlated requirements of routing, and essentially, our security objectives in a very general manner. We represent the output of a routing protocol, which is actually the ensemble of the routing entries of the honest nodes, with a given con-1 In this paper, the rows and the columns of all matrices are numbered from zero. 51 Node X Node A 2 Node A 1 Node Y 1: m 1: jam 1: m 2: m Node X Node Y Node A 1: m 1: m 1: jam 2: m Figure 1: Message modification performed by the cooperation of two adversarial nodes A 1 and A 2 (on the right-hand side) in Scenario 1, and employing overhearing, jamming, and relaying with a single adversarial node A (on the left-hand side) in Scenario 2. Honest nodes are labelled by X and Y . Arrows between nodes illustrate the direction of communication, the sequence of message exchanges are also depicted on these arrows. Dashed arrows illustrate failed message delivery caused by jamming. figuration conf by a matrix T conf with size k + 1 k + 1. 2 T conf i,j = 1, if honest node v i sends every message to an honest node identified by L(v j ) in order to deliver the message to the base station, otherwise let T conf i,j be 0. In the rest of the paper, we shortly refer to the result of a routing protocol with a given configuration as a routing topology, which can be considered as a directed graph described by matrix T conf . In the following, we will omit the index conf of T when the configuration can be unambiguously determined in a given context. In fact, T conf is a random variable, where the ran-domness is caused by the sensor readings initiated randomly by the environment, processing and transmission time of the sensed data, etc. Let us denote the set of all configurations by G. Furthermore , T denotes the set of the routing topologies of all configurations . The security objective function F : G T R assigns a real number to a random routing topology of a configuration . This function intends to distinguish "attacked" topologies from "non-attacked" topologies based on a well-defined security objective. We note that the definition of F is protocol dependent. For example, let us consider routing protocols that build a routing tree, where the root is the base station. We can compare routing trees based on network lifetime by the following security objective function F(conf , T conf ) = 1 k k X i=1 E(v i , conf , T conf ) where E : V G T R assigns the overall energy consumption of the path from a node v i to v 0 (the base station) in a routing tree of a configuration. Since T conf is a random variable, the output of F is a random variable too. If the distribution of this output in the presence of an attacker non-negligibly differs from the distribution when there's no attacker, then the protocol is not secure. If we intend to compare routing trees based on network delay a simple security objective function may be F(conf , T conf ) = 1 k k X i=1 M(v i , conf , T conf ) where M : V G T R assigns the length of the path from a node to v 0 in a routing topology of a configuration. 2 Of course, here we only consider the result of the protocol with respect to the honest nodes, since the adversarial nodes may not follow the protocol rules faithfully. 2.4 Dynamic model Following the simulation paradigm, we define a real-world model and an ideal-world model. The real-world model represents the real operation of the protocol and the ideal-world model describes how the system should work ideally. Both models contain an adversary. The real-world adversary is not constrained apart from requiring it to run in time polynomial . This enables us to be concerned with arbitrary feasible attacks. In addition, the ideal-world adversary is constrained in a way that it cannot modify messages and inject extra ones due to the construction of the ideal-world system . In other words, all attacks that modify or inject any messages is unsuccessful in the ideal-world system. However , the ideal-world adversary can perform attacks that are unavoidable or very costly to defend against (e.g., message deletion). Once the models are defined, the goal is to prove that for any real-world adversary, there exist an ideal-world adversary that can achieve essentially the same effects in the ideal-world model as those achieved by the real-world adversary in the real-world model (i.e., the ideal-world adversary can simulate the real-world adversary). 2.4.1 Real-world model The real-world model that corresponds to a configuration conf = (V, L, E, C) and adversary A is denoted by sys real conf ,A , and it is illustrated on Figure 2. We model the operation of the protocol participants by interactive and probabilistic Turing machines. Correspondingly, we represent the adversary , the honest sensor nodes, and the broadcast nature of the radio communication by machines A, M i , and C, respectively . These machines communicate with each other via common tapes. Each machine must be initialized with some input data (e.g., cryptographic keys, reachability matrix, etc.), which determines its initial state. Moreover, the machines are also provided with some random input (the coin flips to be used during the operation). Once the machines have been initialized , the computation begins. The machines operate in a reactive manner, i.e., they need to be activated in order to perform some computation. When a machine is activated , it reads the content of its input tapes, processes the received data, updates its internal state, writes some output on its output tapes, and goes back to sleep. The machines are activated in rounds by a hypothetic scheduler, and each machine in each round is activated only once. The order of activation is arbitrary with the only restriction that C must be activated at the end of the rounds. 52 Now, we present the machines in more details: Machine C. This machine is intended to model the radio communication. It has input tapes out i and out j , from which it reads messages written by M i and A, resp. It also has output tapes in i and in j , on which it writes messages to M i and A, resp. C is also initialized by matrix E at the beginning of the computation. Messages on tape out i can have the format ( sndr , cont , e, dest), where sndr L is the identifier of the sender, cont is the message content, e is the energy level to be used to determine the range of transmission, and dest is the identifier of the intended destination dest L {}, where indicates broadcast message. Messages on tape out j can have the following formats: (MSG, sndr , cont, e, dest ): MSG message models a normal broadcast message sent by the adversary to machine C with sender identifier sndr L, message content cont , energy level e, and identifier of the intended destination dest L {}. (JAM, e): Special JAM message, that is sent by the adversary to machine C, models the jamming capability of the adversary. When machine C receives a message JAM, it performs the requested jamming by deleting all messages in the indicated range e around the jamming node, which means that those deleted messages are not delivered to the nodes (including the jammer node itself) within the jamming range. (DEL, tar , e): Special DEL message, that is sent by the adversary to machine C, models the modification capability of the adversary. When receiving a message DEL with identifier tar L, machine C does not deliver any messages sent by node v V , where L(v ) = tar , if v is within the indicated range e, except the adversarial node itself that will receive the deleted messages. This models the sophisticated jamming technique that we described in Subsection 2.1. In a more formal way, when reading a message msg in = (MSG, sndr , cont, e, dest) from out j , C determines the nodes which receive the message by calculating the set of nodes V e V , such that for all v V e e v j ,v e. Finally, C processes msg in as follows. 1. if dest L {}, then C writes msg out = ( sndr , cont, dest ) to the input tapes of machines corresponding to honest nodes in V e msg out = (MSG, sndr , cont, dest ) to the input tapes of machines corresponding to adversarial nodes in V e \ {v j } 2. otherwise C discards msg in When reading a message msg in = (JAM, e) from out j , C determines the set of nodes which receive the message by calculating V e V , such that for all v V e e v j ,v e. Afterwards, C does not write any messages within the same round to the input tapes of machines corresponding to V e . When reading a message msg in = (DEL, tar , e) from out j , C determines the set of nodes which receive the message by calculating V e V , such that for all v V e e v j ,v e. Finally, C processes msg in as follows. 1. if there exists v x V e (1 x k), such that L(v x ) = tar , then C does not write any messages within the same round from tape out x to the input tapes of machines corresponding to V e \ {v j } 2. otherwise C discards msg in When reading a message msg in = ( sndr , cont, e, dest ) from out i , C determines the set of nodes which receive the message by calculating V e V , such that for all v V e e v j ,v e. Finally, C processes msg in as follows. 1. if dest L {}, then C writes msg out = ( sndr , cont, dest) to the input tapes of machines corresponding to honest nodes in V e \ {v i } msg out = (MSG, sndr , cont, dest ) to the input tapes of machines corresponding to adversarial nodes in V e 2. otherwise C discards msg in Machine M i . This machine models the operation of honest sensor nodes, and it corresponds to node v i . It has input tape in i and output tape out i , which are shared with machine C. The format of input messages must be ( sndr , cont, dest ), where dest L {}. The format of output messages must be ( sndr , cont, e, dest ), where sndr must be L(v i ), dest L {}, and e indicates the transmission range of the message for C. When this machine reaches one of its final states or there is a time-out during the computation process, it outputs its routing table. Machine A. This machine models the adversary logic. Encapsulating each adversarial node into a single machine allows us to model wormholes inside A. One can imagine that the adversary deploy several antennas in the network field, which are connected to a central adversary logic. In this convention, node v j corresponds to an adversarial antenna, which is modelled by input tape in j and output tape out j . These tapes are shared with machine C. The format of input messages must be msg in = (MSG, sndr , cont , e, dest), where dest L {}. The format of output messages msg out can be (MSG, sndr , cont, e, dest), where dest L {} and e indicates the transmission range of the message ; (JAM, e), where e indicates the range of jamming; (DEL, tar , e), where e indicates the range of selective jamming, and tar L. The computation ends, when all machines M i reach their final states, or there is a time-out. The output of sys real conf ,A is the value of the security objective function F applied to the resulted routing topology defined in Subsection 2.3 and configuration conf . The routing topology is represented by 53 C M 0 ... in 0 out 0 in 1 out 1 in 2 out 2 A M 1 out 1 in 1 M k out k in k in m out m ... C M 0 ... in 0 out 0 in 1 out 1 in 2 out 2 M 1 out 1 in 1 M k out k in k in m out m ... A Figure 2: The real-world model (on the left-hand side) and the ideal-world model (on the right-hand side). the ensemble of the routing entries of machines M i . We denote the output by Out real,F conf ,A (r), where r is the random input of the model. In addition, Out real,F conf ,A will denote the random variable describing Out real,F conf ,A (r) when r is chosen uniformly at random. 2.4.2 Ideal-world model The ideal-world model (illustrated on Figure 2) that corresponds to a configuration conf = (V, L, E, C) and adversary A is denoted by sys ideal conf ,A . The ideal-world model is identical to the real-world model with the exception that the ideal-world adversary cannot modify and inject extra messages . However, he is allowed to simply drop any messages or perform jamming, since these attacks are unavoidable, or at least, they are too costly to defend against. Our model is considered to be ideal in this sense. Comparing to the real-world model, we replace machine C with machine C and machine A with machine A in order to implement our restricted ideal-world adversary. Hence, we only detail the operation of C and A here, since M i are the same as in the real-world model. Receiving an MSG message from machines M i , C internally stores that message with a unique message identifier in its internal store. Delivering any MSG message to A , C also includes the message identifier into the message. A can send an MSG message to C with a different format; it only contains an identifier id and an energy level e. Upon the reception of such a message, C searches for the original message which is associated with identifier id in its internal store, and delivers this stored message using the energy level e. Although A also receives the original message with its associated identifier from C , he is not able to modify that, since C only accepts a message identifier issued by himself and an energy level from A . In other words, A can only delete messages, since A can also send special DEL and JAM messages to C . We elaborate the operation of C and A in a more formal way as follows. A and C communicate via tapes in j and out j . Machine C . It has input tapes out i and out j , from which it reads messages written by M i and A, resp. It also has output tapes in i and in j , on which it writes messages to M i and A, resp. C is also initialized by matrix E. In addition, it sets its internal variable id C to 1 at the beginning of the computation. C interacts with machines M i in a similar way as C does in the real-world model; when reading a message msg in = ( sndr , cont, e, dest ) from out i , C processes msg in identically to C in the real-world model only with one exception: Before writing msg in = (MSG, id C , sndr , cont, dest) to output tapes in j , C internally stores msg in in set S. After writing msg in to output tapes in j , C increments id C by one. Therefore , C knows what messages are passed to A from M i . Messages on out j can have the formats: (MSG, id, e): MSG message models a normal broadcast message sent by the ideal-world adversary to machine C , where e indicates the transmission range of the message identified by id. (JAM, e): Special JAM message, that is sent by the adversary to machine C, models the jamming capability of the ideal-world adversary, where e indicates the range of jamming. (DEL, tar , e): Special DEL message, that is sent by the adversary to machine C, models the modification capability of the ideal-world adversary, where e indicates the range of selective jamming, and tar L. When reading a message msg in = (MSG, id, e) from out j , machine C operates differently from C. C determines the set of nodes which receive the message by calculating V e V , such that for all v V e e v j ,v e. Finally, C processes msg in as follows. 1. if 1 id id C , then C searches the msg = (MSG, id , sndr , cont , dest ) in S such that id equals to id, and C writes msg out = ( sndr , cont , dest ) to the input tapes of machines corresponding to honest nodes in V e msg out = (MSG, id , sndr , cont , dest ) to the input tapes of machines corresponding to adversarial nodes in V e \ {v j } 2. otherwise C discards msg in When reading a message msg in = (JAM, e) or msg in = (DEL, tar , e) from out j , machine C operates the same 54 way as C does in case of the corresponding message formats. Machine A . It has output tapes out j and input tapes in j . The format of messages on input tape in j must be msg in = (MSG, id, sndr , cont, e, dest), where dest L {}. The format of output messages msg out can be (MSG, id, e), where id is a message identifier and e indicates the transmission range of the message identified by id; (JAM, e), where e indicates the range of jamming; (DEL, tar , e), where e indicates the range of selective jamming, and tar L. The computation ends, when all machines M i reach their final states, or there is a time-out. Similar to the real-world model, the output of sys ideal conf ,A is the value of the security objective function F applied to the resulted routing topology and configuration conf . The routing topology is represented by the ensemble of the routing entries of machines M i . We denote the output by Out ideal,F conf ,A (r), where r is the random input of the model. Moreover, Out ideal,F conf ,A will denote the random variable describing Out ideal,F conf ,A (r) when r is chosen uniformly at random. 2.5 Definition of secure routing Let us denote the security parameter of the model by (e.g., is the key length of the cryptographic primitive employed in the routing protocol, such as digital signature, MAC, etc.). Based on the model described in the previous subsections, we define routing security as follows: Definition 1 (Statistical security). A routing protocol is statistically secure with security objective function F, if for any configuration conf and any real-world adversary A, there exists an ideal-world adversary A , such that Out real,F conf ,A is statistically indistinguishable from Out ideal,F conf ,A . Two random variables are statistically indistinguishable if the L 1 distance of their distributions is a negligible function of the security parameter . Intuitively, if a routing protocol is statistically secure, then any system using this routing protocol cannot satisfy its security objectives represented by function F only with a probability that is a negligible function of . This negligible probability is related to the fact that the adversary can always forge the cryptographic primitives (e.g., generate a valid digital signature) with a very small probability depending on the value of . INSECURITY OF TINYOS ROUTING In this section, we present an authenticated routing mechanism based on the well-known TinyOS routing, and we show that it is not secure in our model for a given security objective function representing a very minimal security requirement. 3.1 Operation of an authenticated routing protocol Originally, the authors of TinyOS implemented a very simple routing protocol, where each node uses a globally unique identifier. The base station periodically initiates a routing topology discovery by flooding the network by a beacon message. Upon reception of the first beacon within a single beaconing interval, each sensor node stores the identifier of the node, from which it received the beacon, as its parent (aka. next-hop towards the base station), and then re-broadcasts the beacon after changing the sender identifier to its own identifier. As for each node only one parent is stored, the resulted routing topology is a tree. Every sensor node receiving a data packet forwards that towards the base station by sending the packet to its parent. A lightweight cryptographic extension is employed in [14] in order to authenticate the beacon by the base station. This authenticated variant of TinyOS routing uses Tesla scheme to provide integrity for the beacon; each key is disclosed by the next beacon in the subsequent beaconing interval. We remark that this protocol has only been defined informally that inspired us to present a new protocol, which provides the "same" security as the authenticated routing protocol in [14], but due to its simplicity it fits more in demonstrating the usage of our model. Consequently, the presented attack against this new protocol also works against the protocol in [14]. We must note again that this protocol is only intended to present the usefulness of our model rather than to be considered as a proposal of a new sensor routing protocol. We assume that the base station B has a public-private key pair, where the public key is denoted by K pub . Furthermore , it is assumed that each sensor node is also deployed with K pub , and they are capable to perform digital signature verification with K pub as well as to store some beacons in its internal memory. We note that B never relays messages between sensor nodes. Initially, B creates a beacon, that contains a constant message identifier BEACON, a randomly generated number rnd, the identifier of the base station Id B , and a digital signature sig B generated on the previous elements except Id B . Afterwards , the base station floods the network by broadcasting this beacon: B : msg 1 = (BEACON, rnd, Id B , sig B ) Each sensor node X receiving msg 1 checks whether it has already received a beacon with the same rnd in conjunction with a correct signature before. If it is true, the node discards msg 1 , otherwise it verifies sig B . If the verification is successful, then X sets Id B as its parent, stores msg 1 in its internal memory, and re-broadcasts the beacon by changing the sender identifier Id B to its own identifier Id X : X : msg 2 = (BEACON, rnd, Id X , sig B ) If the signature verification is unsuccessful, then X discards msg 1 . Every sensor node receiving msg 2 performs the same steps what X has done before. Optionally, B can initiate this topology construction periodically by broadcasting a new beacon with different rnd. In the rest, we shortly refer to this protocol as ABEM (Authenticated Beaconing Mechanism). 3.2 Formalization of a simple attack A simple security objective is to guarantee the correctness of all routing entries in the network. Namely, it is desirable that a sender node v i is always able to reach node v j , if v i set L(v j ) as its parent identifier earlier. It means that if 55 node v i sets node L(v j ) as its parent identifier, then E i,j should contain a finite value, or v i as well as v j should have an adversarial neighboring node v 1 and v 2 , resp., such that E i,k+ 1 and E k+ 2 ,j are finite values, where 1 1 , 2 m and 1 = 2 may hold. In order to formalize this minimal security requirement, we introduce the following security objective function F ABEM (conf , T ) = 8 &lt; : 1, if i, j : T i,j E i,j `Q m=1 E i,k+ + Q m=1 E k+ ,j = 0 0, otherwise where we derive matrix E with size n n from E, so that E i,j = 1, if E i,j = , otherwise E i,j = 0. In other words, E i,j = 1, if v i cannot send a message directly to v j , otherwise E i,j = 0. We will show that ABEM is not secure in our model for security objective function F ABEM . In particular, we present a configuration conf and an adversary A, for which there doesn't exist any ideal-world adversary A , such that Out real,F ABEM conf ,A is statistically indistinguishable from Out ideal,F ABEM conf ,A . Equivalently, we show that for a real-world adversary A, F ABEM (conf , T ) = 0 with a probability that is a non-negligible function of in the real-world model, while F ABEM (conf , T ) = 0 with probability zero for every ideal-world adversary A in the ideal-world model, where T describes the routing topology in the ideal-world model. Moreover, the success probability of the real-world adversary A described below is independent from . v 0 , B v 1 , X v 2 , Y v 1 = v 3 v 0 , B v 1 , X v 2 , Y v 1 = v 3 Figure 3: A simple attack against ABEM. v 0 , v 1 , and v 2 are honest nodes with identifiers L(v 0 ) = B, L(v 1 ) = X, and L(v 2 ) = Y , whereas v 1 is an adversarial node. E 1,0 , E 3,0 , E 2,3 are finite values, and E 3,1 = E 2,0 = E 2,1 = . Links are assumed to be symmetric, i.e., E i,j = E j,i . The configuration is illustrated on the left-hand side, where a dashed line denotes a direct link. In the routing topology of the real-world model, on the right-hand side, v 2 sets X as its parent identifier, however, E 2,1 = and E 3,1 = . The configuration conf and the result of the attack is depicted on Figure 3. We assume that the base station broadcasts only a single beacon during the computational process, i.e., only a single beaconing interval is analyzed in our model. At the beginning, the base station B floods the network by a beacon B : msg 1 = (BEACON, rnd, B, sig B ) Both adversarial node v 1 and honest node X receive this beacon, and X sets B as its parent, since the verification of the signature is successful. X modifies the beacon by replacing sender identifier B to X, and broadcasts the resulted beacon: X : msg 2 = (BEACON, rnd, X, sig B ) In parallel, v 1 modifies the beacon by replacing sender identifier B to X, and broadcasts the resulted beacon: v 1 : msg 2 = (BEACON, rnd, X, sig B ) Upon the reception of msg 2 , node Y sets X as its parent, since sig B is correct. In the real-world model, these actions result T 2,1 = 1, which implies that F ABEM (conf , T ) = 0. On the contrary, F ABEM (conf , T ) never equals to 0, where T represents the routing topology in the ideal-world model. Let us assume that F ABEM (conf , T ) = 0, which means that T 1,2 = 1 or T 2,1 = 1. T 1,2 = 1 is only possible, if X receives msg 3 = (BEACON, rnd, Y, sig B ) However, it yields contradiction, since E 3,1 = E 2,1 = , and B never broadcasts msg 3 . Similarly, if T 2,1 = 1 then Y must receive msg 2 , which means that v 1 must broadcast msg 2 . Conversely, B never broadcasts msg 2 , and E 3,1 = . Therefore, v 1 can only broadcast msg 2 , if he successfully modifies msg 1 or forges msg 2 . However, it also contradicts our assumption that the ideal-world adversary cannot modify and inject messages in the ideal-world model. RELATED WORK In [10], the authors map some adversary capabilities and some feasible attacks against routing in wireless sensor networks , and they define routing security implicitly as resistance to (some of) these attacks. Hence, the security of sensor routing is only defined informally, and the countermeasures are only related to specific attacks. In this way, we even cannot compare the sensor routing protocols in terms of security. Another problem with this approach is the lack of a formal model, where the security of sensor routing can be described in a precise and rigorous way. While secure messaging and key-exchange protocols are classical and well-studied problems in traditional networks [3, 15], formal modelling of secure routing in sensor networks has not been considered so far. The adversarial nodes are also classified into the groups of sensor-class and laptop-class nodes in [10], but the capabilities of an adversarial node regarding message manipulations are not discussed. The simulation paradigm is described in [15, 5]. These models were mainly proposed with wired networks in mind typically implemented on the well-known Internet architecture , and the wireless context is not focused there. In our opinion, the multi-hop nature of communications is an inherent characteristic of wireless sensor networks, therefore, it should be explicitly modelled. In more particular, the broadcast nature of communication enables a party to overhear the transmission of a message that was not destined to him, however, this transmission can be received only in a certain range of the sender. The size of this range is determined by the power at which the sender sent the message. Another deviation from [15] is the usage of the security objective function in the definition of security. In [15], the 56 indistinguishability is defined on the view of the honest parties (on their input, states, and output) in the ideal-world and in the real-world models. However, an adversary can distort the states of the honest parties in unavoidable ways, and hence, the classical definition would be too strong and no routing protocol would satisfy it. On the other hand, our model is compliant with [15] considering high-level connections between nodes. In [15], the standard cryptographic system allows us to define each high-level connection as secure (private and authentic), authenticated (only authentic ), and insecure (neither private nor authentic). In this taxonomy, the communication channel between two honest nodes can be either insecure or secure in our model. If an adversarial node is placed in the communication range of one of the communicating nodes, then it is considered to be an insecure channel. If the adversary can reach none of the communicating nodes, the channel between that nodes is hidden from the adversary, and thus, it is considered to be secure. Although some prior works [18, 12] also used formal techniques to model the security of multi-hop routing protocols , these ones were mainly proposed for ad hoc routing. Moreover, the model proposed in [12] is based on CPAL-ES, and the model in [18] is similar to the strand spaces model. Both of these formal techniques differ from the simulation paradigm. Our work is primarily based on [4, 1]. Here, the authors also use the simulation paradigm to prove the security of routing protocols in wireless ad-hoc networks. However, our model differs from the models in [4, 1] in two ways: Adversary model: The adversary in [4] and [1] is assumed to have the same resources and communication capabilities as an ordinary node in the network. Therefore, that adversary model deviates from the so-called Dolev-Yao model in [6]. In our work, the adversary also uses wireless devices to attack the systems, and it is reasonable to assume that the adversary can interfere with communications only within its power range. The adversarial nodes belonging to the sensor-class nodes has the same resources and communication capabilities as an ordinary sensor node, but a more resourced adversarial node (e.g., laptops) may affect the overall communication of an entire part of the network depending on the power range of the resourced adversarial device. That resourced devices also make the adversary able to perform more sophisticated message manipulations. Modelling security objectives: In ad hoc networks, nodes construct routes between a source and a destination [13, 8], whereas sensor nodes should build a complete routing topology for the entire network. In case of sensor networks, the only destination for all nodes is the base station [9]. In addition, sensor nodes are resource constrained, which implies that we also need to model the energy consumption of sensor nodes, since several attacks impacts the network lifetime. These differences from ad hoc networks has yielded a wide range of sensor applications, and thus, sensor routing protocols [9] are much diverse than ad hoc routing protocols. Hence, the security objectives cannot be modelled uniformly for sensor routing protocols. CONCLUSION In this paper, we proposed a formal security model for routing protocols in wireless sensor networks. Our model is based on the well-known simulation paradigm, but it differs from previously proposed models in several important aspects . First of all, the adversary model is carefully adopted to the specific characteristics of wireless sensor networks. In our model, the adversary is not all-powerful, but it can only interfere with communications within its own radio range. A second important contribution is that we defined the output of the dynamic models that represent the ideal and the real operations of the system as a suitable function of the routing state of the honest nodes, instead of just using the routing state itself as the output. We expect that this will allow us to model different types of routing protocols in a common framework. In addition, this approach hides the unavoidable distortions caused by the adversary in the routing state, and in this way, it makes our definition of routing security satisfiable. As an illustrative example, we considered an authenticated version of the TinyOS beaconing routing protocol, and we showed how an attack against this protocol can be represented in our formal model. As we mentioned in the Introduction, this paper is a workin -progress paper. In particular, we have presented neither a new secure routing protocol designed with the help of our formal model, nor a detailed security proof carried out within our model. These are left for future study. We must note, however, that the generality of the simulation paradigm and the fact that we could represent a known attack against the authenticated TinyOS protocol in our model make us confident that we are on the right track. ACKNOWLEDGEMENTS The work described in this paper is based on results of IST FP6 STREP UbiSec&Sens ( http://www.ist-ubisecsens.org). UbiSec&Sens receives research funding from the European Community's Sixth Framework Programme. Apart from this, the European Commission has no responsibility for the content of this paper. The information in this document is provided as is and no guarantee or warranty is given that the information is fit for any particular purpose. The user thereof uses the information at its sole risk and liability. The work presented in this paper has also been partially supported by the Hungarian Scientific Research Fund (contract number T046664). The first author has been further supported by the HSN Lab. The second author has been supported by the Hungarian Ministry of Education (B O2003/70). REFERENCES [1] G. Acs, L. Butty an, and I. Vajda. Provable Security of On-Demand Distance Vector Routing in Wireless Ad Hoc Networks. In Proceedings of the Second European Workshop on Security and Privacy in Ad Hoc and Sensor Networks (ESAS 2005), July 2005. [2] G. Acs, L. Butty an, and I. Vajda. Provably Secure On-demand Source Routing in Mobile Ad Hoc Networks. To appear in IEEE Transactions on Mobile Computing. [3] M. Bellare, R. Canetti, and H. Krawczyk. A modular approach to the design and analysis of authentication 57 and key exchange protocols. In Proceedings of the ACM Symposium on the Theory of Computing, 1998. [4] L. Butty an and I. Vajda. Towards provable security for ad hoc routing protocols. In Proceedings of the ACM Workshop on Security in Ad Hoc and Sensor Networks (SASN), October 2004. [5] R. Canetti. Universally composable security: A new paradigm for cryptographic protocols. In Proceedings of the 42nd IEEE Symposium on Foundations of Computer Science (FOCS), 2001. [6] D. Dolev and A. C. Yao. On the Security of Public Key Protocols. In IEEE Transactions on Information Theory 29 (2), 1983. [7] IEEE Standard for Information technology--Telecommunications and information exchange between systems--Local and metropolitan area networks--Specific requirements. Part 15.4: Wireless Medium Access Control (MAC) and Physical Layer (PHY) Specifications for Low-Rate Wireless Personal Area Networks (LR-WPANs), 2003. [8] D. Johnson and D. Maltz. Dynamic source routing in ad hoc wireless networks. In Mobile Computing, edited by Tomasz Imielinski and Hank Korth, Chapter 5, pages 153-181. Kluwer Academic Publisher, 1996. [9] J. N. Al-Karaki and A. E. Kamal. Routing techniques in wireless sensor networks: a survey. In IEEE Wireless Communications, Volume 11, pp. 6-28, 2004. [10] C. Karlof, D. Wagner. Secure routing in wireless sensor networks: attacks and countermeasures. In Ad Hoc Networks, Volume 1, 2003. [11] Q. Li, J. Aslam, and D. Rus. Hierarchical Power-aware Routing in Sensor Networks. In Proceedings of the DIMACS Workshop on Pervasive Networking, May, 2001. [12] J. Marshall. An Analysis of the Secure Routing Protocol for mobile ad hoc network route discovery: using intuitive reasoning and formal verification to identify flaws. MSc thesis, Department of Computer Science, Florida State University, April 2003. [13] C. Perkins and E. Royer. Ad hoc on-demand distance vector routing. In Proceedings of the IEEE Workshop on Mobile Computing Systems and Applications, pp. 90-100, February 1999. [14] A. Perrig, R. Szewczyk, V. Wen, D. Culler, J. D. Tygar. SPINS: Security Protocols for Sensor Networks. In Wireless Networks Journal (WINE), Volume 8, September 2002. [15] B. Pfitzman and M. Waidner. A model for asynchronous reactive systems and its application to secure message transmission. In Proceedings of the 22nd IEEE Symposium on Security & Privacy, 2001. [16] S. Singh, M. Woo, and C. Raghavendra. Power-Aware Routing in Mobile Ad Hoc Networks. In Proceedings of the Fourth Annual ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom '98), Oct. 1998. [17] W. Xu, W. Trappe, Y. Zhang and T. Wood. The Feasibility of Launching and Detecting Jamming Attacks in Wireless Networks. In Proceedings of the 6th ACM International Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc'05), May 2005. [18] S. Yang and J. Baras. Modeling vulnerabilities of ad hoc routing protocols. In Proceedings of the ACM Workshop on Security of Ad Hoc and Sensor Networks, October 2003. 58

Simulatability;Adversary Model;Routing Protocols;Sensor Networks;Provable Security

142

Obfuscated Databases and Group Privacy

We investigate whether it is possible to encrypt a database and then give it away in such a form that users can still access it, but only in a restricted way. In contrast to conventional privacy mechanisms that aim to prevent any access to individual records, we aim to restrict the set of queries that can be feasibly evaluated on the encrypted database. We start with a simple form of database obfuscation which makes database records indistinguishable from lookup functions . The only feasible operation on an obfuscated record is to look up some attribute Y by supplying the value of another attribute X that appears in the same record (i.e., someone who does not know X cannot feasibly retrieve Y ). We then (i) generalize our construction to conjunctions of equality tests on any attributes of the database, and (ii) achieve a new property we call group privacy. This property ensures that it is easy to retrieve individual records or small subsets of records from the encrypted database by identifying them precisely, but "mass harvesting" queries matching a large number of records are computationally infeasible. Our constructions are non-interactive. The database is transformed in such a way that all queries except those ex-plicitly allowed by the privacy policy become computationally infeasible, i.e., our solutions do not rely on any access-control software or hardware.

INTRODUCTION Conventional privacy mechanisms usually provide all-or-nothing privacy. For example, secure multi-party computation schemes enable two or more parties to compute some joint function while revealing no information about their respective inputs except what is leaked by the result of the computation. Privacy-preserving data mining aims to com-pletely hide individual records while computing global statistical properties of the database. Search on encrypted data and private information retrieval enable the user to retrieve data from an untrusted server without revealing the query. In this paper, we investigate a different concept of privacy. Consider a data owner who wants to distribute a database to potential users. Instead of hiding individual data entries, he wants to obfuscate the database so that only certain queries can be evaluated on it, i.e., the goal is to ensure that the database, after it has been given out to users, can be accessed only in the ways permitted by the privacy policy. Note that there is no interaction between the data owner and the user when the latter accesses the obfuscated database. Our constructions show how to obfuscate the database before distributing it to users so that only the queries permitted by the policy are computationally feasible. This concept of privacy is incomparable to conventional definitions because, depending on the policy, a permitted query may or even should reveal individual data entries. For example, a college alumni directory may be obfuscated in such a way that someone who already knows a person's name and year of graduation is able to look up that person's email address, yet spammers cannot indiscriminately harvest addresses listed in the directory. Employees of a credit bureau need to have access to customers' records so that they can respond to reports of fraudulent transactions, yet one may want to restrict the bureau's ability to compile a list of customers' addresses and sell it to a third party. We develop provably secure obfuscation techniques for several types of queries. We do not assume that users of the obfuscated database access it through a trusted third party, nor that they use trusted or "tamper-proof" access-control software or hardware (in practice, such schemes are vulnerable to circumvention and reverse-engineering, while trusted third parties are scarce and often impractical). Our constructions are cryptographically strong, i.e., they assume an adversary who is limited only by his computational power. We prove security in the standard "virtual black-box" model for obfuscation proposed by Barak et al. [2]. Intuitively , a database is securely obfuscated if the view of any efficient adversary with access to the obfuscation can be efficiently simulated by a simulator who has access only to the ideal functionality, which is secure by definition. The ideal functionality can be thought of as the desired privacy policy for the database. One of our contributions is coming up with several ideal functionalities that capture interesting privacy policies for databases. 102 Directed-access databases. Our "warm-up" construction is a directed-access database. Some attributes of the database are designated as query attributes, and the rest as data attributes. The database is securely obfuscated if, for any record, it is infeasible to retrieve the values of the data attributes without supplying the values of the query attributes, yet a user who knows the query attributes can easily retrieve the corresponding data attributes. To illustrate by example, a directed-access obfuscation of a telephone directory has the property that it is easy to look up the phone number corresponding to a particular name-address pair, but queries such as "retrieve all phone numbers stored in the directory" or "retrieve all names" are computationally infeasible. Such a directory is secure against abusive harvesting, but still provides useful functionality . Note that it may be possible to efficiently enumerate all name-address pairs because these fields have less entropy than regular cryptographic keys, and thus learn the entire database through the permitted queries. Because the database is accessed only in permitted ways, this does not violate the standard definition of obfuscation. Below, we give some examples where it is not feasible to enumerate all possible values for query attributes. The directed-access property of a single database record can be modeled as a point function, i.e., a function from {0, 1} n to {0, 1} that returns 1 on exactly one input x (in our case, query attributes are the arguments of the point function). Directed-access obfuscation guarantees that the adversary's view of any obfuscated record can be efficiently simulated with access only to this point function. Therefore , for this "warm-up" problem, we can use obfuscation techniques for point functions such as [22]. Informally, we encrypt the data attributes with a key derived from hashed query attributes. The only computationally feasible way to retrieve the data attributes is to supply the corresponding query attributes. If the retriever does not know the right query attributes, no information can be extracted at all. Group-exponential databases. We then consider a more interesting privacy policy, which requires that computational cost of access be exponential in the number of database records retrieved. We refer to this new concept of privacy as group privacy. It ensures that users of the obfuscated database can retrieve individual records or small subsets of records by identifying them precisely, i.e., by submitting queries which are satisfied only by these records. Queries matching a large number of records are infeasible. We generalize the idea of directed access to queries consisting of conjunctions of equality tests on query attributes, and then to any boolean circuit over attribute equalities. The user can evaluate any query of the form attribute j 1 = value 1 . . .attribute j t = value t , as long as it is satisfied by a small number of records. Our construction is significantly more general than simple keyword search on encrypted data because the value of any query attribute or a conjunction thereof can be used as the "keyword" for searching the obfuscated database, and the obfuscator does not need to know what queries will be evaluated on the database. To distinguish between "small" and "large" queries, suppose the query predicate is satisfied by n records. Our construction uses a form of secret sharing that forces the retriever to guess n bits before he can access the data attributes in any matching record. (If n=1, i.e., the record is unique, the retriever still has to guess 1 bit, but this simply means that with probability 1 2 he has to repeat the query.) The policy that requires the retriever to uniquely identify a single record, i.e., forbids any query that is satisfied by multiple records, can also be easily implemented using our techniques. Our construction can be viewed as the noninteractive analog of hash-reversal "client puzzles" used to prevent denial of service in network security [21], but, unlike client puzzles, it comes with a rigorous proof of security. For example, consider an airline passenger database in which every record contains the passenger's name, flight number, date, and ticket purchase details. In our construction , if the retriever knows the name and date that uniquely identify a particular record (e.g., because this information was supplied in a court-issued warrant), he (almost) immediately learns the key that encrypts the purchase details in the obfuscated record. If the passenger traveled on k flights on that date, the retriever learns the key except for k bits. Since k is small, guessing k bits is still feasible. If, however, the retriever only knows the date and the flight number, he learns the key except for m bits, where m is the number of passengers on the flight, and retrieval of these passengers' purchase details is infeasible. A database obfuscated using our method has the group privacy property in the following sense. It can be accessed only via queries permitted by the privacy policy. The probability of successfully evaluating a permitted query is inversely exponential in the number of records that satisfy the query predicate. In particular, to extract a large number of records from the database, the retriever must know a pri-ori specific information that uniquely identifies each of the records, or small subsets thereof. The obfuscated database itself does not help him obtain this information. In obfuscated databases with group privacy, computational cost of access depends on the amount of information retrieved. Therefore, group privacy can be thought of as a step towards a formal cryptographic model for "economics of privacy." It is complementary to the existing concepts of privacy, and appears to be a good fit for applications such as public directories and customer relationship management (CRM) databases, where the database user may need to access an individual record for a legitimate business purpose, but should be prevented from extracting large subsets of records for resale and abusive marketing. While our constructions for group privacy are provably secure in the "virtual black-box" sense of [2], the cost of this rigorous security is a quadratic blowup in the size of the obfuscated database, rendering the technique impractical for large datasets. We also present some heuristic techniques to decrease the size of the obfuscated database, and believe that further progress in this area is possible. Alternative privacy policies. Defining rigorous privacy policies that capture intuitive "database privacy" is an important challenge, and we hope that this work will serve as a starting point in the discussion. For example, the group privacy policy that we use in our constructions permits the retriever to learn whether a given attribute of a database record is equal to a particular value. While this leaks more information than may be desirable, we conjecture that the privacy policy without this oracle is unrealizable. We also consider privacy policies that permit any query rather than just boolean circuits of equality tests on attributes . We show that this policy is vacuous: regardless of the database contents, any user can efficiently extract 103 the entire database by policy-compliant queries. Therefore, even if the obfuscation satisfies the virtual black-box property , it serves no useful purpose. Of course, there are many types of queries that are more general than boolean circuits of equality tests on attributes. Exact characterization of non-vacuous, yet realizable privacy policies is a challenging task, and a topic of future research. Organization of the paper. We discuss related work in section 2. The ideas are illustrated with a "warm-up" construction in section 3. In section 4, we explain group privacy and the corresponding obfuscation technique. In section 5, we generalize the class of queries to boolean circuits over attribute equalities. In section 6, we show that policies which permit arbitrary queries are vacuous, and give an informal argument that a policy that does not allow the retriever to verify his guesses of individual attribute values cannot be realized . Conclusions are in section 7. All proofs will appear in the full version of the paper. RELATED WORK This paper uses the "virtual black-box" model of obfuscation due to Barak et al. [2]. In addition to the impossibility result for general-purpose obfuscation, [2] demonstrates several classes of circuits that cannot be obfuscated. We focus on a different class of circuits. To the best of our knowledge, the first provably secure constructions for "virtual black-box" obfuscation were proposed by Canetti et el. [5, 6] in the context of "perfectly one-way" hash functions, which can be viewed as obfuscators for point functions (a.k.a. oracle indicators or delta functions). Dodis and Smith [15] recently showed how to construct noise-tolerant "perfectly one-way" hash functions. which they used to obfuscate proximity queries with "en-tropic security." It is not clear how to apply techniques of [15] in our setting. In section 6, we present strong evidence that our privacy definitions may not be realizable if queries other than equality tests are permitted. Lynn et al. [22] construct obfuscators for point functions (and simple extensions, such as public regular expressions with point functions as symbols) in the random oracle model. The main advantage of [22] is that it allows the adversary partial information about the preimage of the hash function, i.e., secrets do not need to have high entropy. This feature is essential in our constructions, too, thus we also use the random oracle model. Wee [27] proposed a construction for a weaker notion of point function obfuscation, along with the impossibility result for uniformly black-box obfuscation. This impossibility result suggests that the use of random oracles in our proofs (in particular, the simulator's ability to choose the random oracle) is essential. Many ad-hoc obfuscation schemes have been proposed in the literature [1, 10, 9, 12, 13, 11]. Typically, these schemes contain neither a cryptographic definition of security, nor proofs, except for theoretical work on software protection with hardware restrictions on the adversary [19, 20]. Forcing the adversary to pay some computational cost for accessing a resource is a well-known technique for preventing malicious resource exhaustion (a.k.a. denial of service attacks). This approach, usually in the form of presenting a puzzle to the adversary and forcing him to solve it, has been proposed for combating junk email [16], website metering [17], prevention of TCP SYN flooding attacks [21], protecting Web protocols [14], and many other applications. Puzzles based on hash reversal, where the adversary must discover the preimage of a given hash value where he already knows some of the bits, are an especially popular technique [21, 14, 26], albeit without any proof of security. Our techniques are similar, but our task is substantially harder in the context of non-interactive obfuscation. The obfuscation problem is superficially similar to the problem of private information retrieval [3, 8, 18] and keyword search on encrypted data [25, 4]. These techniques are concerned, however, with retrieving data from an untrusted server, whereas we are concerned with encrypting the data and then giving them away, while preserving some control over what users can do with them. A recent paper by Chawla et al. [7] also considers database privacy in a non-interactive setting, but their objective is complementary to ours. Their definitions aim to capture privacy of data, while ours aim to make access to the database indistinguishable from access to a certain ideal functionality. DIRECTED-ACCESS DATABASES As a warm-up example, we show how to construct directed-access databases in which every record is indistinguishable from a lookup function. The constructions and theorems in this section are mainly intended to illustrate the ideas. Let X be a set of tuples x , Y a set of tuples y , and Y = Y {}. Let D X Y be the database. We want to obfuscate each record of D so that the only operation that a user can perform on it is to retrieve y if he knows x . We use the standard approach in secure multi-party computation , and formally define this privacy policy in terms of an ideal functionality. The ideal functionality is an (imaginary ) trusted third party that permits only policy-compliant database accesses. An obfuscation algorithm is secure if any access to the obfuscated database can be efficiently simulated with access only to the ideal functionality. This means that the user can extract no more information from the obfuscated database than he would be able to extract had all of his accesses been filtered by the trusted third party. Definition 1. (Directed-access privacy policy) For database D, define the corresponding directed-access functionality DA D as the function that, for any input x X such that x , y 1 , . . . , x , y m D, outputs { y 1 , . . . , y m }. Intuitively, a directed-access database is indistinguishable from a lookup function. Given the query attributes of an individual record ( x ), it is easy to learn the data attributes ( y ), but the database cannot be feasibly accessed in any other way. In particular, it is not feasible to discover the value of y without first discovering a corresponding x . Moreover, it is not feasible to harvest all y values from the database without first discovering all values of x . This definition does not say that, if set X is small, it is infeasible to efficiently enumerate all possible values of x and stage a dictionary attack on the obfuscated database. It does guarantee that even for this attack, the attacker is unable to evaluate any query forbidden by the privacy policy. In applications where X cannot be efficiently enumerated (e.g., X is a set of secret keywords known only to some users of the obfuscated database), nothing can be retrieved from the obfuscated database by users who don't know the keywords. Observe that x can contain multiple attributes, 104 and thus multiple keywords may be required for access to y in the obfuscated database. Directed-access databases are easy to construct in the random oracle model, since lookup functionality is essentially a point function on query attributes, and random oracles naturally provide an obfuscation for point functions [22]. The obfuscation algorithm OB da takes D and replaces every record x i , y i D with hash (r i 1 || x i ), hash(r i 2 || x i ) y i , r i 1 , r i 2 where r i 1,2 are random numbers, || is concatenation, and hash is a hash function implementing the random oracle. Theorem 1. (Directed-access obfuscation is "virtual black-box") Let OB da be as described above. For any probabilistic polynomial-time adversarial algorithm A, there exists a probabilistic polynomial-time simulator algorithm S and a negligible function of the security parameter k such that for any database D: |P(A(OB da (D)) = 1) - P(S DA D (1 |D| ) = 1)| (k) where probability P is taken over random oracles (implemented as hash functions), as well as the the randomness of A and S. Intuitively, this theorem holds because retrieving y i requires finding the (partial) pre-image of hash(r i 2 , x i ). The standard definition of obfuscation in [2] also requires that there exist an efficient retrieval algorithm that, given some x , extracts the corresponding y from the obfuscation OB da (D). Clearly, our construction has this property . Someone who knows x simply finds the record(s) in which the first value is equal to hash(r i 1 || x ), computes hash (r i 2 || x ) and uses it as the key to decrypt y . GROUP-EXPONENTIAL DATABASES For the purposes of this section, we restrict our attention to queries P that are conjunctions of equality tests over attributes (in section 5, we show how this extends to arbitrary boolean circuits over equality tests). For this class of queries, we show how to obfuscate the database so that evaluation of the query is exponential in the size of the answer to the query. Intuitively, this means that only precise query predicates, i.e., those that are satisfied by a small number of records, can be efficiently computed. "Mass harvesting" queries, i.e., predicates that are satisfied by a large number of records, are computationally infeasible. Recall that our goal is to restrict how the database can be accessed. For some databases, it may be possible to efficiently enumerate all possible combinations of query attributes and learn the entire database by querying it on every combination. For databases where the values of query attributes are drawn from a large set, our construction prevents the retriever from extracting any records that he cannot describe precisely. In either case, we guarantee that the database can be accessed only through the interface permitted by the privacy policy, without any trust assumptions about the retriever's computing environment. In our construction, each data attribute is encrypted with a key derived from a randomly generated secret. We use a different secret for each record. The secret itself is split into several (unequal) shares, one per each query attribute. Each share is then encrypted itself, using a key derived from the output of the hash function on the value of the corresponding query attribute. If the retriever knows the correct value only for some query attribute a, he must guess the missing shares. The size of the revealed share in bits is inversely related to the number of other records in the database that have the same value of attribute a. This provides protection against queries on frequently occurring attribute values. 4.1 Group privacy policy We define the privacy policy in terms of an ideal functionality , which consists of two parts. When given an index of a particular query attribute and a candidate value, it responds whether the guess is correct, i.e., whether this value indeed appears in the corresponding attribute of the original database. When given a predicate, it evaluates this predicate on every record in the database. For each record on which the predicate is true, it returns this record's data attributes with probability 2 -q , where q is the total number of records in the database that satisfy the predicate. if no more information can be extracted this ideal functionality. Definition 2. (Group privacy policy) Let X be a set and Y a set of tuples. Let D be the database 1 , 2 , . . . N where i = {x i 1 , x i 2 , . . . , x im , y i } X m Y. Let P : X m {0, 1} be a predicate of the form X j 1 = x j 1 X j 2 = x j 2 . . . X j t = x j t . Let D [P] = { i D | P(x i 1 , x i 2 , . . . , x im ) = 1} be the subset of records on which P is true. The group-exponential functionality GP D consists of two functions: - C D (x, i, j) is 1 if x = x ij and 0 otherwise, where 1 i N, 1 j m. - R D (P) = 1iN { i, i }, where i = y i with probability 2 -|D [P] | if P( i ) with probability 1 - 2 -|D [P] | if P( i ) if P( i ) Probability is taken over the internal coin tosses of GP D . Informally, function C answers whether the jth attribute of the ith record is equal to x, while function R returns all records that satisfy some predicate P, but only with probability inversely exponential in the number of such records. It may appear that function C is unnecessary. Moreover, it leaks additional information, making our privacy policy weaker than it might have been. In section 6, we argue that it cannot be simply eliminated, because the resulting functionality would be unrealizable. Of course, there may exist policies that permit some function C which leaks less information than C, but it is unclear what C might be. We discuss several alternatives to our definition in section 6. We note that data attributes are drawn from a set of tuples Y because there may be multiple data attributes that need to be obfuscated. Also observe that we have no restrictions on the values of query attributes, i.e., the same m -tuple of query attributes may appear in multiple records, with different or identical data attributes. 4.2 Obfuscating the database We now present the algorithm OB gp , which, given any database D, produces its obfuscation. For notational convenience , we use a set of random hash functions H : {0, 1} {0, 1} k . Given any hash function H, these can be implemented simply as H(||x). To convert the k-bit hash function output into a key as long as the plaintext to which it is 105 applied, we use a set of pseudo-random number generators prg , : {0, 1} k {0, 1} (this implements random oracles with unbounded output length). Let N be the number of records in the database. For each row i , 1 i N , generate a random N -bit secret r i = r i 1 ||r i 2 || . . . ||r iN , where r ij R {0, 1}. This secret will be used to protect the data attribute y i of this record. Note that there is 1 bit in r i for each record of the database. Next, split r i into m shares corresponding to query attributes . If the retriever can supply the correct value of the jth attribute, he will learn the jth share (1 j m). Denote the share corresponding to the jth attribute as s ij . Shares are also N bits long, i.e., s ij = s ij 1 || . . . ||s ijN . Each of the N bits of s ij has a corresponding bit in r i , which in its turn corresponds to one of the N records in the database. For each p s.t. 1 p N , set s ijp = r ip if x ij = x pj , and set s ijp = 0 otherwise. In other words, the jth share s ij consists of all bits of r i except those corresponding to the records where the value of the jth attribute is the same. An example can be found in section 4.4. The result of this construction is that shares corresponding to commonly occurring attribute values will be missing many bits of r i , while a share corresponding to an attribute that uniquely identifies one record will contain all bits of r i except one. Intuitively, this guarantees group privacy. If the retriever can supply query attribute values that uniquely identify a single record or a small subset of records, he will learn the shares that reveal all bits of the secret r i except for a few, which he can easily guess. If the retriever cannot describe precisely what he is looking for and supplies attribute values that are common in the database, many of the bits of r i will be missing in the shares that he learns, and guessing all of them will be computationally infeasible. Shares corresponding to different query attributes may overlap. For example, suppose that we are obfuscating a database in which two records have the same value of attribute X 1 if and only if they have the same value of attribute X 2 . In this case, for any record in the database, the share revealed if the retriever supplies the correct value of X 1 will be exactly the same as the share revealed if the retriever supplies the value of X 2 . The retriver gains nothing by supplying X 2 in conjunction with X 1 because this does not help him narrow the set of records that match his query. To construct the obfuscated database, we encrypt each share with a pseudo-random key derived from the value of the corresponding query attribute, and encrypt the data attribute with a key derived from the secret r i . More precisely, we replace each record i = x i 1 , . . . , x im , y i of the original database with the obfuscated record v i 1 , w i 1 , v i 2 , w i 2 , . . . , v im , w im , u i , z i where - v ij = H 1,i,j (x ij ). This enables the retriever to verify that he supplied the correct value for the jth query attribute. - w ij = prg 1,i,j (H 2,i,j (x ij )) s ij . This is the jth share of the secret r i , encrypted with a key derived from the value of the jth query attribute. - u i = H 3,i (r i ). This enables the retriever to verify that he computed the correct secret r i . - z i = prg 2,i (H 4,i (r i )) y i . This is the data attribute y i , encrypted with a key derived from the secret r i . Clearly, algorithm OB gp runs in time polynomial in N (the size of the database). The size of the resulting obfuscation is N 2 m . Even though it is within a polynomial factor of N (and thus satisfies the definition of [2]), quadratic blowup means that our technique is impractical for large databases. This issue is discussed further in section 4.5. We claim that OB gp produces a secure obfuscation of D, i.e., it is not feasible to extract any more information from OB gp (D) than permitted by the privacy policy GP D . Theorem 2. (Group-exponential obfuscation is "virtual black-box") For any probabilistic polynomial-time (adversarial) algorithm A, there exists a probabilistic polynomial-time simulator algorithm S and a negligible function of the security parameter k s.t. for any database D: |P(A(OB gp (D)) = 1) - P(S GP D (1 |D| ) = 1)| (k) Remark. An improper implementation of the random oracles in the above construction could violate privacy under composition of obfuscation, i.e., when more than one database is obfuscated and published. For instance, if the hash of some attribute is the same in two databases, then the adversary learns that the attributes are equal without having to guess their value. To prevent this, the same hash function may not be used more than once. One way to achieve this is to pick H i (.) = H(r i ||.) where r i R {0, 1} k , and publish r i along with the obfuscation. This is an example of the pitfalls inherent in the random oracle model. 4.3 Accessing the obfuscated database We now explain how the retriever can efficiently evaluate queries on the obfuscated database. Recall that the privacy policy restricts the retriever to queries consisting of conjunctions of equality tests on query attributes, i.e., every query predicate P has the form X j 1 = x j 1 . . . X j t = x j t , where j 1 , . . . , j t are some indices between 1 and m. The retriever processes the obfuscated database record by record. The ith record of the obfuscated database (1 i N ) has the form v i 1 , w i 1 , v i 2 , w i 2 , . . . , v im , w im , u i , z i . The retriever's goal is to compute the N -bit secret r i so that he can decrypt the ciphertext z i and recover the value of y i . First, the retriever recovers as many shares as he can from the ith record. Recall from the construction of section 4.2 that each w ij is a ciphertext of some share, but the only way to decrypt it is to supply the corresponding query attribute value x ij . Let range over the indices of attributes supplied by the retriever as part of the query, i.e., {j 1 , . . . , j t }. For each , if H 1,i, (x ) = v i , then the retriever extracts the corresponding share s i = prg 1,i, (H 2,i, (x )) w i . If H 1,i, (x ) = v i , this means that the retriever supplied the wrong attribute value, and he learns nothing about the corresponding share. Let S be the set of recovered shares. Each recovered share s S reveals only some bits of r i , and, as mentioned before, bits revealed by different shares may overlap. For each p s.t. 1 p N , the retriever sets the corresponding bit r ip of the candidate secret r i as follows: r ip = s p if s S s.t. v p = H 1,1, (x ) random otherwise Informally, if at least one of recovered shares s contains the pth bit of r i (this can be verified by checking that the value of th attribute is not the same in the pth record of the database -- see construction in section 4.2), then this 106 bit is indeed to the pth bit of the secret r i . Otherwise, the retriever must guess the pth bit randomly. Once a candidate r i is constructed, the retriever checks whether H 3,i (r i ) = u i . If not, the missing bits must have been guessed incorrectly, and the retriever has to try another choice for these bits. If H 3,i (r i ) = u i , then the retriever decrypts the data attribute y i = prg 2,i (H 4,i (r i )) z i . The obfuscation algorithm of section 4.2 guarantees that the number of missing bits is exactly equal to the number of records satisfied by the query P. This provides the desired group privacy property. If the retriever supplies a query which is satisfied by a single record, then he will only have to guess one bit to decrypt the data attributes. If a query is satisfied by two records, then two bits must be guessed, and so on. For queries satisfied by a large number of records, the number of bits to guess will be infeasible large. 4.4 Example Consider a toy airline passenger database with 4 records, where the query attributes are "Last name" and "Flight," and the data attribute (in bold) is "Purchase details." Last name Flight Purchase details Smith 88 Acme Travel, Visa 4390XXXX Brown 500 Airline counter, cash Jones 88 Nonrevenue Smith 1492 Travel.com, AmEx 3735XXXX Because N = 4, we need to create a 4-bit secret to protect each data attribute. (4-bit secrets can be easily guessed, of course. We assume that in real examples N would be sufficiently large, and use 4 records in this example only to simplify the explanations.) Let = 1 2 3 4 be the secret for the first data attribute, and , , the secrets for the other data attributes, respectively. For simplicity, we use a special symbol "?" to indicate the missing bits that the retriever must guess. In the actual construction, each of these bits is equal to 0, but the retriever knows that he must guess the ith bit of the jth share if the value of the jth attribute in the current record is equal to the value of the jth attribute in the ith record. Consider the first record. Each of the two query attributes, "Last name" and "Flight," encrypts a 4-bit share. The share encrypted with the value of the "Last name" attribute (i.e., "Smith") is missing the 1st and 4th bits because the 1st and 4th records in the database have the same value of this attribute . (Obviously, all shares associated the ith record have the ith bit missing). The share encrypted with the value of the "Flight" attribute is missing the 1st and 3rd bits. H 111 ("Smith"), prg 1,1,1 (H 211 ("Smith")) (? 2 3 ?), H 112 ("88"), prg 1,1,2 (H 212 ("88")) (? 2 ? 4 ), H 31 ( 1 2 3 4 ), prg 2,1 (H 41 ( 1 2 3 4 )) ("Acme. . . ") H 121 ("Brown"), prg 1,2,1 (H 221 ("Brown")) ( 1 ? 3 4 ), H 122 ("500"), prg 1,2,2 (H 222 ("500")) ( 1 ? 3 4 ), H 32 ( 1 2 3 4 ), prg 2,2 (H 42 ( 1 2 3 4 )) ("Airline. . . ") H 131 ("Jones"), prg 1,3,1 (H 231 ("Jones")) ( 1 2 ? 4 ), H 132 ("88"), prg 1,3,2 (H 232 ("88")) (? 2 ? 4 ), H 33 ( 1 2 3 4 ), prg 2,3 (H 43 ( 1 2 3 4 )) ("Nonrev. . . ") H 141 ("Smith"), prg 1,4,1 (H 241 ("Smith")) (? 2 3 ?), H 142 ("1492"), prg 1,4,2 (H 242 ("1492")) ( 1 2 ? 4 ), H 34 ( 1 2 3 4 ), prg 2,4 (H 44 ( 1 2 3 4 )) ("Travel.com. . . ") Suppose the retriever knows only that the flight number is 88. There are 2 records in the database that match this predicate . From the first obfuscated record, he recovers ? 2 ? 4 and from the third obfuscated record, ? 2 ? 4 . The retriever learns which bits he must guess by computing H 2i2 ("88") for 1 i 4, and checking whether the result is equal to v i 2 from the ith obfuscated record. In both cases, the retriever learns that he must guess 2 bits (1st and 3rd) in order to reconstruct the secret and decrypt the data attribute. Now suppose the retriever knows that the flight number is 88 and the last name is Smith. There is only 1 record in the database that satisfies this predicate. From the first part of the first obfuscated record, the retriever can recover ? 2 3 ?, and from the second part ? 2 ? 4 (note how the shares overlap). Combining them, he learns ? 2 3 4 , so he needs to guess only 1 bit to decrypt the data attribute. It may appear that this toy example is potentially vulnerable to a dictionary attack, since it is conceivable that all combinations of last names and flight numbers can be efficiently enumerated with enough computing power. Note, however, that this "attack" does not violate the definition of secure obfuscation because the retriever must supply the name-flight pair before he can recover the purchase details. Therefore , the obfuscated database is only accessed via queries permitted by the privacy policy. In databases where values are drawn from a large set, even this "attack" is infeasible. 4.5 Efficiency The algorithm of section 4.2 produces obfuscations which are a factor of (N ) larger than original databases. Thus, while our results establish feasibility of database obfuscation and group privacy, they are not directly applicable to real-world databases. This appears to be a recurring problem in the field of database privacy: the cryptography community has very strict definitions of security but loose notions of efficiency (typically polynomial time and space), whereas the database community has very strict efficiency requirements but loose security (typically heuristic or statistical). As a result, many proposed schemes are either too inefficient, or too insecure for practical use. A possible compromise might be to start with a provably secure but inefficient construction and employ heuristic techniques to improve its efficiency. In this spirit, we now propose some modifications to reduce the size of the obfuscated database without providing a security proof. The presentation is informal due to lack of space; see the full version of the paper for a more rigorous version. The obfuscation algorithm is modified as follows. For each record i, we split r i into N k "blocks" of k bits each, padding the last block if necessary (k is the security parameter). Instead of generating the bits randomly, we create a binary tree of depth log N k . A key of length k is associated with each node of the tree, with the property the two "children" keys are derived from the "parent" key (e.g., using a size-doubling pseudo-random generator). This is similar to a Merkle tree in which keys are derived in the reverse direction. The edge of tree (minus the padding of the last block) is r i . Let us denote the j th attribute of the i th record by i, j . Say that i, j is entitled to the secret bit r i j if x ij = x i j , and i, j is entitled to an entire block B if it is entitled to each secret bit r i j in that block. Intuitively, if an entire block is entitled, then we encode it efficiently using the "reverse Merkle" tree described above; if it is partially entitled , then we fall back on our original construction. Thus, 107 let N ij be the minimal set of nodes in the tree which are sufficient for reconstructing all entitled blocks (i.e., every entitled block has among its parents an element of N ij ), and only these blocks. Then the share s ij consists of (a suitable encoding of) N ij together with the remaining bits r i j to which i, j is entitled. These are the entitled bits from any block which also includes non-entitled bits. In the worst case, this algorithm does not decrease the blowup in the size of the obfuscated database. This occurs when for every query attribute j of every record i, there are (N ) records i for which the value of the query attribute is the same, i.e., x ij = x i j . If we assume a less pathological database, however, we can get a better upper bound. If there is a threshold t such that for any (i, j) there are at most t records i for which x ij = x i j , then the size of the key material (which causes the blowup in the size of the obfuscated database) is O(mN t(k log N k )) bits (recall that m is the number of attributes). This bound is tight only for small values of t, and the new algorithm does no worse than the original one even when t = (N ). When we consider that each of the mN entries of the original database is several bits long, the size of the obfuscated database could be acceptably small for practical use. It must be noted that this obfuscation reveals the size of the share, and thus, for a given attribute of a given record, it leaks information about the number of other records whose attribute value is the same (but not which records they are). This opens two research questions: - Is there a provably secure database obfuscation algorithm that produces obfuscations that are smaller than O(N 2 ). - Can the heuristic described in this section be improved to obtain acceptable lower bounds in the worst case? ARBITRARY PREDICATES OVER EQUALITIES ON ATTRIBUTES We now consider queries formed by taking an arbitrary predicate P over m boolean variables b 1 , b 2 . . . b m , and substituting (X j = x j ) for b j , where X j is a query attribute, and x j X {} is a candidate value for this attribute, drawn from the domain X of query attribute values. The special value denotes that the value of the X j attribute is ignored when evaluating the predicate. The class of queries considered in section 4 is a partial case of this definition, where P = 1jm b j . The group-exponential property is similar to definition 2 except for the domain of P. Let C be a boolean circuit computing P. We assume that C is a monotone circuit, i.e., a poly-size directed acyclic graph where each node is an AND, OR or FANOUT gate. AND and OR gates have two inputs and one output each, while FANOUT gates have one input and two outputs. Circuit C has m inputs, one per each query attribute. Below, we show how to generalize our obfuscation technique to non-monotone circuits. Obfuscation algorithm. The algorithm is similar to the one in section 4, and consists of generating a random secret to encrypt each data attribute, splitting this secret into (unequal) shares, and encrypting these shares under the keys derived from the values of query attributes. As before, let H : {0, 1} {0, 1} k be a set of random hash functions and prg , : {0, 1} k {0, 1} a set of pseudo-random generators. For each record i in the database, do the following: Generate a block of uniformly random bits {r ilEt }, where 1 l N , E ranges over all edges of the circuit C, and 1 t k, where k is the length of the hash functions' output. Denote r iEt = r i 1Et ||r i 2Et || . . . ||r iN Et -r ilE = r ilE 1 ||r ilE 2 || . . . ||r ilEk Then, for each query attribute X j : Output v ij = H 1,i,j (x ij ) Let E j be the input edge in the circuit C whose input is the X j = x j test. Define the bits of the corresponding share s iljt = r ilE j t if x ij = x lj , and 0 otherwise. Encrypt the resulting share using a key derived from x ij , i.e., output w ij = prg 1,i,j (H 2,i,j (x ij )) ( s i 1j ||-s i 2j || . . . ||-s iN j ). Let E out be the output edge in the circuit C. Output u i = H 3,i (r iE out 0 ) Output z i = prg 2,i (H 4,i (r iE out 0 )) y i . The previous procedure obfuscated only the output edge of C. Repeat the following step recursively for each gate G C, whose output edge (or both of whose output edges, for a FANOUT gate) have been obfuscated . Stop when all edges have been obfuscated: If G is an AND gate, let E 0 and E 1 be the input edges and E the output edge. For each l, set -r ilE 0 = -r ilE 1 = -r ilE . If G is an OR gate, then, for each l, generate random -r ilE 0 R {0, 1} k and set -r ilE 1 = -r ilE 0 -r ilE . If G is a FANOUT gate, let E 0 and E 1 be the output edges and E the input edge. For each l, generate random -r ilE 0 , -r ilE 1 R {0, 1} k and output ilE 0 = H 5,i,l,E 0 (-r ilE ) -r ilE 0 and ilE 1 = H 5,i,l,E 1 (-r ilE ) -r ilE 1 Retrieval algorithm. Let Q be the query predicate in which specific values of x j or have been plugged into all X j = x j expressions in the leaves of the circuit C. The retrieval algorithm consists of two functions: C ob (OB gp (D), x, i, j), which enables the retriever to check whether the jth query attribute of the ith record is equal to x , and R ob (OB gp (D), Q, i), which attempts to retrieve the value of the obfuscated data attribute in the ith record. Define C ob (OB gp (D), x, i, j) = 1 if H 1,i,j (x) = v ij and 0 otherwise. Evaluate Q( i ) using C ob . If Q OB gp ( i ), then R ob (OB gp (D), Q, i) =. For each l and each circuit edge E, set -r ilE =?? . . .? (i.e., none of the bits of the secret are initially known). For each query attribute j, let E j be the input edge of the circuit associated with the equality test for this attribute . If Q contains this test, i.e., if Q contains X j = 108 x j for some candidate value x j (rather than X j = ), then set ( s i 1j || . . . ||-s iN j ) = w ij prg 1,i,j (H 2,i,j (x ij )), i.e., decrypt the secret bits with the key derived from the value of the jth attribute. For each l, if C ob (x ij , l, j ) = 0, then set -r ilE j = s ilj , i.e., use only those of the decrypted bits that are true bits of the secret -r ilE . So far, only the input gates of the circuit have been visited. Find a gate all of whose input edges have been visited, and repeat the following step for every gate until the output edge E out has been visited. If E is the output of an AND gate with inputs E 0 and E 1 , then, for each l, if -r ilE 0 =?, set -r ilE = -r ilE 0 ; if -r ilE 1 =?, set -r ilE = -r ilE 1 . E is the output of an OR gate with inputs E 0 and E 1 . For each l, if -r ilE 0 =? and -r ilE 1 =?, set -r ilE = -r ilE 0 -r ilE 1 . E is the output of a FANOUT gate with input E 0 . For each l, if -r ilE 0 =?, set r ilE = ilE 0 H 5,i,l,E 0 (-r ilE 0 ). For each l, if r ilE out 0 =?, this means that the corresponding secret bit must be guessed. Choose random r ilE out 0 R {0, 1}. If H 3,i (r iE out 0 ) = u i , this means that the retriever successfully reconstructed the secret. In this case, define R ob (OB gp (D), Q, i) = prg 2,i (H 4,i (r iE out 0 )) z i . Otherwise , define R ob (OB gp (D), Q, i) =. Theorem 3. The obfuscation algorithm for arbitrary predicates over equalities on attributes satisfies the virtual black-box property. 5.1 Obfuscating non-monotone circuits Given a non-monotone circuit C, let C be the monotone circuit whose leaves are literals and negated literals formed by "pushing down" all the NOT gates. Observe that C has at most twice as many gates as C. Also, C can be considered a monotone circuit over the 2m predicates X 1 = x 1 , X 2 = x 2 , . . . , X m = x m , X 1 = x 1 , X 2 = x 2 , . . . X m = x m . Observe that a predicate of the form X j = x j is meaningful only when x j = x ij for some record i. This is because if x j = x ij for any record i, then X j = x j matches all the records. Hence there exists a circuit C (obtained by setting the leaf in C corresponding to the predicate X j = x j to true) that evaluates to the same value as C for every record in the database. Given that x j = x ij for some record i, the predicate X j = x j is equivalent to the predicate X j = x ij for some value of i. C can thus be viewed as a monotone circuit over the m + mN attribute equality predicates X 1 = x 1 , X 2 = x 2 , . . . , X m = x m , and X j = x ij for each i and j. It follows that a database D with N records and m columns can be transformed into a database D with N records and m+mN columns such that obfuscating D over the circuit C is equivalent to obfuscating D over the monotone circuit C. ALTERNATIVE PRIVACY POLICIES In general, a privacy policy can be any computable, possibly randomized, joint function of the database and the query. Clearly, it may be useful to consider generalizations of our privacy policies in several directions. First, we discuss alternatives to definition 2 that may be used to model the requirement that accessing individual records should be easy, but mass harvesting of records should be hard. To motivate this discussion, let us consider a small database with, say, 10 or 20 records. For such a database, the group-exponential property is meaningless. Even if all records match the adversary's query, he can easily try all 2 10 or 2 20 possibilities for the random bits r ik because database accesses are noninteractive. This does not in any way violate our definition of privacy. Exactly the same attack is possible against the ideal functionality , therefore, the simulation argument goes through, showing that the obfuscated database leaks no more information than the ideal functionality. It is thus natural to seek an alternative privacy definition that will make the above attack infeasible when N is small (especially when N &lt; k, the security parameter). Our construction can be easily modified to support a wide variety of (monotonically decreasing) functions capturing the dependence between the probability of the ideal functionality returning the protected attributes and the number of records matching the query. For example, the following threshold ideal functionality can be implemented using a threshold (n-t)-out-of-n secret sharing scheme [24]. - C D (x, i, j) is 1 if x = x ij and 0 otherwise, where 1 i N, 1 j m. - R D (P) = 1iN { i, i }, where i = y i if P( i ) and |D [P] | t if P( i ) and |D [P] | &gt; t if P( i ) The adversary can evaluate the query if there are at most t matching records, but learns nothing otherwise. The details of the construction are deferred to the full version. We may also consider which query language should be permitted by the privacy policy. We demonstrated how to obfuscate databases in accordance with any privacy policy that permits evaluation of some predicate consisting of equality tests over database attributes. Such queries can be considered a generalization of "partial match" searches [23], which is a common query model in the database literature. Also, our algorithms can be easily modified to support policies that forbid some attributes from having as a legal value, i.e., policies that require the retriever to supply the correct value for one or more designated attributes before he can extract anything from the obfuscated database. It is worth asking if we can allow predicates over primitives looser than exact attribute equality (e.g., proximity queries of [15] are an interesting class). We present strong evidence that this is impossible with our privacy definitions. In fact, even using ideal functionalities (IF) that are more restrictive than the one we have used does not seem to help. Recall that the IF considered in section 4 consists of two functions: C D (it tells the retriever whether his guess of a particular query attribute value is correct) and R D (it evaluates the query with the inverse-exponential probability). We will call this IF the permissive IF. We define two more IFs. The strict IF is like the permissive IF except that it doesn't have the function C. The semi-permissive IF falls in between the two. It, too, 109 doesn't have the function C, but its retrieval function R leaks slightly more information. Instead of the same symbol , function R of the semi-permissive IF gives different responses depending on whether it failed to evaluate the query because it matches no records (no-matches) or because it matches too many records, and the probability came out to the retriever's disadvantage (too-many-matches). Define R D (P) as 1iN R (P, i), where R is as follows: If P( i ), then R (P, i) = . If P( i ), then R (P, i) = y i with probability 2 -|D [P] | and with probability 1 - 2 -|D [P] | . Observe that if, for any privacy policy allowing single-attribute equality tests, i.e., if all queries of the form X j = x j are permitted, then the semi-permissive IF can simulate the permissive IF. Of course, the permissive IF can always simulate the semi-permissive IF. We say that a privacy policy leaks query attributes if all x ij can be computed (with overwhelming probability) simply by accessing the corresponding ideal functionality I D , i.e., there exists a probabilistic poly-time oracle algorithm A s.t., for any database D, P(A I D , O (i, j) = x ij ) 1 - (k). Note that the order of quantifiers has been changed: the algorithm A is now independent of the database. This captures the idea that A has no knowledge of the specific query attributes, yet successfully retrieves them with access only to the ideal functionality. Such a policy, even if securely realized, provides no meaningful privacy. We have the following results (proofs omitted): If X = {1, 2, . . . M } and queries consisting of conjunctions over inequalities are allowed, then the semi-permissive IF leaks query attributes. Each of the x ij can be separately computed by binary search using queries of the form X j x low X j x high . If arbitrary PPT-computable queries are allowed, then even the strict IF leaks query attributes. Note that a policy does not have to leak all query attributes to be intuitively useless or vacuous. For instance, a policy which allows the retriever to evaluate conjunctions of inequalities on the first m - 1 query attributes, and allows no queries involving the last attribute, is vacuous for the semi-permissive IF. Therefore, we give a stronger criterion for vacuousness, which formalizes the notion that "all information contained in the IF can be extracted without knowing anything about the query attributes". Note that the definition below applies to arbitrary privacy policies, for it makes no reference to query or data attributes. Definition 3. (Vacuous privacy policy) We say that an ideal functionality I D is vacuous if there exists an efficient extractor Ext such that for any PPT algorithm A there exists a simulator S so that for any database D: |P(A I D (1 k ) = 1) - P(S(Ext I D (1 k ))) = 1)| = (k) In other words, we first extract all useful information from I D without any specific knowledge of the database, throw away I D , and use the extracted information to simulate I D against an arbitrary adversary. As a special case, if Ext can recover the entire database D from I D , then the functionality can be simulated, because the privacy policy is required to be computable and the simulator is not required to be computationally bounded (if we consider only privacy policies which are computable in probabilistic polynomial time, then we can define vacuousness with a PPT simulator as well). At the other extreme, the ideal functionality that permits no queries is also simulatable: Ext simply outputs nothing. The reader may verify that the IF in the all-but-one -query-attribute example above is also vacuous. Theorem 4. The strict ideal functionality that permits arbitrary queries is vacuous. Finally, we consider what happens if we use the strict IF but don't increase the power of the query language. We conjecture the existence of very simple languages, including a language that contains only conjunctions of equality tests on attributes, which are unrealizable even for single-record databases in the sense that there is no efficient obfuscation algorithm that would make the database indistinguishable from the corresponding IF. This can be seen as justification for the choice of the permissive, rather than strict IF for our constructions. conjecture 1. The strict IF for the following query language cannot be realized even for single-record databases: 2k i =1 (X 2i-1 = x 2i-1 X 2i = x 2i ) where i x i {0, 1}. Note that the only constraint on the database is that its size should be polynomial in the security parameter k, and therefore we are allowed to have 2k query attributes. We expect that a proof of this conjecture will also yield a proof of the following conjecture: conjecture 2. The strict IF for a query language consisting of conjunction of equality tests on k query attributes is unrealizable even for single-record databases. These conjectures are interesting from another perspective . They can be interpreted as statements about the impossibility of circuit obfuscation in the random oracle model. They also motivate the question: given a query language, it is possible to achieve the group-exponential property with the strict IF provided there exists an obfuscation algorithm for this query language on a single record? In other words, given a class of predicates over single records and an efficient obfuscator for the corresponding circuit class, does there exist an obfuscator for the entire database that realizes the group-exponential ideal functionality for that query language? We discuss this question in the full version of the paper. CONCLUSIONS We introduced a new concept of database privacy, which is based on permitted queries rather than secrecy of individual records, and realized it using provably secure obfuscation techniques. This is but a first step in investigating the connection between obfuscation and database privacy. While our constructions are secure in the "virtual black-box" model for obfuscation, the blowup in the size of the obfuscated database may render our techniques impractical for large databases. Our query language permits any predicate over equalities on database attributes, but other query languages may also be realizable. We define group privacy in terms of a particular ideal functionality, but there may be 110 other functionalities that better capture intuitive security against "mass harvesting" queries. In general, investigating which ideal functionalities for database privacy can be securely realized is an important topic of future research. Finally, all proofs in this paper are carried out in the random oracle model. Whether privacy-via-obfuscation can be achieved in the plain model is another research challenge. REFERENCES [1] D. Aucsmith. Tamper resistant software: an implementation. In Proc. 1st International Workshop on Information Hiding, volume 1174 of LNCS, pages 317333. Springer, 1996. [2] B. Barak, O. Goldreich, R. Impagliazzo, S. Rudich, A. Sahai, S. Vadhan, and K. Yang. On the (im)possibility of obfuscating programs. In Proc. Advances in Cryptology - CRYPTO 2001, volume 2139 of LNCS, pages 118. Springer, 2001. [3] D. Beaver, J. Feigenbaum, J. Kilian, and P. Rogaway. Locally random reductions: improvements and applications. J. Cryptology, 10:1736, 1997. [4] D. Boneh, G. Di Crescenzo, R. Ostrovsky, and G. Persiano. Public key encryption with keyword search. In Proc. Advances in Cryptology EUROCRYPT 2004, volume 3027 of LNCS, pages 506522. Springer, 2004. [5] R. Canetti. Towards realizing random oracles: hash functions that hide all partial information. In Proc. Advances in Cryptology - CRYPTO 1997, volume 1294 of LNCS, pages 455469. Springer, 1997. [6] R. Canetti, D. Micciancio, and O. Reingold. Perfectly one-way probabilistic hash functions. In Proc. 30th Annual ACM Symposium on Theory of Computing (STOC), pages 131140. ACM, 1998. [7] S. Chawla, C. Dwork, F. McSherry, A. Smith, and H. Wee. Towards privacy in public databases. In Proc. 2nd Theory of Cryptography Conference (TCC), volume 3378 of LNCS, pages 363385. Springer, 2005. [8] B. Chor, E. Kushilevitz, O. Goldreich, and M. Sudan. Private information retrieval. J. ACM, 45(6):965981, 1998. [9] S. Chow, P. Eisen, H. Johnson, and P. van Oorschot. White-box cryptography and an AES implementation. In 9th Annual International Workshop on Selected Areas in Cryptography (SAC), volume 2595 of LNCS, pages 250270. Springer, 2003. [10] S. Chow, P. Eisen, H. Johnson, and P. van Oorschot. A white-box DES implementation for DRM applications. In ACM Digital Rights Management Workshop, volume 2696 of LNCS, pages 115. Springer, 2003. [11] C. Collberg and C. Thomborson. Watermarking, tamper-proofing, and obfuscation - tools for software protection. IEEE Transactions on Software Engineering, 28(8):735746, 2002. [12] C. Collberg, C. Thomborson, and D. Low. A taxonomy of obfuscating transformations. Technical Report 148, Department of Computer Sciences, The University of Auckland, July 1997. [13] C. Collberg, C. Thomborson, and D. Low. Manufacturing cheap, resilient, and stealthy opaque constructs. In Proc. 25th ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL), pages 184196. ACM, 1998. [14] D. Dean and A. Stubblefield. Using client puzzles to protect TLS. In Proc. 10th USENIX Security Symposium, pages 18. USENIX, 2001. [15] Y. Dodis and A. Smith. Correcting errors without leaking partial information. In Proc. 37th Annual ACM Symposium on Theory of Computing (STOC), pages 654663. ACM, 2005. [16] C. Dwork and M. Naor. Pricing via processing or combatting junk mail. In Proc. Advances in Cryptology - CRYPTO 1992, volume 740 of LNCS, pages 139147. Springer, 1993. [17] M. Franklin and D. Malkhi. Auditable metering with lightweight security. J. Computer Security, 6(4):237255, 1998. [18] Y. Gertner, Y. Ishai, E. Kushilevitz, and T. Malkin. Protecting data privacy in private information retrieval schemes. In Proc. 30th Annual ACM Symposium on Theory of Computing (STOC), pages 151160. ACM, 1998. [19] O. Goldreich and R. Ostrovsky. Software protection and simulation on oblivious rams. J. ACM, 43(3):431473, 1996. [20] Y. Ishai, A. Sahai, and D. Wagner. Private circuits: securing hardware against probing attacks. In Proc. Advances in Cryptology - CRYPTO 2003, volume 2729 of LNCS, pages 463481. Springer, 2003. [21] A. Juels and J. Brainard. Client puzzles: a cryptographic defense against connection depletion. In Proc. Network and Distributed System Security Symposium (NDSS), pages 151165. The Internet Society, 1999. [22] B. Lynn, M. Prabhakaran, and A. Sahai. Positive results and techniques for obfuscation. In Proc. Advances in Cryptology - EUROCRYPT 2004, volume 3027 of LNCS, pages 2039. Springer, 2004. [23] R. Rivest. Partial-match retrieval algorithms. SIAM Journal of Computing, 5(1):1950, 1976. [24] A. Shamir. How to share a secret. Communications of the ACM, 22(11):612613, 1979. [25] D. Song, D. Wagner, and A. Perrig. Practical techniques for searches on encrypted data. In Proc. IEEE Symposium on Security and Privacy, pages 4455. IEEE Computer Society, 2000. [26] X. Wang and M. Reiter. Defending against denial-of-service attacks with puzzle auctions. In Proc. IEEE Symposium on Security and Privacy, pages 7892. IEEE Computer Society, 2003. [27] H. Wee. On obfuscating point functions. In Proc. 37th Annual ACM Symposium on Theory of Computing (STOC), pages 523532. ACM, 2005. 111

Database privacy;Obfuscation

143

On the Complexity of Computing Peer Agreements for Consistent Query Answering in Peer-to-Peer Data Integration Systems

Peer-to-Peer (P2P ) data integration systems have recently attracted significant attention for their ability to manage and share data dispersed over different peer sources. While integrating data for answering user queries, it often happens that inconsistencies arise, because some integrity constraints specified on peers' global schemas may be violated. In these cases, we may give semantics to the inconsistent system by suitably "repairing" the retrieved data, as typically done in the context of traditional data integration systems. However , some specific features of P2P systems, such as peer autonomy and peer preferences (e.g., different source trusting ), should be properly addressed to make the whole approach effective. In this paper, we face these issues that were only marginally considered in the literature. We first present a formal framework for reasoning about autonomous peers that exploit individual preference criteria in repairing the data. The idea is that queries should be answered over the best possible database repairs with respect to the preferences of all peers, i.e., the states on which they are able to find an agreement. Then, we investigate the computational complexity of dealing with peer agreements and of answering queries in P2P data integration systems. It turns out that considering peer preferences makes these problems only mildly harder than in traditional data integration systems.

INTRODUCTION Peer-to-Peer (P2P ) data integration systems are networks of autonomous peers that have recently emerged as an effective architecture for decentralized data sharing, integration, and querying. Indeed, P2P systems offer transparent access to the data stored at (the sources of) each peer p, by means of the global schema equipped with p for modeling its domain of interest; moreover, pair of peers with the same domain of interest one peer and the system is in charge of accessing each peer containing relevant data separately, and combining local results into a global answer by suitably exploiting the mapping rules. P2P systems can be considered the natural evolution of traditional data integration systems, which have received considerable attention in the last few years, and which have already become a key technology for managing enormous amounts of information dispersed over many data sources. In fact, P2P systems have attracted significant attention recently, both in the development of efficient distributed algorithms for the retrieval of relevant information and for answering user queries (see, e.g., [9, 21, 12, 13]), and in the investigation of its theoretical underpinnings (see, e.g., [16, 3, 20, 11, 9, 5]). In this paper, we continue along this latter line of research, by investigating some important theoretical issues. In particular , we consider an expressive framework where integrity constraints are specified on peer schemas in order to enhance their expressiveness, so that each peer can be in fact considered a completely specified data integration system. In this scenario, it may happen that data at different peers are mutually inconsistent, i.e., some integrity constraints are violated after the integration is carried out; then, a "repair" for the P2P system has to be computed [5, 17]. Roughly speaking, repairs may be viewed as insertions or deletions of tuples at the peers that are able to lead the system to a consistent state. Our aim is to deal with data integration in P2P systems, by extending some of the ideas described in previous studies on merging mutually inconsistent databases into a single consistent theory [2, 14] and on repairing individual data integration systems [8, 6, 4, 10]. 36 Indeed, in order to be effective in this framework, the repair approach should consider the peculiarities of P2P systems and, specifically, the following two issues: In practical applications, peers often have an a-priori knowledge about the reliability of the sources that, in turn, determines their criteria for computing repairs. That is, peers will rarely delete tuples coming from highly reliable sources, and will try to solve conflicts by updating the less reliable sources only. Peers are autonomous and not benevolent: they rarely disregard their individual preferences in order to find an agreement with other peers on the way the repair should be carried out. Therefore, the presence of possibly contrasting interests of selfish peers should be accounted for, when answering user queries. Despite the wide interest in this field, none of the approaches in the literature considered the issue of modeling the autonomy of the peers in providing a semantics for the system, and therefore they implicitly assume that all the peers act cooperatively in the network. Moreover, the possibility of modeling peer preferences has been rarely considered in previous studies, even though it has been widely recognized to be a central issue for the design of quality-aware integration systems (cf. [17]). Indeed, the first and almost isolated attempt is in [5], where the authors considered trust relationships among peers in a simplified setting in which the system does not transitively propagate information through peers. Actually, an extension to the case of transitive propagations is also argued, but peers autonomy is not considered, and query answering is undecidable in presence of loops. In this paper, we face the above issues by introducing a formal framework for reasoning about autonomous peers that exploit individual preference criteria in repairing data. In summary, our contributions are the following: We preliminary introduce a framework for P2P data integration systems, where each peer is equipped with integrity constraints on its global schema. The model is simple yet very expressive, since each peer is assumed to be in turn a data integration system. The semantics of a P2P system is defined in terms of suitable databases for the peers, called models. We show that checking whether a system has a model can be done efficiently. We propose an approach to the repair of inconsistent P2P systems that focuses on data stored at the sources, rather than on the global schema (following the approach described by [15] for the standard data integration setting). This is particularly suited for dealing with peers, as their preferences are typically expressed over the sources. Indeed, if repairs were considered on the global schema, suitable reformulations and translation of the preferences would be required. We investigate the effect of considering individual preferences on the semantics of P2P database integration systems. The idea is that queries should be answered over the best possible database repairs with respect to the preferences of all peers, i.e., over the states on which they are able to find an agreement. Unfortu-nately , but not surprisingly, it turns out that considering autonomous peers gives rise to scenarios where they are not able to find any agreement on the way the integration should be done. The above result motivates the subsequent study of the complexity of dealing with peer agreements and of answering queries in such P2P data integration systems . We show that checking whether a given database is an agreed repair is a difficult task, since it is complete for the class co-NP. Moreover, the complexity of computing an agreement turns out to be complete for the functional class FPNP. Finally, we study the complexity of computing consistent answers and show that this problem is P 2 -complete. It follows that our approach for handling preferences in P2P systems is just mildly harder than the basic data integration framework , where in fact query answering lies at the first level of the polynomial hierarchy [8], as well. The rest of the paper is organized as follows. In Section 2, we briefly present some preliminaries on relational databases. In Section 3, we introduce a simple formalization of P2P data integration systems and in the subsequent section we enrich it to take care of peers' preferences. The computational complexity of the concept of agreement in query answering is studied in Section 5. Finally, in Section 6 we draw our conclusions. PRELIMINARIES ON RELATIONAL DATABASES We recall the basic notions of the relational model with integrity constraints. For further background on relational database theory, we refer the reader to [1]. We assume a (possibly infinite) fixed database domain whose elements can be referenced by constants c 1 ,. . . , c n under the unique name assumption, i.e. different constants denote different objects. These elements are assumed to be shared by all the peers and are, in fact, the constants that can appear in the P2P system. A relational schema (or simply schema) RS is a pair , , where: is a set of relation symbols, each with an associated arity that indicates the number of its attributes, and is a set of integrity constraints, i.e., (first-order) assertions that have to be satisfied by each database instance. We deal with quantified constraints, i.e., first order formulas of the form: ~x. l i=1 A i ~y. m j=1 B j n k=1 k , (1) where l+m &gt; 0, n 0, A 1 , . . . A l and B 1 , . . . B m are positive literals, 1 , . . . n are built-in literals, and ~ x and ~ y are lists of distinct variables. Actually, to keep things simple, we shall assume throughout the paper that ~ y is empty, thereby dealing with universally quantified constraints. We recall here that this kind of constraint covers most of the classical constraints issued on a relational schema, such as keys, functional dependencies, and exclusion dependencies. A brief discussion on how to generalize the results in the paper to other classes of constraints is reported in Section 6. A database instance (or simply database) DB for a schema RS = , is a set of facts of the form r(t) where r is a relation of arity n in and t is an n-tuple of constants from . We denote as r DB the set {t | r(t) DB}. A database DB for a schema RS is said to be consistent with RS if it satisfies (in the first order logic sense) all constraints expressed on RS. 37 Figure 1: The P2P system P r in Example 1. A relational query (or simply query ) over RS is a formula that is intended to extract tuples of elements from the underlying domain of constants . We assume that queries over RS = , are Unions of Conjunctive Queries (UCQs), i.e., formulas of the form {~x | ~y 1 . conj 1 (~ x, ~ y 1 ) ~y m . conj m (~ x, ~ y m ) } where, for each i {1, . . . , m}, conj i (~ x, ~ y i ) is a conjunction of atoms whose predicate symbols are in , and involve ~ x = X 1 , . . . , X n and ~ y i = Y i,1 , . . . , Y i,n i , where n is the arity of the query, and each X k and each Y i, is either a variable or a constant in . Given a database DB for RS, the answer to a UCQ Q over DB, denoted Q DB , is the set of n-tuples of constants c 1 , . . . , c n such that, when substituting each X i with c i , the formula ~y 1 . conj 1 (~ x, ~ y 1 ) ~y m . conj m (~ x, ~ y m ) evaluates to true on DB. DATA INTEGRATION IN P2P SYSTEMS In this section, we introduce a simple framework for dealing with P2P systems. The model is not meant to be a novel comprehensive formalization, since our aim here is to face the problem of finding agreement among peers rather than to investigate new syntactic modeling features. Therefore, our approach takes basically the same perspective as [9, 11, 5, 17]. 3.1 Basic Framework A P2P system P is a tuple P, I, N , map , where P is a non-empty set of distinct peers and I, N and map are functions whose meaning will be explained below. First, each peer p P is equipped with its own data integration system I(p), which is formalized as a triple G p , S p , M p . Basically, S p is meant to denote the set of sources to which p is allowed to access and is in fact modeled as a relational schema of the form S p = p , , i.e., there are no integrity constraints on the sources. The structure of the global schema is, instead, represented by means of the schema G p = p , p , whereas the relationships between the sources and the global schema are specified by M p , which is a set of local mapping assertions between G p and S p . We assume that each assertion is of the form Q S p Q G p , where Q S p and Q G p are two conjunctive queries of the same arity over the source schema S p and the peer schema G p , respectively. Example 1 Let us introduce three peers, namely p 1 , p 2 , and p 3 , that constitute the P2P scenario that will be used as a running example throughout this paper to illustrate technical definitions. The global schema G p 1 of peer p 1 consists of the relation predicate secretary (Employee, Manager ) (without constraints ), the source schema S p 1 consists of the relation symbol s 1 , and the set M p 1 of the local mapping assertions is {X, Y | s 1 (X, Y )} {X, Y | secretary (X, Y )}. As for peer p 2 , the schema G p 2 consists of the relation financial (Employee, Manager ) (without constraints), the source schema consists of the relation symbol s 2 , and M p 2 = {X, Y | s 2 (X, Y )} {X, Y | financial(X, Y )}. The schema G p 3 of peer p 3 consists of the relations employee(Name, Dept) and boss(Employee, Manager ), whose set of constraints contains the assertions (quantifiers are omitted) employee (X, Y ) boss(X 1 , Y 1 ) X = Y 1 and boss(X, Y ) boss(X 1 , Y 1 ) Y 1 = X, stating that managers are never employees; the source schema S p 3 comprises the relation symbols s 3 ; and, the set of the local mapping assertions is {X, Y | s 3 (X, Y )} {X, Y | employee(X, Y )}. P Each peer p P in a P2P system P = P, I, N , map is also equipped with the neighborhood function N providing a set of peers N (p) P - {p} containing the peers (called neighbors) who potentially have some information of interest to p. Intuitively, the neighborhood relation determines the structure of a P2P system P. Such a structure is better described by the dependency graph G( P) of P, i.e., by a directed graph having P as its set of vertices and {(p, q) | q P p N (q)} as its set of edges. In particular, a peer q is in N (p) iff p is interested in the data exported by q by means of its global schema, i.e., some of the global relations of p can be populated by means of the data coming from q besides the data coming from the sources of p itself. To this aim, map(p) defines the set of peer mapping assertions of p. Each assertion is an expression of the form Q q Q p , where the peer q N (p) is a neighbor of p, and Q q and Q p are two conjunctive queries of the same arity over schemas G q and G p , respectively. Example 1 (contd.) Let P r = P r , I r , N r , map r be a P2P system, where P r consists of three peers p 1 , p 2 and p 3 , such that N r (p 1 ) = N r (p 2 ) = and N r (p 3 ) = {p 1 , p 2 }. Figure 1 summarizes the structure of the system P r by showing, for each peer, its global schema, its source schema, and its local and peer mapping assertions. In particular, notice that the mapping assertions are such that: map(p 1 ) = map(p 2 ) = , and map(p 3 ) = {X, Y | financial(X, Y ))} {X, Y | boss(X, Y )} {X, Y | secretary(X, Y )} {X, Y | boss(X, Y )}. P 38 A source database for a P2P system P is a function D assigning to each peer p P such that I(p) = G p , S p , M p a database instance D(p) for S p . A global database for P is a function B assigning to each peer p a database instance B(p) for G p . Usually, we are interested in global databases that can be "retrieved" from a given source, as formalized below. Given a source database D for P, a retrieved global database for D is a global database B that satisfies the mapping assertions M p of each peer p, i.e., B is such that: p P and (Q S p Q G p ) M p , it is the case that Q D(p) S p Q B(p) G p . We denote by ret ( P, D) the set of all the retrieved global databases for D in the system P. Notice that in the definition above we are considering sound mappings: data retrieved from the sources by the mapping views are assumed to be a subset of the data that satisfy the corresponding global relation. This is a classical assumption in data integration, where sources in general do not provide all the intended extensions of the global schema, hence extracted data are to be considered sound but not necessarily complete. Example 1 (contd.) Let D r be a source database for the P2P system P r such that D r (p 1 ) is {s 1 (Albert, Bill)}, D r (p 2 ) consists of {s 2 (John, Mary), s 2 (Mary, Tom)}, and D r (p 3 ) = {s 3 (Mary, D1)}. Consider also the global database B r such that B r (p 1 ) = {secretary (Albert, Bill)}, B r (p 2 ) = {financial(John, Mary), financial(Mary, Tom)} and B r (p 3 ) = {employee(Mary, D1)}. Then, it is easy to see that B r is a retrieved database for D r in P r , i.e., B r ret(P r , D r ). Note that a global database B whose peer schema for some peer p {p 1 , p 2 , p 3 } is a superset of B r (p) is in ret (P r , D r ) as well - we simply say that B is a superset of B r . P 3.2 Models of Peer-to-Peer Systems Given a source database D, it is particular important to investigate whether it is possible to retrieve from D a database which satisfies the semantics of the network. Therefore, we next define a suitable notion of model for a P2P system. The approach has been inspired by the au-toepistemic approach of [9]; in particular, we assume that peers propagate through mapping assertions only the values they really trust. Definition 2 Let P = P, I, N , map be a P2P system, p P a peer with I(p) = G p , S p , M p and G p = p , p , and D a source instance for P. Then, a p-model for P w.r.t. D is a maximal nonempty set of global databases M ret(P, D) such that: 1. for each B M, B(p) satisfies the constraints in p , and 2. for each assertion Q q Q p map(p), it holds: B M Q B (q) q B M Q B (p) p . P Thus, according to Condition 1, any databases in the p-model satisfies all the integrity constraints issued over the global schema of p; moreover, Condition 2 guarantees that peers communicate only those values that belong to all models , i.e., a cautious approach to the propagation has been pursued. Finally we point out that, as for local mapping assertions , peer mapping assertions are assumed to be sound. Now, given that each peer singles out its models, a notion of model for the whole system can be easily stated. Definition 3 Let P = P, I, N , map be a P2P system. A model for P w.r.t. D is a maximal nonempty set M ret ( P, D) of global databases such that, for each p P , M is a p-model. If a model for P w.r.t. D exists, we say that D satisfies P, denoted by D |= P. P For instance, in our running example, D r does not satisfy P r ; indeed, the peer mapping assertions constrain the schema of p 3 to contain in every global database (retrieved from D r ) the tuples boss(Albert, Bill), boss(John, Mary), boss(Mary, Tom), and employee (Mary, D1) that violate the integrity constraints over p 3 , since Mary results to be both an employee and a manger. Therefore, retrieving data from D r leads to an inconsistent scenario. We conclude by noticing that deciding whether a P2P system admits a model can be done efficiently. The result can be proven by modifying the techniques in [9], in order to first evaluate all the mappings in the network and then check for the satisfaction of the integrity constraints over peer schemas. Theorem 4 Let P = P, I, N , map be a P2P system, and D be a database instance for P. Then, deciding whether there is a model for P w.r.t. D, i.e., D |= P, is feasible in polynomial time. DEALING WITH AUTONOMOUS PEERS As shown in our running example, in general data stored in local and autonomous sources are not required to satisfy constraints expressed on the global schema (for example when a key dependency on G is violated by data retrieved from the sources). Thus, a P2P system may be unsatisfiable w.r.t. a source database D. In this section, we face the problem of solving inconsistencies in P2P systems. Specifically, we introduce a semantics for "repairing" a P2P system. To this aim, we first provide a model for peer preferences, and then show the impact of these individual preferences on the cost of reaching a global agreed repair. 4.1 Peer Preferences and Repairs Let P = P, I, N , map be a P2P system, and D be a source database instance for P. Next, we define a repair weighting function w p (P,D) for each peer p, encoding its preferences on candidate repairs of D. Formally, w p (P,D) is a polynomially-computable function assigning, to each source database instance D, a natural number that is a measure of the preference of p on having D as a repair for D (the lower the number, the more preferred the repair). As a quite simple, yet natural example of weighting function , we can consider the evaluation of the number of deletions performed to the peer's sources. In this case, we have that w p (P,D) ( D ) = |D (p) - D(p)|, which in fact corresponds to the size of the difference between D and D restricted to tuples of peer p. This weighting function is called cardinality-based in the following. Example 1 (contd.) Consider the source databases D r 1 , D r 2 , and D r 3 such that: D r 1 (p 1 ) = D r 2 (p 1 ) = D r 3 (p 1 ) = D r (p 1 ), 39 D r 1 (p 2 ) = {s 2 (John, Mary)}, D r 2 (p 2 ) = {s 2 (Mary, Tom)}, D r 3 (p 2 ) = {}, D r 1 (p 3 ) = {}, D r 2 (p 3 ) = {s 3 (Mary, D1)}, and D r 3 (p 3 ) = {s 3 (Mary, D1)}. Assume that, for each peer p, w p (P r ,D r ) ( D) = |D(p) D r (p)|, i.e., she prefers source repairs where the minimum number of tuples is deleted from D r (p). Then, w p 1 (P r ,D r ) ( D r 1 ) = w p 1 (P r ,D r ) ( D r 2 ) = w p 1 (P r ,D r ) ( D r 3 ) = 0; w p 2 (P r ,D r ) ( D r 1 ) = w p 2 (P r ,D r ) ( D r 2 ) = 1; w p 2 (P r ,D r ) ( D r 3 ) = 2; w p 3 (P r ,D r ) ( D r 1 ) = 1; w p 3 (P r ,D r ) ( D r 2 ) = w p 3 (P r ,D r ) ( D r 3 ) = 0. P The problem of solving inconsistency in "classical" data integration systems has been traditionally faced by providing a semantics in terms of the repairs of the global databases that the mapping forces to be in the semantic of the system [4, 7, 6]. Repairs are obtained by means of addition and deletion of tuples according to some minimality criterion. We next propose a generalization of these approaches to the P2P framework, which takes into account peers preferences . To this aim, we focus on finding the proper set of facts at the sources that imply as a consequence a global database satisfying all integrity constraints. Basically, such a way of proceeding allows us to easily take into account information on preferences when trying to solve inconsistency, since repairing is performed by directly focusing on those sources, whose integration has caused inconsistency. Definition 5 (Repair) Let P be a P2P system, p a peer, and D and D two source databases. We say that D is p-minimal if D |= P, and there exists no source database D such that w p (P,D) ( D ) &lt; w p (P,D) ( D ) and D |= P. Then, D is a repair for P w.r.t. D if D is p-minimal for each peer p. P Example 1 (contd.) It is easy to see that D r 1 , D r 2 , and D r 3 satisfy P r and they are both p 1 -minimal. Indeed, peer p 1 has no preferences among the three databases, since w p 1 (P r ,D r ) ( D r 1 ) = w p 1 (P r ,D r ) ( D r 2 ) = w p 1 (P r ,D r ) ( D r 3 ) = 0. Moreover, D r 1 and D r 2 are equally preferred by p 2 , whereas D r 2 and D r 3 are equally preferred by p 3 . Therefore, all peers agree on D r 2 , which is thus a repair for D r w.r.t. P r . However , neither D r 3 is p 2 -minimal, nor D r 1 is p 3 -minimal, and thus they are not repairs. P We next define the semantics of a P2P system, in terms of models for those sources on which all the peers agree. Definition 6 (Agreement) Let P = P, I, N , map be a P2P system, and D be an instance for P. The agreement for P w.r.t. D is the set of all of its models w.r.t. some repair, and will be denoted by Agr ( P, D). P Example 1 (contd.) D r 2 is p-minimal, for each peer p, and it is easy to see that the set Agr ( P r , D r ) contains all databases belonging to some model for P r w.r.t. D r 2 . In particular, it contains the supersets (satisfying the constraints) of the database B r 2 such that B r 2 (p 1 ) = {secretary (Albert, Bill)}, B r 2 (p 2 ) = {financial(Mary, Tom)} and B r 2 (p 3 ) = {boss(Albert, Bill), boss (Mary, Tom), employee(Mary, D1)}. Moreover, no other global database is in Agr ( P r , D r ). P We can finally characterize the answer to a user query in terms of the repairs for the system. Definition 7 Let P = P, I, N , map be a P2P system, let D be a source database for it, and let Q be a query over the schema of a peer p. Then, the answer to Q is the evaluation of the query over all the possible agreed databases: ans(Q, p, P, D) = B Agr(P,D) Q B(p) p . P For instance, in our running example, the answer to the user query {X | boss(X, Y )} posed over peer p 3 , which asks for all employees that have a boss, is { Albert , Mary }, since this query is evaluated over the supersets of the database B r 2 retrieved from D r 2 only. We conclude the section by noticing that Agr ( P, D) is just a formal characterization of the semantics of a P2P system. Usually, we are not interested in computing such a set; and, in fact, for practical applications, suitable techniques and optimization algorithms should be investigated to handle inconsistency at query time (in the spirit of, e.g., [10]). 4.2 The Price of Autonomy Given the framework presented so far, we are in the position of studying the effects of having autonomous peers repairing their source databases according to their own preferences . We next show that, in some cases, peers might not find an agreement on the way the repair has to be carried out. This is a somehow expected consequence of having selfish interested peers in the absence of a global coordination. Proposition 8 There exists a P2P system P and a source database D such that there is no agreement, i.e., Agr(P , D) is empty. Proof [Sketch]. Consider the P2P system P = P , I , N , map , where P consists of the peers challenger (short: c) and duplicator (short: d), that are mutually connected , i.e., N (c) = {d} and N (d) = {c}. Peer c is such that I (c) = G c , S c , M c , where the schema G c consists of predicates r c (X) and mr d (X) with constraints r c (X) r c (Y ) X = Y and r c (X) mr d (Y ) X = Y ; the source schema consists of the relation symbol s c ; and M c contains only the assertion {X | s c (X)} {X | r c (X)}. Peer d is such that I (d) = G d , S d , M d , where the schema G d consists of predicates r d (X) and mr c (X) with constraints r d (X) r d (Y ) X = Y and r d (X) mr c (Y ) X = Y ; the source schema consists of the relation symbol s d ; and M d contains only the assertion {X | s d (X)} {X | r d (X)}. Finally, map(c) contains the assertion {X | r c (X))} {X | mr c (X)}, while map(d) contains the assertion {X | r d (X))} {X | mr d (X)}. Let D be a source database for P such that D(c) = {s c (0), s c (1) } and D(d) = {s d (0), s d (1) }. We build four source databases, say D 1 , D 2 , D 3 and D 4 , that satisfy P. They are such that: D 1 (c) = {}, D 1 (d) = {s d (0) }; D 2 (c) = {}, D 2 (d) = {s d (1) }; D 3 (c) = {s c (0) }, D 3 (d) = {}; D 4 (c) = {s c (1) }, D 4 (d) = {}. Notice that all the other databases satisfying P are proper subsets of these ones. Then, by assuming that each peer wants to minimize the number of deletions in D, there exists no source database satisfying P that is both c-minimal and d-minimal. THE COMPLEXITY OF QUERY ANSWERING In the light of Proposition 8, it is particulary relevant to investigate the complexity of dealing with peer agreements 40 and query answering in such P2P data integration systems. In this section, we first present some basic problems arising in the proposed framework, and subsequently analyze their computational complexity. This analysis is a fundamental premise to devise effective and optimized implementations. 5.1 Problems Given a P2P system P and a source database D for P, we consider the following problems: RepairChecking: given a source instance D , is D a repair for P w.r.t. D? AgreementExistence: is Agr(P, D) = ? AnyAgreementComputation: compute a database B in the agreement Agr ( P, D), if any. QueryOutputTuple: given a query Q over a peer schema G p and a tuple t, is t ans(Q, p, P, D)? Intuitively, RepairChecking is the very basic problem of assessing whether a source instance at hand satisfies the data integration system. Then, AgreementExistence (and its corresponding computational version AnyAgreementComputation) asks for singling out scenarios where some agrement can be in fact computed . Finally, QueryOutputTuple represents the problem characterizing the intrinsic complexity of a query answering in the proposed framework; indeed, it is the problem of deciding the membership of a given tuple in the result of query evaluation. 5.2 Results Our first result is that checking whether all the peers are satisfied by a given source database is a difficult task that is unlikely to be feasible in polynomial time. Theorem 9 RepairChecking is co-NP-complete. Hardness holds even for cardinality-based weighting functions. Proof [Sketch]. Membership. Consider the complementary problem of deciding whether there exists a peer p such that D is not p-minimal. This problem is feasible in NP by guessing a source database D and checking in that 1. D |= P , and 2. there exists a peer p such that w p (P,D) ( D ) &lt; w p (P,D) ( D ). In particular, 1. is feasible in polynomial time because of Theorem 4, and 2. is feasible in polynomial time because our weighting functions are polynomially computable. Hardness. Recall that deciding whether a Boolean formula in conjunctive normal form = C 1 . . . C m over the variables X 1 , . . . , X n is not satisfiable, i.e., deciding whether there exists no truth assignments to the variables making each clause C j true, is a co-NP-hard problem. We built a P2P system P such that: P contains a peer x i for each variable X i , a peer c j for each clause C j , and the distinguished peer e. The source schema of x i (resp. c j ) consists of the unary relation s x i (resp. s c j ), whereas the global schema consists of the unary relation r x i (resp. r c j ). The source schema of e consists of the unary relations s e and s a , whereas its global schema consists of the unary relations r e and r a . For each source relation, say s , P() contains a local mapping assertion of the form {X | s (X)} {X | r (X)}. Each global relation of the form r x i is equipped with the constraint r x i (X 1 ) r x i (X 2 ) X 1 = X 2 , stating that each relation must contain one atom at most. Each global relation of the form r c j is equipped with the constraint r c j (tx i ) r c j (fx i ) , where is the empty disjunction , stating that for each variable x i , r c j cannot contain both tx i and fx i at the same time. Moreover, peer e has also the constraint r e (X 1 ) r a (X 2 ) X 1 = X 2 . Consider the source database D for P such that: D (x i ) = {s x i (tx i ), s x i (fx i ) }; for each x i occurring in c j , D (c j ) = {s c j (tx i ), s c j (fx i ) }; and D (e) = {s e (t), s e (f), s a (t)}. Notice that due to the constraints issued over peers schemas, any source database D , with D |= P , is such that |D (x i ) | 1, for each x i . Therefore , the restriction of D to the peers of the form x i is in one-to-one correspondence with a truth-value assignment for , denoted by (D ). Intuitively, the atom s x i (tx i ) (resp. s x i (fx i )) means that variable X i is set to true (resp. false), whereas the atom s c j (tx i ) means that the clause C j is true, witnessed by the assignment for the variable X i occurring in c j . Finally, the peers mapping assertions in P are defined as follows. For each variable X i occurring positively (resp. negatively) in the clause C j there are exactly two mappings of the form {r x i (tx i ) } {r c j (tx i ) } and {r x i (fx i ) } {r c j (fx i ) } (resp. {r x i (fx i ) } {r c j (tx i ) } and {r x i (tx i ) } {r c j (fx i ) }); moreover, for each clause C j containing variables X j 1 , ..., X j k , there exists a mapping {r c j (fx j 1 ) r c j (fx j k ) } {r e (f)}. Figure 2 shows on the upper part the dependency graph G( P ) for the formula = (X 1 X 2 ) (X 3 ) (X 1 X 3 X 4 ) (X 4 ) (X 5 X 6 X 7 ) (X 4 X 6 X 8 ). Assume that each peer wants to minimize the number of deletions in D . Then, given a source database D minimal w.r.t. each peer in P but e, we can show that the above mappings encode an evaluation of the assignment (D ). In particular, it is easy to see that (D ) is a satisfying assignment for if and only if D (e) contains the facts {s e (t), s a (t)}, i.e., one fact is deleted from the source of e only. Assume, now, that D is such that D (e) = {s e (f)}, i.e., two facts are deleted from the source of e. Then, D is also e-minimal if and only if is not satisfiable. P Given the above complexity result, one can easily see that AnyAgreementComputation is feasible in the functional version of P 2 . Indeed, we can guess in NP a source instance D, build in polynomial time a model B for P w.r.t. D (by construction in Theorem 4), and check in co-NP that D is minimal for each peer. Actually, we can do much better. In fact, we next show that the problem is complete for the polynomial time closure of NP, and thus remains at the first level of the polynomial hierarchy. Theorem 10 AnyAgreementComputation is FPNP-complete . Hardness holds even for cardinality-based weighting functions. Proof [Sketch]. Membership. The problem can be solved by processing peers in a sequential manner. For each peer in P, we can find the minimum value of the associated preference function by means of a binary search, in which at each step we guess in NP a database instance and verify that such a preference holds. After having collected the minimum values for all peers, we conclude with a final guess to get a repair D, and a subsequent check that actually each peer gets its minimum possible value for P w.r.t. D. 41 Figure 2: Constructions in Proofs of Complexity Results . Finally, a model for P w.r.t. D can be build in polynomial time (again, by construction in Theorem 4). Hardness. Let be a boolean formula in conjunctive normal form = C 1 . . . C m over the variables X 1 , . . . , X n . Assume that each clause, say C j , is equipped with a weight w j (natural number). Let be an assignment for the variables in . Its weight is the sum of the weights of all the clauses satisfied in . The problem of computing the maximum weight over any truth assignment, called MAX - WEIGHT - SAT, is FPNP-complete. Consider again the construction in Theorem 9, and modify P as follows. The source schema of peer e consists of the relation s w , whereas its global schema consists of the relations r w and r v , and of the constraint r v (X) r w (X, Y ) . The local mappings of e is {X, Y | s w (X, Y )} {X, Y | r w (X, Y )}. Moreover, for each clause c j over variables X j 1 , ..., X j k , map(e) contains the assertion {r c j (fx j 1 ) r c j (fx j k ) } {r v (fc j ) }. Let P be such a modified P2P system. Notice that G( P ) coincides with G( P ) (see again Figure 2). Consider now the database instance D for P obtained by modifying D such that D (e) contains the atoms s w (fc j , 1), s w (fc j , 2), ...s w (fc j , w j ) for each clause c j . Intuitively , peer e stores w j distinct atoms for each clause c j . Let D be a source instance that satisfies P . As in Theorem 9, the restriction of D over the variables is in one-to-one correspondence with a truth assignment for , denoted by (D ). Then, it is easy to see that peer e must delete in D all the w j distinct atoms corresponding to a clause C j that is not satisfied by the assignment (D ). Therefore , |D (e)| = i|C i is false in (D ) w i . Hence, the result easily follows, since computing the source instance that is e-minimal , say D, determines the maximum weight over any assignment for as ( i w i ) - |D(e)|. P We next focus on the AgreementExistence problem. Note that membership of this problem in P 2 is easy to proven, after the above theorem. However, the reduction for the hardness part we shall exploit here is rather different. Theorem 11 AgreementExistence is P 2 -complete. Hardness holds even for cardinality-based weighting functions. Proof [Sketch]. Membership is shown with the same line of reasoning of Theorem 10. For the hardness, consider again MAX - WEIGHT - SAT, and the P 2 -complete problem of deciding whether it has a unique solution. Let P be the P2P system built in Theorem 10, and let P be a copy of it, obtained by replacing each element (both relations and peers) r in P by r . Then, consider the system ~ P obtained as the union of P , P and a fresh peer u. Figure 2 shows the dependency graph G( ~ P ). The local schema of u is empty, while its global schema consists of the unary relation r u with the constraint n i=1 r u (bad i ) . The mapping assertions are as follows. For each variable X i in , map(u) contains {r x i (tx i ) r x i (tx i ) } {r u (bad i ) } and {r x i (fx i ) r x i (fx i ) } {r u (bad i ) }. It is worthwhile noting that, for the sake of simplicity, the mapping assertions are slightly more general than those allowed in the usual definition of P2P systems, since they involve joins among different peers. However, this is only a syntactical facility, as such a mapping can be easily simulated by introducing a suitable dummy peer. The idea of the reduction is that, if the same assignment that maximizes the weight of the satisfied clauses is selected for both P and P , then r u (bad i ) is pushed to u (for each i), thereby violating the constraint. Thus, there is a (nonempty ) agreement in ~ P if and only if there are at least two such assignments. P We conclude our investigation by observing that query answering is at least as hard as AgreementExistence. Indeed , intuitively, if peers are not able to find an agreement in an inconsistent P2P system, then the answer to any given query will be empty. Moreover, membership can be proven by the same line of reasoning of Theorem 10, and we thus get the following result. Theorem 12 QueryOutputTuple is P 2 -complete. Hardness holds even for cardinality-based weighting functions. CONCLUSIONS In this paper, we investigated some important theoretical issues in P2P data integration systems. Specifically, we introduced a setting in which peers take into account their own preferences over data sources, in order to integrate data if some inconsistency arise. This seems a natural setting for such kind of systems, which has not been previously investigated in the literature. It turns out that there are scenarios where peers do not find any agreement on the way the repair should be carried out, and where some kind of centralized coordination is required. Actually, our results show that this coordination comes with a cost and some basic problems are unlikely to be tractable. However, the complexity of the problems studied in this paper are only mildly harder than the corresponding problems in traditional data integration systems. 42 This is an important feature of our approach, that paves the way for possible easy implementations, based on available systems. In particular, the prototypical implementation appears viable with minor efforts if done on top of integration systems that exploit a declarative approach to data integration (e.g., [18], where logic programs serve as executable logic specifications for the repair computation). Indeed, our complexity results show that logic engines able to express all problems in the second level of the polynomial hierarchy, such as the DLV system [19], suffices for managing the framework, once we provide appropriate logic specifications. A number of interesting research questions arise from this work. First, it is natural to ask whether the framework can be extended to the presence of existentially quantified constraints. This can be easily done for some special syntactic fragments, such as for non key-conflicting schemas, i.e., global schemas enriched with inclusion dependencies and keys, for which decidability in the context of data integration systems has been proven in [7]. To this aim, one has to modify the algorithm in [9] to propagate information in a P2P system by accounting for mapping assertion as well as for inclusion dependencies, and eventually check that after such propagation no key has been violated. We conclude by noticing that an avenue of further research is to consider more sophisticated peer-agreement semantics, besides the Pareto-like approach described here. For instance , we may think of some applications where peers may form cooperating groups, or do not cooperate at all. Another line of research may lead to enrich the setting by further kinds of peer preferences criteria, by replacing or complementing the weighting functions proposed in this paper. Acknowledgments The work was partially supported by the European Commission under project IST-2001-33570 INFOMIX. Francesco Scarcello's work was also supported by ICAR-CNR , Rende, Italy. REFERENCES [1] Serge Abiteboul, Richard Hull, and Victor Vianu. Foundations of Databases. Addison Wesley Publ. Co., Reading, Massachussetts, 1995. [2] Marcelo Arenas, Leopoldo E. Bertossi, and Jan Chomicki. Consistent query answers in inconsistent databases. In Proc. of PODS'99, pages 6879, 1999. [3] P. Bernstein, F. Giunchiglia, A. Kementsietsidis, J. Mylopoulos, L. Serafini, and I. Zaihrayeu. Data management for peer-to-peer computing: A vision. In Workshop on the Web and Databases, WebDB, 2002. [4] Leopoldo Bertossi, Jan Chomicki, Alvaro Cortes, and Claudio Gutierrez. Consistent answers from integrated data sources. In Proc. of FQAS'02, pages 7185, 2002. [5] Leopoldo E. Bertossi and Loreto Bravo. Query answering in peer-to-peer data exchange systems. In Proc. of EDBT Workshops 2004, pages 476485, 2004. [6] Loreto Bravo and Leopoldo Bertossi. Logic programming for consistently querying data integration systems. In Proc. of IJCAI'03, pages 1015, 2003. [7] Andrea Cal`i, Domenico Lembo, and Riccardo Rosati. On the decidability and complexity of query answering over inconsistent and incomplete databases. In Proc. of PODS'03, pages 260271, 2003. [8] Andrea Cal`i, Domenico Lembo, and Riccardo Rosati. Query rewriting and answering under constraints in data integration systems. In Proc. of IJCAI'03, pages 1621, 2003. [9] Diego Calvanese, Giuseppe De Giacomo, Maurizio Lenzerini, and Riccardo Rosati. Logical foundations of peer-to-peer data integration. In Proc. of PODS'04, pages 241251, 2004. [10] Thomas Eiter, Michael Fink, Gianluigi Greco, and Domenico Lembo. Efficient evaluation of logic programs for querying data integration systems. In Proc. of ICLP'03, pages 348364, 2003. [11] Enrico Franconi, Gabriel Kuper, Andrei Lopatenko, and Luciano Serafini. A robust logical and computational characterisation of peer-to-peer database systems. In Proc. of DBISP2P'03, pages 6476, 2003. [12] Enrico Franconi, Gabriel Kuper, Andrei Lopatenko, and Ilya Zaihrayeu. A distributed algorithm for robust data sharing and updates in p2p database networks. In Proc. of P2P&DB'04, pages 446455, 2004. [13] Enrico Franconi, Gabriel Kuper, Andrei Lopatenko, and Ilya Zaihrayeu. Queries and updates in the codb peer to peer database system. In Proc. of VLDB'04, pages 12771280, 2004. [14] Gianluigi Greco, Sergio Greco, and Ester Zumpano. A logic programming approach to the integration, repairing and querying of inconsistent databases. In Proc. of ICLP'01, pages 348364. Springer, 2001. [15] Gianluigi Greco and Domenico Lembo. Data integration with prefernces among sources. In Proc. of ER'04, pages 231244, 2004. [16] Alon Y. Halevy, Zachary G. Ives, Peter Mork, and Igor Tatarinov. Piazza: data management infrastructure for semantic web applications. In Proc. of WWW'03, pages 556567, 2003. [17] Maurizio Lenzerini. Quality-aware peer-to-peer data integration. In Proc. of IQIS'04, 2004. [18] Nicola Leone, Thomas Eiter, Wolfgang Faber, Michael Fink, Georg Gottlob, Gianluigi Greco, Giovambattista Ianni, Edyta Kalka, Domenico Lembo, Maurizio Lenzerini, Vincenzino Lio, Bartosz Nowicki, Riccardo Rosati, Marco Ruzzi, Witold Staniszkis, and Giorgio Terracina. The INFOMIX system for advanced integration of incomplete and inconsistent data. In Proc. of SIGMOD'05, pages 915917, 2005. [19] Nicola Leone, Gerald Pfeifer, Wolfgang Faber, Thomas Eiter, Georg Gottlob, Simona Perri, and Francesco Scarcello. The DLV System for Knowledge Representation and Reasoning. ACM Transaction on Cumputational Logic. To appear. [20] Luciano Serafini, Fausto Giunchiglia, John Mylopoulos, and Philip A. Bernstein. Local relational model: A logical formalization of database coordination. In Fourth International and Interdisciplinary Conference on Modeling and Using Context, CONTEXT 2003, pages 286299, 2003. [21] Igor Tatarinov and Alon Halevy. Efficient query reformulation in peer data management systems. In Proc. of SIGMOD'04, pages 539550, 2004. 43

Peer-to-Peer Systems;Data Integration Systems

144

On the Discovery of Significant Statistical Quantitative Rules

In this paper we study market share rules, rules that have a certain market share statistic associated with them. Such rules are particularly relevant for decision making from a business perspective. Motivated by market share rules, in this paper we consider statistical quantitative rules (SQ rules) that are quantitative rules in which the RHS can be any statistic that is computed for the segment satisfying the LHS of the rule. Building on prior work, we present a statistical approach for learning all significant SQ rules, i.e., SQ rules for which a desired statistic lies outside a confidence interval computed for this rule. In particular we show how resampling techniques can be effectively used to learn significant rules. Since our method considers the significance of a large number of rules in parallel, it is susceptible to learning a certain number of &quot;false&quot; rules. To address this, we present a technique that can determine the number of significant SQ rules that can be expected by chance alone, and suggest that this number can be used to determine a &quot;false discovery rate&quot; for the learning procedure. We apply our methods to online consumer purchase data and report the results.

INTRODUCTION Rule discovery is widely used in data mining for learning interesting patterns. Some of the early approaches for rule learning were in the machine learning literature [11, 12, 21]. More recently there have been many algorithms [1, 25, 28, 31] proposed in the data mining literature, most of which are based on the concept of association rules [1]. While all these various approaches have been successfully used in many applications [8, 22, 24], there are still situations that these types of rules do not capture. The problem studied in this paper is motivated by market share rules, a specific type of rule that cannot be represented as association rules. Informally, a market share rule is a rule that specifies the market share of a product or a firm under some conditions. The results we report in this paper are from real user-level Web browsing data provided to us by comScore Networks. The data consists of browsing behavior of 100,000 users over 6 months. In addition to customer specific attributes, two attributes in a transaction that are used to compute the market share are the site at which a purchase was made and the purchase amount. Consider the example rules below that we discovered from the data: (1) Household Size = 3 35K &lt; Income &lt; 50K ISP = Dialup marketshare Expedia = 27.76%, support = 2.1% (2) Region = North East Household Size = 1 marketshare Expedia = 25.15%, support = 2.2% (3) Education = College Region = West 50 &lt; Household Eldest Age &lt; 55 marketshare Expedia = 2.92%, support=2.2% (4) 18 &lt; Household Eldest Age &lt; 20 marketshare Expedia = 8.16%, support = 2.4% The market share for a specific site, e.g. Expedia.com, is computed as the dollar value of flight ticket purchases (satisfying the LHS of the rule) made at Expedia.com, divided by the total dollar value of all flight ticket purchases satisfying the LHS. The discovered rules suggest that Expedia seems to be doing particularly well among the single households in the North East region (rule 2), while it cedes some market in the segment of teenagers (rule 4). Rules such as these are particularly relevant for business since they suggest natural actions that may be taken. For example, it may be worth investigating the higher market share segments to study if there is something particularly good that is being done, which is not being done in the lower market share segments. More generally, "market share" is an example of a statistic that is computed based on the segment satisfying the antecedent of the rule. Besides market share, various other quantitative statistics on the set of transactions satisfying the LHS of a rule can be computed, including mean and variance of an attribute. Prior work on learning quantitative association rules [2, 33] studied the discovery of rules with statistics such as the mean, variance, or minimum/maximum of a single attribute on the RHS of a rule. In this paper we generalize the structure of the rules considered in [2] to rules in which the RHS can be any quantitative statistic that can be computed for the subset of data satisfying the LHS. This statistic can even be computed based on multiple attributes. We term such rules as statistical quantitative rules (SQ rules). With respect to learning SQ rules from data, we formulate the problem as learning significant SQ rules that have adequate support. We define an SQ rule to be significant if the specific statistic computed for the rule lies outside a certain confidence interval. This confidence interval represents a range in which the statistic can be expected by chance alone. This is an important range to identify if the rules discovered are to be interpreted as suggesting fundamental relationships between the LHS and the market share. For example, by chance alone if it is highly likely that the market share of Expedia is between 25% and 30% for any subset of data, then it is not at all clear that the rule relating income and Expedia's market share (rule 1 in the example) is identifying a fundamental relationship between income and the market share. While prior work [6, 9] has used confidence intervals to identify significant rules, most of these approaches are either parametric or specific for binary data. Building on prior work in this paper we present a statistical approach for learning significant SQ rules that is entirely non-parametric. In particular we show how resampling techniques, such as permutation, can be effectively used to learn confidence intervals for rules. Based on these confidence intervals, significant rules can be identified. However, since our method considers the significance of a large number of rules in parallel, for a given significance level it is susceptible to learning a certain number of false rules. To address this we present an intuitive resampling technique that can determine the number of false rules, and argue that this number can be used to determine a &quot;false discovery rate&quot; for the learning procedure. The practical significance of this approach is that we learn significant SQ rules from data and specify what the false discovery rate exactly is. The paper is organized as follows. We first define SQ rules in the next section. Section 3 presents an algorithm for computing confidence intervals and Section 4 presents an algorithm for learning significant SQ rules. In Section 5 we explain how the false discovery rate for our approach can be computed. We present detailed experimental results on real web browsing data in Section 6 followed by a literature review and conclusions. STATISTICAL QUANTITATIVE RULES In this section we define SQ rules and significant SQ rules. Let A= {A 1 , A 2 ,..., A n } be a set of attributes that will be used to describe segments and B = {B 1 , B 2 ,..., B m } be another set of attributes that will be used to compute various statistics that describe the segment. Let dom(A i ) and dom(B j ) represent the set of values that can be taken by attribute A i and B j respectively, for any A i A and B j B. Let D be a dataset of N transactions where each transaction is of the form {A 1 = a 1 , A 2 = a 2 ,..., A n = a n , B 1 = b 1 , B 2 = b 2 ,..., B m = b m } where a i dom(A i ) and b j dom(B j ). Let an atomic condition be a proposition of the form value 1 A i value 2 for ordered attributes and A i = value for unordered attributes where value, value 1 , value 2 belong to the finite set of discrete values taken by A i in D. Finally, let an itemset represent a conjunction of atomic conditions. Definition 2.1 (SQ rule). Given (i) sets of attributes A and B, (ii) a dataset D and (iii) a function f that computes a desired statistic of interest on any subset of data, an SQ rule is a rule of the form: X f(D X ) = statistic, support = sup 1 (2.1) where X is an itemset involving attributes in A only, D X is the subset of D satisfying X, the function f computes some statistic from the values of the B attributes in the subset D X , and support is the number of transactions in D satisfying X. Note that the statistic on the RHS of the rule can be computed using the values of multiple attributes. The following examples are listed to demonstrate different types of rules that an SQ rule can represent. For ease of exposition we use the name of the desired statistic in the RHS instead of referring to it as f(D X ). 1. Quantitative association rules [2]: population-subset mean or variance values for the subset (2.2) Quantitative association rules are a popular representation for rules in the data mining literature in which the RHS of a rule represents the mean or variance of some attribute. Example: Education = graduate Mean(purchase) = $15.00. (2.2) is a special case of (2.1), where f(subset) is the mean of some attribute B j in the subset of data. 2. Market share rules: Let {A 1 , A 2 ,..., A n , MSV, P} be a set of attributes in a dataset D. MSV (Market Share Variable) is a special categorical attribute for which the market share values are to be computed. P is a special continuous variable that is the basis for the market share computation for MSV. For example, each transaction T k may represent a book 2 purchased online. A 1 through A n may represent attributes of the customer who makes the purchase, such as income, region of residence and household size. For each transaction, MSV is the variable indicating the online book retailer where the purchase was made. dom(MSV) may be {Amazon, Barnes&Noble, Ebay} and P is the price of the book purchased. For a specific v dom(MSV) a market share statistic can be computed as described below. Market share rules have the following form: X marketshare v = msh, support = sup (2.3) where X is an itemset consisting of attributes in {A 1 , A 2 ,..., A n } and marketshare v is a statistic that represents the market share of a specific v dom(MSV). This is computed as follows. Let D X represent the subset of transactions satisfying X and D X, MSV=v 1 In association rules, support is the number of transactions satisfying both LHS and RHS of a rule. In SQ rules, since the RHS is not an itemset, we define support as the number of transactions satisfying the LHS of a rule only. 2 The provider, comScore Networks categorizes each purchase into categories such as "book", "travel" and "consumer electronics". Hence we can generate datasets in which all transactions represent purchases in a single category, and this helps in the generation of market share rules representing specific categories. 375 Research Track Paper represent the subset of transactions satisfying (X MSV = v). Then marketshare v is computed as sum(P, D X, MSV=v ) / sum(P, D X ), where sum(P, D) is the sum of all the values of attribute P in the transactions in D. Market share rules naturally occur in various applications, including online purchases at various Web sites, sales applications, and knowledge management applications. The examples presented in the introduction are real market share rules discovered in online purchase data. The following additional examples illustrate the versatility and usefulness of market share rules. Within a specific product category (e.g. shoes) Footlocker sells competing brands of shoes. In their transaction data, the brand of the shoe can be the MSV and the purchase price is P. Consider a dataset of patents associated with some area (e.g. hard disks). Each record may consist of several attributes describing a patent, including one attribute (MSV) which represents the organization to which the patent belongs and another attribute that is always 1 (representing P and indicating a granted patent) in the data. For a specific organization, e.g. IBM, market share rules will represent the percentage of patents that belong to IBM under some conditions involving other attributes of the patent. Definition 2.1 differs from the definition of quantitative rule [2, 33] as follows. First, it is not limited to mean and variance statistics and assumes a much broader class of statistics, including the market share statistics. Second, unlike quantitative rules, the statistic of interest in the RHS of a rule can be computed based on multiple attributes. Definition 2.2 (Significant SQ rule). For a given significance level (0, 1), let (stat L , stat H ) be the (1 ) confidence interval for a desired statistic, where this confidence interval represents the range in which the statistic can be expected by chance alone. An SQ rule X f(D X ) = statistic, support = sup is significant if statistic lies outside the range (stat L , stat H ). The main objective of this paper is to discover all significant SQ rules. The first challenge in learning significant SQ rules is in constructing a confidence interval for the desired statistic such that this interval represents a range of values for the RHS statistic that can be expected by chance alone. In the next section we present an algorithm for learning these confidence intervals. COMPUTING CONF INTERVALS The first question that needs to be addressed is what is meant by "a range for the statistic that can be expected by chance alone". In this section we start by addressing this question and outline a procedure by which such a range can be computed. Next we will point out the computational challenge in implementing such a procedure for learning these intervals for several SQ rules and then outline three observations that will substantially help address the computational problems. Based on these observations we present a resampling-based algorithm for computing the confidence intervals. 3.1 Motivation and outline of a procedure For a given SQ rule, the desired confidence interval theoretically represents the range in which the statistic can be expected when there is no fundamental relationship between the LHS of the rule and the statistic. More precisely, since the statistic is computed from the values of the B attributes, the confidence interval represents the range in which the statistic can be expected when the A attributes are truly independent of the B attributes. Without making any parametric distributional assumptions, such a confidence interval can be generated using the classical nonparametric technique of permutation. Indeed permutation-based approaches have been commonly used to generate confidence intervals in the statistics literature [16]. If R is the set of all attributes in a dataset, the basic idea in permutation is to create multiple datasets by randomly permuting the values of some attributes R i R. Such a permutation would create a dataset in which R i is independent of (R R i ), but would maintain the distributions of R i and (R R i ) in the permutation dataset to be the same as the distributions of these attributes in the original dataset. Table 3.1 illustrates one example of a permutation dataset D in which the B attributes are randomly permuted. Since a desired statistic can be computed on each permutation dataset, a distribution for the statistic can be computed based on its values from the multiple permutation datasets. A confidence interval can then be computed from this distribution. Table 3.1 Dataset permutation Original dataset D: Permutation dataset D : A 1 A 2 B 1 B 2 A 1 A 2 B 1 B 2 1 2 3 8 1 2 5 6 1 3 5 6 1 3 7 4 2 3 7 4 2 3 3 8 As mentioned above, this is a commonly used procedure in nonparametric statistics. The reason this procedure makes sense is as follows. Even if there is a relationship between the LHS of an SQ rule and the statistic on the RHS, by holding the A attributes fixed and randomly re-ordering the values of the B attributes the relationship is destroyed and the A attributes and B attributes are now independent of each other. Repeating this procedure many times provides many datasets in which the A attributes and B attributes are independent of each other, while maintaining the distributions of the A and B attributes to be the same as their distributions in the original dataset. The values for the statistic computed from the many permutation datasets is used to construct a distribution for the statistic that can be expected when the A attributes are truly independent of the B attributes. Specifically, for the same itemset X, compare the following two SQ rules in D and D , D: X f( X D ) = stat D , support = sup D (3.1) D : X f( X D ) = stat D , support = sup D (3.2) First note that the supports of the rules are the same since the number of records satisfying X in the permutation dataset is the same as the original dataset. We will use this observation to build a more efficient method for computing confidence intervals 376 Research Track Paper shortly. A confidence interval for the rule in (3.1) can be computed using the following nave procedure. 1. Create permutation dataset D from the original dataset D and compute stat D (as mentioned earlier in Section 2, the function f computes this number based on the records satisfying X). 2. Repeat step 1 N perm &gt; 1000 times 3 , sort all the N perm stat D values in an ascending order (stat D -1 , stat D -2 ,..., stat D -Nperm ) and let the /2 th and (1 /2) th percentiles 4 from this list be stat D -L and stat D -H . The N perm values computed above represents a distribution for the statistic that can be expected by chance alone, while the percentile values from this distribution determine a specific confidence interval. (Below we use the terms "distribution" and "confidence interval" frequently.) 3. The (1 ) confidence interval for the SQ rule in Equation (3.1) is (stat D -L , stat D -H ). 3.2 Computational challenges and solutions Computing these confidence intervals for multiple candidate SQ rules creates several computational problems which we will address in this section. For example, if we need to test 10,000 potential significant rules (which is a reasonably small number for data mining tasks), then we would need to repeat the above steps 10,000 times, and this means generating permutation datasets 10,000 N perm &gt; 10 7 times and to compute the desired statistic in each permutation dataset. The following observations substantially reduce the computational complexity of the procedure. 1. Sampling can be used instead of creating permutation datasets. For the SQ rule in Equation (3.1), computing stat D on a permutation dataset is really equivalent to computing stat D based on a random sample of sup D records in D. This is the case since none of the A attributes play a role in the computation of the statistic. Permuting the dataset, identifying the (sup D ) records where X holds, and then computing the statistic on this subset achieves the same effect as picking a random sample of sup D records from D and computing the statistic on this random subset. Hence to determine the confidence interval for the SQ rule in Equation (3.1), instead of permuting the dataset N perm times, it is enough to sample sup D records from D for N perm times. 2. Some of the candidate SQ rules have the same support values as other rules. Based on this observation, confidence intervals for two SQ rules with the same support can be approximated by the same interval. This is the case since for a given rule the interval is generated by sampling sup D records many times and if another rule has support = sup D then the interval for that rule will be similar if the same procedure is repeated (it will not be exactly the 3 N perm is typically a big number. If we let N perm = N!, which is the number of all possible permutations, we will be implementing a Monte Carlo test. On large datasets, such a test is impractical. For a statistic like market share whose value is limited by 0 and 1, N perm &gt; 1000 makes the distribution precise to the third decimal place. In our experiments, N perm = 1999. 4 Since we do not have any prior assumption about the expected value of the statistic we use a two sided p-value. same because of randomization). Therefore, fewer confidence intervals need to be generated. 3. It is adequate to generate a fixed number of intervals, independent of the number of rules considered. We observe that the interval for an SQ rule with support = sup D can be approximated by an interval computed by sampling sup E records where sup E is "reasonably close" to sup D . This is a heuristic that we use to considerably reduce the complexity of the procedure. Denote N Rule as the number of rules to be tested. If all rules have different support values, we need to construct N Rule distributions. Instead, we would construct a fixed number N dist distributions, such that for rule "X f(D X ) = statistic, support = sup", statistic is compared with the distribution that is constructed by sampling the closest number of transactions to sup. This heuristic is more meaningful when we consider support in terms of percentage of transactions satisfying LHS of a rule, which is a number between 0 and 1. 3.3 Algorithm CIComp Based on the above observations, we present in Figure 3.1 algorithm CIComp for constructing N dist distributions and determining the (1 ) confidence intervals for a given significance level. Input: dataset D with N transactions, the number of distributions N dist , the number of points in each distribution N perm , a function f that computes the desired statistic, and significance level . Output: N dist distributions and significance thresholds. 1 for ( dist = 1; dist N dist ; dist ++ ) { 2 N sample = dist / N dist N ; 3 for ( perm = 1; perm &lt; N perm ; perm ++ ) { 4 S = N sample transactions from D sampled without replacements 5 ; 5 stat[ dist ][ perm ] = f ( S ); 6 } 7 sort(stat[ dist ]); 8 LowerCI[ dist ] = stat[ dist ][( N perm + 1) /2]; 9 UpperCI[ dist ] = stat[ dist ][( N perm + 1) (1 /2)]; 10 } 11 Output stat[][], LowerCI[], UpperCI[] Figure 3.1 Algorithm CIComp In the above algorithm, N dist , N perm , and are user-defined parameters. is usually chosen to be 5%, 2.5% or 1%. For N dist and N perm , the larger they are, the more precise the distributions will be. Let N = 1000, N dist = 100, N perm = 999, and = 5%. We use these numbers as an example to explain the algorithm. For step 2, the first distribution corresponds to N sample = dist/N dist N = 1/100 1000 = 10 transactions. Step 3 to 6 computes N perm = 999 statistics for 10 randomly sampled transactions from dataset D. Then we sort these 999 statistics and pick /2 and 1 /2 percentiles, which are the 25 th and 975 th numbers in the distribution, as the lower and upper thresholds for the (1 ) confidence interval. Steps 2 through 9 are repeated N dist = 100 times to get the desired number of distributions and confidence intervals. 5 If the sampling is done with replacement then the interval will be the bootstrap confidence interval. The two intervals will essentially be the same when the support of the itemset is small. 377 Research Track Paper The computation complexity of the algorithm in Figure 3.1 is O(N N perm N dist ), whereas the complexity of nave method is O(N N perm N rule ). Note that N dist can be fixed to a reasonable small number, e.g. 100, whereas N rule is the number of rules that are being tested and can easily be orders of magnitude more than N dist . DISCOVERING SQ RULES Given the distributions and confidence intervals, discovering all significant statistical rules is straightforward. Algorithm SigSQrules is presented in Figure 4.1. Input: dataset D with N transactions, sets of attributes A and B, N dist , stat[][], LowerCI[], and UpperCI[] from algorithm CIComp , a function f that computes the desired statistic, minimum support minsup and a large itemset generation procedure largeitemsets . Output: set of Significant rules, sigrules. 1 L = largeitemsets ( D , A , minsup ) # generates large itemsets involving attributes in A 2 sigrules = {} 3 forall (itemsets x L) { 4 x.stat = f ( D x ) // statistic computed on transactions satisfying x 5 dist = round( support(x) / N N dist ) 6 if x .stat (LowerCI[ dist ], UpperCI[ dist ]) { // x f ( D x ) = x . stat is significant 7 x.pvalue = 2 percentile of x.stat in stat[ dist ][1.. N perm ] 8 sigrules = sigrules { x f ( D x ) = x . stat , support = support ( x )} 9 } 10 } Figure 4.1 Algorithm SigSQrules Given N dist distributions constructed from the algorithm CIComp, we use the above algorithm to discover all significant SQ rules. We continue to use the example N = 1000, N dist = 100, and N perm = 999 to describe the steps in Figure 4.1. Note that the attributes in A represent the attributes in the dataset that are used to describe segments for which statistics can be computed. Step 1 uses any large itemset generation procedure in rule discovery literature to generate all large itemsets involving attributes in A. The exact procedure used will depend on whether the attributes in A are all categorical or not. If they are, then Apriori algorithm can be used to learn all large itemsets. If some of them are continuous then other methods such as the ones described in [31] can be used. Step 4 computes the statistic function for each large itemset, x. In step 5, we find out which distribution is to be used for significance test. For example, if support(x) = 23, then support(x)/N N dist = (23/1000) 100 = 2.3, and hence dist will be round(2.3) = 2. We would compare x.stat with its corresponding confidence interval (LowerCI[2], UpperCI[2]) in step 6. If x.stat is outside of the confidence interval, the rule is significant, and we use step 7 to calculate its 2-side p-value. If x.stat is the qth percentile, the 2-side p-value is 2 min(q%, 1 q%). The p-value is not only a value to understand how significant a rule is, but is also useful for determining the false discovery rate in Section 5. Note that the confidence interval used to test significance of a rule is approximate since we do not compute this interval for the exact value of the support of this rule. Instead we use the closest interval (which was pre-computed as described in Section 3.2) corresponding to this support value. In future research we will quantify the effects of this approximation. We would also like to point out that in many cases (see below) the computation of the statistic can be done efficiently within the itemset generation procedure (largeitems) itself. This can be used to modify the algorithm to make it more efficient once a specific itemset generation procedure is used. This is the case if the function f that computes the statistic on transactions T 1 , T 2 ,..., T s is a recursive function on s, that is, f(T 1 , T 2 ,..., T s ) = g(f(T 1 , T 2 ,..., T s-1 ), f(T s ), s) (4.1) Many statistics, such as mean and market share, are recursive. For example, Mean(T 1 , T 2 ,..., T s ) = [Mean(T 1 , T 2 ,..., T s 1 ) (s 1) + Mean(T s )] / s. In this section we presented an algorithm SigSQrules for generating significant SQ rules. However, as mentioned in the introduction, for any given level of significance for a rule, the fact that thousands of rules are evaluated for their significance makes it possible to discover a certain number of false rules. This is the well known multiple hypothesis testing problem [4]. While it is difficult to eliminate this problem, it is possible to quantify this effect. In the next section we discuss the problem of false discovery in detail and present an algorithm for determining the false discovery rate associated with the discovery of significant SQ rules. FALSE DISCOVERY OF SQ RULES As mentioned above, when multiple rules are tested in parallel for significance, it is possible to learn a number of "false" rules by chance alone. Indeed, this is a problem for many rule discovery methods in the data mining literature. The false discovery rate (FDR) is the expected percentage of false rules among all the discovered rules. Prior work in statistics has taken two approaches to deal with the multiple hypothesis testing problem [4, 17, 34]. One approach attempts to lower the false discovery rate by adjusting the significance level at which each rule is tested. As we will describe below, this approach is not suitable for data mining since it will result in very few rules being discovered. The second approach assumes that a given number of false discoveries should be expected, and focuses on estimating what the false discovery rate (FDR) exactly is. This is more useful for data mining, since it permits the discovery of a reasonable number of rules, but at the same time computes a FDR that can give users an idea of what percentage of the discovered rules are spurious. In this section, we first review key ideas related to the multiple hypotheses testing problem and then present a nonparametric method to determine false discovery rate for our procedure. For significance tests for a single rule, the significance level is defined as the probability of discovering a significant rule when the LHS and RHS of the rule are actually independent of each other; in other words, is the probability of a false (spurious) discovery. For example, on a random dataset where all attributes are independent, if we test 10,000 rules, then by definition of , we expect 10,000 5% = 500 false discoveries by pure chance alone. When some of the attributes are dependent on each other, as is the case for most datasets on which rule discovery methods are used, the above approach cannot be used to get an expectation for the number of false rules. In such cases, two approaches are 378 Research Track Paper possible. In statistics, a measure called familywise error rate (FWER) is defined as the probability of getting at least one false rule output. Most conventional approaches in statistics that deals with the multiple hypotheses testing problem use different methods to control FWER by lowering significance level for individual rule, ind . For example, Bonferroni-type procedures would have ind = / the number of rules tested, which is 5% / 10,000 = 5 10 -6 . However, when the number of hypotheses tested is large (as is the case in data mining algorithms), extreme low value, e.g. 5 10 -6 , will result in very few rules discovered. The other type of approach, as taken recently in [4] estimates the false discovery rate (FDR), the expectation of the proportion of false discoveries in all discoveries. Table 5.1 Confusion matrix of the number of rules Non-Significant Rules Significant Rules LHS independent of RHS a b LHS dependent on RHS c d In Table 5.1, the number of rules tested is (a + b + c + d), out of which (a + b) is the number of rules where the LHS of the rules is truly independent of the RHS, and (c + d) is the number of rules where there is a real relationship between the LHS and the RHS of the rules. The columns determine how many tested rules are output as significant or non-significant. The two terms FDR and FWER can be defined precisely as FDR = Exp(b / b + d) and FWER = Prob(b &gt;0). We adopt FDR estimation in this section because it effectively estimates false discoveries without rejecting too many discovered rules. However, the method proposed in the literature [4, 7, 35] for FDR cannot be used for large scale rule discovery because of the following two reasons: first, the assumption that statistics of the rules tested are independent from each other (which some of the approaches use) is not true. For example, rules A 1 = 1 Mean(D A1 = 1 ) and A 1 = 1 A 2 = 2 Mean(D A1 = 1 A2 = 2 ) are not independent. In fact a large number of rules are related to each other in rule discovery because their LHS share common conditions and RHS come from the same attributes. Second, methods in statistics draw conclusions based on the number of rules tested (= a + b + c + d), however, as indicated in [25], a and c are unknown values due to the filtering by support constraint. Without making any assumptions, below we present another permutation-based method to estimate the FDR for our procedure for learning significant SQ rules. Denote N sig ( ) to be the number of significant rules discovered from dataset D when the significant level = . In Table 5.1, N sig ( ) = b + d. Similar to the procedure described in Section 3, by keeping the values in attributes A intact and permuting the B attributes, we get a permutation dataset D . Since we remove any relationship between A and B attributes by this procedure, all the LHS and RHS statistic of each rule tested in D are independent. If we apply the significant rule discovery algorithm SigSQrules, the number of significant rules discovered from D when the significant level = will be one instance of false discovery, that is, N sig-perm ( ) = b. It is easy to see that by creating multiple permutation datasets, we can estimate the expectation of the number of false discoveries and thus compute a false discovery rate FDR = Exp(N sig-perm ( )) / N sig ( ). We will describe the steps how FDR( ) can be estimated in detail in the Appendix. In this section, we described the problem of multiple hypotheses testing and pointed out that for any given significance level a certain number of significant SQ rules will be discovered by chance alone. We then described an intuitive permutation based procedure to compute the false discovery rate. From a practical point of view the procedure described above can be used in conjunction with SigSQrules to discover a set of significant SQ rules and provide a number representing the percentage of these rules that are likely to be spurious. EXPERIMENTS In this section we present results from learning significant market share rules, a specific type of SQ rules. We started with user-level online purchase data gathered by comScore Networks, a market data vendor. The data consist of 100,000 users' online browsing and purchase behavior over a period of six months. The market data vendor tracks all online purchases explicitly by parsing the content of all pages delivered to each user. Starting from the raw data we created a dataset of purchases where each transaction represents a purchase made at an online retailer. Attributes of the transaction include user demographics, the site at which the purchase was made, the primary category (e.g. books, home electronics etc) of the site, the product purchased and the price paid for the product. Within a specific category, e.g. books, significant market share rules would be particularly interesting to discover. We selected many datasets with purchases belonging to each specific category and applied our method to learn several interesting significant market share rules. For space limitations we do not present all the results, but report the results for learning market share rules for the top three retailers in the online book industry. Specifically the dataset consists of all transactions in which a book was purchased at any site and we use the methods presented in the paper to learn market share rules for the top 3 sites Amazon.com, Barnes&Noble and Ebay. The total number of transactions was 26,439 records and we limit the rules to having at most five items on the LHS of a rule. 6.1 Rule Examples Among the most significant market share rules (as determined by the p-values of these rules), we picked four rules to list that were particularly interesting for each online retailer. Amazon.com (1) Education = High School marketshare Amazon = 42.72%, support = 20.7%, CI = (46.07%, 50.92%) (2) Region = West Household Size = 2 marketshare Amazon = 57.93%, support = 7.9%, CI = (44.36%, 52.50%) (3) Region = South Household Size = 4 marketshare Amazon = 38.54%, support = 5.4%, CI = (43.76%, 53.39%) (4) 35 &lt; Household Eldest Age &lt; 40 ISP = Broadband marketshare Amazon = 60.52%, support = 4.3%, CI = (42.88%, 53.99%) Barnesandnoble.com (1) Education = Graduate Household Size = 2 marketshare BN = 13.12%, support = 6.0%, CI = (16.81%, 25.68%) (2) 50 &lt; Household Eldest Age &lt; 55 Income &gt; 100K marketshare BN = 30.28%, support = 4.2%, CI = (16.05%, 26.79%) (3) Region = South Household Size = 3 Child = Yes marketshare BN = 13.27%, support = 4.2%, CI = (16.68%, 26.10%) (4) Region = South 60 &lt; Household Eldest Age &lt; 65 marketshare BN = 39.84%, support = 2.8%, CI = (15.55%, 27.10%) 379 Research Track Paper Ebay.com (1) Education = College Region = South marketshare Ebay = 8.28%, support = 6.9%, CI = (11.70%, 17.71%) (2) Education = College Region = North Central marketshare Ebay = 21.77%, support = 4.0%, CI = (11.05%, 18.29%) (3) Region = South Income &gt; 100K marketshare Ebay = 4.83%, support = 2.9%, CI = (9.54%, 20.46%) (4) 18 &lt; Household Eldest Age &lt; 20 marketshare Ebay = 27.50%, support = 2.8%, CI = (10.12%, 19.70%) Rule (4) for Amazon.com indicates that it is doing particularly well in households with middle-aged heads that have broadband access. The market share for Amazon.com in this segment lies significantly outside the confidence interval computed for the rule. On the other hand, rule (1) for Barnesandnoble.com shows that they are doing poorly selling to a segment which perhaps represents well educated couples. Given that this is a large segment (support = 6%), this rule suggests that they could try and examine why this is the case and how they can achieve greater penetration in this segment. In Ebay's case, all four rules are very interesting. Rule (4) indicates that they have high market share among teenagers, while rule (3) describes a segment they clearly have trouble penetrating. For many other categories too (travel and home electronics in particular) the significant SQ rules that we learned were highly interesting. As these examples suggest, these rules can be insightful, identify interesting segments and have significant business potential. 6.2 Varying support and significance levels To test how the methods perform as the minimum support and significance levels vary, for one site we generated significant SQ rules for many values of the minimum support and significance level parameters. Figures 6.1 and 6.2 show how the number of significant rules and the false discovery rate vary with support. As the minimum support threshold is lowered the number of significant SQ rules discovered increases. However the FDR increases as the support threshold is lowered, suggesting a tradeoff between discovering many significant rules while keeping the FDR low. A practical outcome is that it may be desirable to have higher minimum supports (to keep FDR low), but not too high that very few rules are discovered. Figures 6.3 and 6.4 illustrate a similar tradeoff for the significance level parameter. As decreases FDR is lower, but this results in fewer number of significant rules being discovered. Again, the implication is that it may be desirable to have a low (to keep FDR low) but not too low that very few rules are discovered. 0 500 1000 1500 2000 2500 0.0% 2.0% 4.0% 6.0% 8.0% 10.0% support # of s i gn i f i c an t r u l e s = 10% = 5% = 2.5% = 1% 0.0% 5.0% 10.0% 15.0% 20.0% 25.0% 30.0% 35.0% 0.0% 2.0% 4.0% 6.0% 8.0% 10.0% support FD R = 10% = 5% = 2.5% = 1% Figure 6.1. Effect of support on # of rules Figure 6.2. Effect of support on FDR 0.00% 5.00% 10.00% 15.00% 20.00% 25.00% 30.00% 0.00% 2.00% 4.00% 6.00% 8.00% 10.00% significance level FD R support = 1% support = 2% support = 5% support = 10% 0 200 400 600 800 1000 1200 1400 0.0% 2.0% 4.0% 6.0% 8.0% 10.0% significance level # o f s i g n ificant r u le s support = 1% support = 2% support = 5% support = 10% Figure 6.3. Effect of on FDR Figure 6.4. Effect of on # of rules 380 Research Track Paper 6.3 Summary results for online book retailers Based on this general tradeoff we chose minimum support of 2% and chose an of 2.5% in order to report summary results for the three sites. Table 6.1 summarizes the number of significant rules discovered and the false discovery rates of the procedure. As the values in the table and the examples above show, our procedure can be used effectively to learn a good set of significant SQ rules while keeping the false discovery rates reasonable. Table 6.1 Summary of results Web site Significant Rules False Discovery Rate Amazon 651 6.30% Barnesandnoble 393 9.67% Ebay 679 5.60% In this section we first presented compelling examples of rules discovered that illustrate the potential of learning significant market share rules. We then examined how the number of significant rules discovered and the false discovery rate changes with the support and significance level ( ) parameters. The results of this analysis suggested a tradeoff between generating significant rules and keeping the false discovery rate low. Based on this tradeoff we identified a specific value of the support and significance parameters and showed the number of rules discovered for these values. RELATED WORK We compare our work with the literature based on three aspects: rule structure, rule significance, and methodology. Rule structure. Rule discovery methods on a quantitative dataset can be traced back to [29], where rules of the form x 1 &lt; A &lt; x 2 y 1 &lt; B &lt; y 2 are discovered. [31] extends the structure to be conjunctions of multiple conditions on both antecedent and consequent of a rule, and proposes their discovery method based on the Apriori algorithm [1]. Although rules in [31] are important, partitions like y 1 &lt; B &lt; y 2 for continuous attributes on the RHS of a rule only gives partial description of the subset satisfying the LHS of the rule and partial descriptions sometimes are misleading. Observing this problem, [2] introduces a new structure where the consequent of a rule is Mean(D X ) or Variance(D X ) to summarize the behavior of the subset satisfying the antecedent. [33] further extends the form of the consequent of the rule, such that it can be of Min(D X ), Max(D X ), or Sum(D X ). Our rule structure is based on prior work: the antecedent is conjunctions of conditions, while the consequent can be any aggregate function f on multiple attributes to describe the behavior of the subset satisfying the antecedent. Rule significance. Any combination of attributes with conditions can potentially form a rule. Researchers use different measurements, e.g. support and confidence, to select only important rules from all possible rules. Based on the support and confidence framework, many metrics have been developed, such as gain [15], conviction [10], unexpectedness [27]. Although these metrics can be generalized to rules where the antecedent and consequent are both conjunctions of the form value 1 &lt; Attribute &lt; value 2 for quantitative datasets, they are not applicable for rules whose consequent is a function, such as Mean(D X ), or in general, f(D X ). To solve this non-trivial problem, we use statistical significance tests to evaluate rules, so that the consequent of a rule is not expected by chance alone. In the data mining literature, statistical significance tests are commonly used in many applications. For example, chi-square ( 2 ) is a statistic to test correlations between attributes in binary or categorical data, and it has been applied to discover correlation rules [9], actionable rules [23], and contrast sets [3, 32]. For sparse data, [35, 36] employ Fisher's Exact Test to detect anomaly patterns for disease outbreaks. As mentioned in Section 3, these two tests are special cases of our significance test when we apply our significance definition to categorical data. For quantitative rules in [2], the authors use a standard Z-test to determine the significance of inequality of means between a subset D X and its complement D D X . [33] defines a new measurement, impact, to evaluate quantitative rules, where impact can identify those groups that contribute most to some outcome, such as profits or costs. For areas other than rule discovery, standard Z-tests with log-linear models is used in Exploratory Data Analysis for OLAP data cubes [30]. Our significance test is different from the above primarily because (i) our significance definition is applicable to any user-defined aggregate function f(D X ), and (ii) we using nonparametric methods to construct distributions and confidence intervals, in which f(D X ) is expected from random effects alone. Methodology. Nonparametric statistics is philosophically related to data mining, in that both methods typically make no assumptions on distributions of data or test statistics. Even with known distribution of a statistic, nonparametric methods are useful to estimate parameters of the distribution [13]. Nonparametric methods are widely used on testing models that are built from data: as earliest in [18], the author uses randomization tests to tackle a model overfitting problem; [20] compares bootstrap and cross-validation for model accuracy estimation; for decision trees, [14, 26] use permutation tests to select attributes based on 2 2 contingency tables. Rule discovery is to learn local features, which is inherently different from models. Although we have seen methods using parametric hypothesis testing approach to learning rules from dataset [5, 6], no prior work has been found on discovering large number of rules based on nonparametric significance tests. The problem of multiple hypothesis testing/multiple comparison is well known in rule discovery, a good review of which can be found in [19]. On sparse binary data, [25] shows that with proper support and confidence control, very few false rules will be discovered. However, rule discovery on quantitative data faces much more complicated challenges, and conventional p-value adjustment methods cannot be directly applied. To solve this problem, we employ false discovery rate [4] metric to estimate the number of false rules discovered due to testing a large number of rules. In data mining, FDR has been shown useful in [7, 36] for categorical data with known number of hypotheses, and we extend it to quantitative rules with resampling methods. CONCLUSION In this paper we defined a new category of rules, SQ rules, and the significance of SQ rules, on quantitative data. Then we presented a permutation-based algorithm for learning significant SQ rules. Furthermore, we show how an explicit false discovery rate can be estimated for our procedure, which makes the approach useful from a practical perspective. We presented experiments in which we discovered market share rules, a specific 381 Research Track Paper type of SQ rules, in real online purchase datasets and demonstrated that our approach can be used to learn interesting rules from data. We would also like to point out that it is possible to compute the false discovery rate (FDR) for several possible significance levels in an efficient manner (without creating permutation datasets for each significance level). Although a detailed presentation of this is beyond the scope of this paper, in the appendix we provide an overview of how this can be done. One main advantage of being able to do this is that significant SQ rules can be discovered at a chosen significance level that is computed from some desired FDR. Hence rather than just estimating FDR we may be able to discover significant rules given a specific FDR. However this needs to be studied in greater detail in future work. REFERENCES [1] Agrawal, R. and Srikant, R., Fast Algorithms for Mining Association Rules, in Proceedings of the 20th International Conference on Very Large Databases, Santiago, Chile, 1994. [2] Aumann, Y. and Lindell, Y., A Statistical Theory for Quantitative Association Rules, in Proceedings of The Fifth ACM SIGKDD Int'l Conference on Knowledge Discovery and Data Mining, pp. 261-270, San Diego, CA, 1999. [3] Bay, S. D. and Pazzani, M. J., Detecting Change in Categorical Data: Mining Contrast Sets, in Proceedings of the Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 302 - 306, San Diego, CA, 1999. [4] Benjamini, Y. and Hochberg, Y., Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing, Journal of Royal Statistical Society B, vol. 57, iss. 1, pp. 289-300, 1995. [5] Bolton, R. and Adams, N., An Iterative Hypothesis-Testing Strategy for Pattern Discovery, in Proceedings of the Ninth ACM SIGKDD Int'l Conference on Knowledge Discovery and Data Mining, pp. 49-58, Washington, DC, 2003. [6] Bolton, R. J. and Hand, D. J., Significance Tests for Patterns in Continuous Data, in Proceedings of the 2001 IEEE International Conference on Data Mining, pp. 67-74, San Jose, CA, 2001. [7] Bolton, R. J., Hand, D. J., and Adams, N. M., Determining Hit Rate in Pattern Search, in Pattern Detection and Discovery, ESF Exploratory Workshop, pp. 36-48, London, UK, 2002. [8] Brijs, T., Swinnen, G., Vanhoof, K., and Wets, G., Using Association Rules for Product Assortment: Decisions Case Study, in Proceedings of the Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 254-260, San Diego, CA, 1999. [9] Brin, S., Motwani, R., and Silverstein, C., Beyond Market Baskets: Generalizing Association Rules to Correlations, in Proceedings of the ACM SIGMOD/PODS '97 Joint Conference, pp. 265-276, Tucson, AZ, 1997. [10] Brin, S., Motwani, R., Ullman, J. D., and Tsur, S., Dynamic Itemset Counting and Implication Rules for Market Basket Data, in Proceedings ACM SIGMOD International Conference on Management of Data (SIGMOD'97), pp. 255-264 , Tucson, AZ, 1997. [11] Clark, P. and Niblett, T., The Cn2 Induction Algorithm, Machine Learning, vol. 3, pp. 261-283, 1989. [12] Clearwater, S. and Provost, F., Rl4: A Tool for Knowledge-Based Induction, in Procs. of the Second International IEEE Conference on Tools for Artificial Intelligence, pp. 24-30, 1990. [13] Efron, B. and Tibshirani, R. J., An Introduction to the Bootstrap. New York, NY: Chapman & Hall, 1993. [14] Frank, E. and Witten, I. H., Using a Permutation Test for Attribute Selection in Decision Trees, in Proceedings of 15th Int'l Conference on Machine Learning, pp. 152-160, 1998. [15] Fukuda, T., Morimoto, Y., Morishita, S., and Tokuyama, T., Data Mining Using Two-Dimensional Optimized Association Rules: Scheme, Algorithms and Visualization, in Proceedings of the 1996 ACM SIGMOD International Conference on Management of Data (SIGMOD'96), pp. 13-23 , Montreal, Quebec, Canada, 1996. [16] Good, P., Permutation Tests: A Practical Guide to Resampling Methods for Testing Hypotheses - 2nd Edition. New York: Springer, 2000. [17] Hsu, J. C., Multiple Comparisons - Theory and Methods. London, UK: Chapman & Hall, 1996. [18] Jensen, D., Knowledge Discovery through Induction with Randomization Testing, in Proceedings of the 1991 Knowledge Discovery in Databases Workshop, pp. 148-159, Menlo Park, 1991. [19] Jensen, D. and Cohen, P. R., Multiple Comparisons in Induction Algorithms, Machine Learning, vol. 38, pp. 309-338 , 2000. [20] Kohavi, R., A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection, in Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence, pp. 1137-1143, San Mateo, CA, 1995. [21] Lee, Y., Buchanan, B. G., and Aronis, J. M., Knowledge-Based Learning in Exploratory Science: Learning Rules to Predict Rodent Carcinogenicity, Machine Learning, vol. 30, pp. 217-240, 1998. [22] Ling, C. X. and Li, C., Data Mining for Direct Marketing: Problems and Solutions, in Proceedings of the Fourth International Conference on Knowledge Discovery and Data Mining, pp. 73-79, New York, NY, 1998. [23] Liu, B., Hsu, W., and Ma, Y., Identifying Non-Actionable Association Rules, in Proceedings of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 329-334, San Francisco, CA, 2001. [24] Mani, D. R., Drew, J., Betz, A., and Datta, P., Statistics and Data Mining Techniques for Lifetime Value Modeling, in Proceedings of the Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 94-103, San Diego, CA, 1999. [25] Megiddo, N. and Srikant, R., Discovering Predictive Association Rules, in Proceedings of the Fourth International Conference on Knowledge Discovery and Data Mining, pp. 274-278, New York, NY, 1998. [26] Oates, T. and Jensen, D., Large Datasets Lead to Overly Complex Models: An Explanation and a Solution, in Proceedings of the Fourth International Conference on Knowledge Discovery and Data Mining, pp. 294-298, Menlo Park, CA, 1998. [27] Padmanabhan, B. and Tuzhilin, A., A Belief-Driven Method for Discovering Unexpected Patterns, in Proceedings of the Fourth International Conference on Knowledge Discovery and Data Mining, pp. 94-100, New York, NY, 1998. 382 Research Track Paper [28] Padmanabhan, B. and Tuzhilin, A., Small Is Beautiful: Discovering the Minimal Set of Unexpected Patterns, in Proceedings of the Sixth ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 54-63 , Boston, MA, 2000. [29] Piatesky-Shapiro, G., Discovery, Analysis, and Presentation of Strong Rules, in Knowledge Discovery in Databases, Piatesky-Shapiro, G. and Frawley, W. J., Eds. Menlo Park, CA: AAAI/MIT Press, pp. 229-248, 1991. [30] Sarawagi, S., Agrawal, R., and Megiddo, N., Discovery-Driven Exploration of Olap Data Cubes, in Proceedings of the Sixth International Conference on Extending Database Technology (EDBT'98), pp. 168-182, Valencia, Spain, 1998. [31] Srikant, R. and Agrawal, R., Mining Quantitative Association Rules in Large Relational Tables, in Proceedings of the 1996 ACM SIGMOD International Conference on Management of Data, 1996. [32] Webb, G., Butler, S., and Newlands, D., On Detecting Differences between Groups, in Proceedings of the Ninth ACM SIGKDD Int'l Conference on Knowledge Discovery and Data Mining, pp. 256-265, Washington, DC, 2003. [33] Webb, G. I., Discovering Associations with Numeric Variables, in Proceedings of The Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, 2001. [34] Westfall, P. H. and Young, S. S., Resampling-Based Multiple Testing - Examples and Methods for P-Value Adjustment. New York, NY: John Wiley & Sons, Inc, 1993. [35] Wong, W.-K., Moore, A., Cooper, G., and Wagner, M., Rule-Based Anomaly Pattern Detection for Detecting Disease Outbreaks, in Proceedings of the Eighteenth National Conference on Artificial Intelligence (AAAI-2002), Edmonton, Canada, 2002. [36] Wong, W.-K., Moore, A., Cooper, G., and Wagner, M., Bayesian Network Anomaly Pattern Detection for Disease Outbreaks, in Proceedings of the Twentieth International Conference on Machine Learning (ICML-2003), Washington, DC, 2003. APPENDIX: Discovering false discovery rates for multiple significance levels Let us continue to use the example N perm = 999 and = 5%. On the dataset D, from the algorithm SigSQrules we generate significant rules as well as each rule's p-value. Because there are N perm values in each distribution, the smallest possible p-value from the permutation tests is 1/(N perm + 1) = 0.001, and all possible p-values are S = { 1/(N perm + 1) = 0.001, 2/(N perm + 1) = 0.002, ... = 0.05 }. Let N sig [ ind ] be the number of significant rules whose pvalue is no larger than ind S. For example, if there are 50 rules whose p-value = 0.001, and 30 rules whose p-value = 0.002, then N sig [0.001] = 50 and N sig [0.002] = 50 + 30 = 80. Without further permutation tests, with N sig [] we know how many rules will be discovered if we lower the significance level from to ind . For example, if ind = 0.002, there are only N sig [0.002] = 80 rules whose p-value is no larger than ind = 0.002, therefore we expect to discover 80 rules. Similarly, for each permutation dataset D , at each significance level ind &lt; we can compute the number of significant rules and their p-values by applying SigSQrules only once. Note that all discoveries from D are false discoveries, because the relationships between A and B are removed. Let N sig-perm [i][ ind ] be the number of discoveries from permutation datasets D [i]. For example, N sig-perm [1][0.002] = 20 means we have 20 discoveries from the permutation dataset D [1] at ind = 0.002. We implement this procedure on multiple permutation datasets, and Median( N sig-perm [][ ind ]) is the estimate of false discoveries at each significance level ind on permutation datasets. Therefore, FDR( ind ) = Median(N sig-perm [][ ind ]) / N sig [ ind ]. We use Median to estimate the expectation, which conforms to nonparametric statistical considerations (median is the best estimator for expectation when the underlying distribution is unknown). Empirically, we showed in Figure 6.3 that FDR( ind ) is an increasing function on ind . It means that by decreasing ind , we can control FDR( ind ) to a smaller value. We are not always guaranteed, though, to be able to set an individual significance level such that FDR &lt; 5%. It is possible that even when we decrease ind to a level that almost no rules are discovered, FDR is still much larger than 5%. In other words, there are always a large proportion of spurious rules discovered from some datasets. For example, if attributes independent based on a test statistic, then Median(N sig-perm [][ ind ]) N sig [ ind ] for all significance levels, and FDR 1. We want to point out that this is a desirable property of our method on controlling FDR, because there are many real-world datasets whose attributes are truly independent from each other. Traditional methods cannot estimate how many rules should be discovered, but with our technique, we can draw the conclusion that, there is no rule to be discovered because none of the rules is better than chance. This nonparametric method to estimate and control FDR is applicable to quantitative datasets and broad types of rules. 383 Research Track Paper

nonparametric methods;resampling;Rule discovery;market share rules;statistical quantitative rules

145

Optimal transmission range for cluster-based wireless sensor networks with mixed communication modes

Prolonging the network lifetime is one of the most important designing objectives in wireless sensor networks (WSN). We consider a heterogeneous cluster-based WSN, which consists of two types of nodes: powerful cluster-heads and basic sensor nodes. All the nodes are randomly deployed in a specific area. To better balance the energy dissipation, we use a simple mixed communication modes where the sensor nodes can communicate with cluster-heads in either single-hop or multi-hop mode. Given the initial energy of the basic sensor nodes, we derive the optimal communication range and identify the optimal mixed communication mode to maximize the WSN's lifetime through optimizations. Moreover, we also extend our model from 2-D space to 3-D space.

INTRODUCTION A WIRELESS sensor network consists of a large amount of sensor nodes, which have wireless communication capability and some level of ability for signal processing. Distributed wireless sensor networks enable a variety of applications for sensing and controlling the physical world [1], [2]. One of the most important applications is the monitor of a specific geographical area (e.g., to detect and monitor the environmental changes in forests) by spreading a great number of wireless sensor nodes across the area [3][6]. Because of the sensor nodes' self constraints (generally tiny size, low-energy supply, weak computation ability, etc.), it is challenging to develop a scalable, robust, and long-lived wireless sensor network. Much research effort has focused on this area which result in many new technologies and methods to address these problems in recent years. The combination of clustering and data-fusion is one of the most effective approaches to construct the large-scale and energy-efficient data gathering sensor networks [7][9]. In particular, the authors of [9] develop a distributed algorithm called Low-Energy Adaptive Clustering Hierarchy (LEACH) for homogeneous sensor networks where each sensor elects itself as a cluster-head with some probability and The research reported in this paper was supported in part by the U.S. National Science Foundation CAREER Award under Grant ECS-0348694. the cluster reconfiguration scheme is used to balance the energy load. The LEACH allows only single-hop clusters to be constructed. On the other hand, in [10] we proposed the similar clustering algorithms where sensors communicate with their cluster-heads in multi-hop mode. However, in these homogeneous sensor networks, the requirement that every node is capable of aggregating data leads to the extra hardware cost for all the nodes. Instead of using homogeneous sensor nodes and the cluster reconfiguration scheme, the authors of [11] focus on the heterogeneous sensor networks in which there are two types of nodes: supernodes and basic sensor nodes. The supernodes act as the cluster-heads. The basic sensor nodes communicate with their closest cluster-heads via multi-hop mode. The authors of [11] formulate an optimization problem to minimize the network cost which depends on the nodes' densities and initial energies. In addition, The authors of [12] obtain the upper bounds on the lifetime of a sensor network through all possible routes/ communication modes. However, it is complicated to implement a distributed scheme to achieve the upper bound of the WSN lifetime because it is required to know the distance between every two sensor nodes in their scheme. The authors of [13] develop a simpler, but sub-optimal, scheme where the nodes employ the mixed communication modes: single-hop mode and multi-hop mode periodically. This mixed communication modes can better balance the energy load efficiently over WSNs. However, the authors of [13] do not obtain the optimal communication range for the multi-hop mode which is a critically important parameter for the mixed communication modes scheme. Also, their analytical model can only deal with the case of grid deployment, where the nodes are placed along the grids, without considering the random deployment. In order to further increase the network lifetime of heterogeneous WSNs by remedying the deficiencies in the aforementioned pervious work, we develop the analytical models to determine the optimal transmission range of the sensor nodes and identify the optimal communication modes in this paper. In our models, the basic sensor nodes are allowed to communicate with their cluster-heads with mixed communication Proceedings of the 2006 International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM'06) 0-7695-2593-8/06 $20.00 2006 IEEE Cluster-head Basic sensor node Zone1 Zone 2 Zone 3 Zone 4 R Mobile base station Cluster Fig. 1. An example of the wireless sensor data-gathering networks. In each round, the aircraft hovers above the cluster-heads in the monitored area to collect the aggregated data. Within each cluster, the basic sensor nodes can communicate with its cluster-head in either single-hop or multi-hop. modes instead of only multi-hop mode or only single-hop mode. Applying the derived optimal transmission range and communication modes, we also study how the other WSN parameters (e.g., the density and initial energy of the cluster-heads , etc.) affect the network lifetime through simulation experiments. Moreover, simulation results verify our analytical models. We also extend our model from 2-D space to 3-D space. The rest of the paper is organized as follows. Section II develops our proposed models and formulates the design procedure as an optimization problem. Section III solves the formulated optimization problem. Section IV presents the numerical and experimental results. Section V addresses the extended 3-D model. The paper concludes with Section VI. SYSTEM MODEL We study the following WSN scenario in this paper. A heterogeneous sensor network consisting of two types of sensor nodes, i) the more powerful but more expensive cluster-head nodes with density of 1 and ii) the inexpensive basic sensor nodes with density of 2 , is deployed in a specific area. The density of the basic sensor nodes is determined by the application requirements. A basic sensor node joins the cluster whose cluster-head is the closest in hops or distance to this basic sensor node. In each round, the cluster-heads send the aggregated data to the mobile base station (e.g., an aircraft or a satellite) after the cluster-heads receive and process the raw data from the basic sensor node. Fig. 1 shows an example of this type of sensor networks. The definition of network lifetime used in this paper is the period in rounds from the time when the network starts working to the time when the first node dies [15]. Notice that energy dissipation is not uniform over the cluster-based WSNs, implying that some nodes in specific zones (e.g., the nodes which are close to the cluster-heads need to consume more energy for relaying traffic of other nodes in multi-hop mode) drain out their energy faster than others. Thus, the network lifetime of these critical nodes decides how long the WSN can survive. Hence, maximizing the network lifetime is equivalent to minimizing the energy consumption of the critical sensor nodes if the initial energy of sensor nodes is given. In this paper, our optimization objective is not to minimize the total energy consumption by all the sensor nodes, but to minimize the energy consumption of the critical nodes to prolong the network lifetime. A. Node Architectures and Energy Models A wireless sensor node typically consists of the following three parts: 1) the sensor component, 2) the transceiver component , and 3) the signal processing component. In this paper, we have the following assumptions for each component. 1) For the sensor component: Assume that the sensor nodes sense constant amount of information every round. The energy consumed in sensing is simply E sense (l) = l, where is the power consumption for sensing a bit of data and l is the length in bits of the information which a sensor node should sense in every data-gathering round. In general the value of l is a constant. 2) For the transceiver component: We use a simple model for the radio hardware energy consumption [16]. The energy E tx (r, l) and the energy E rx (l) required for a node to transmit and receive a packet of l bits over r distance, respectively, can be expressed as follows. E tx (r, l) = (r n + )l E rx (l) = l (1) where r n accounts for the radiated power necessary to transmit over a distance of r between source and destination and is the energy dissipated in the transmitter circuit (PLLs, VCOs, etc) which depends on the digital coding, modulation, etc. The value of path loss exponent n depends on the surrounding environment [16]. In general, = 10pJ/bit/m 2 when n = 2, = 0.0013pJ/bit/m 4 when n = 4, = 50nJ/bit [9]. 3) For the signal processing component: This component conducts data-fusion. Because the signal processing component usually consists of complicated and expensive gear such as Digital Signal Proces-sors (DSP), Field Programmable Gate Arrays (FPGA), etc., the basic sensor nodes do not contain the signal processing component. Only the cluster-heads have these components and the ability for data-fusion. The energy spent in aggregating k streams of l bits raw information into a single stream is determined by E aggr (k, l) = kl (2) where the typical value of is 5nJ/bit/stream [17]. B. Mixed Communication Modes Notice that in the multi-hop mode, the closer the distance between the sensor node and its cluster-head, the more energy the sensor node consumes since the inner nodes are required Proceedings of the 2006 International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM'06) 0-7695-2593-8/06 $20.00 2006 IEEE to relay the more traffics than that for the outer nodes. On the other hand, for the case of pure single-hop mode, the basic sensor nodes which are closer to their cluster-heads dissipate less energy than those farther from the cluster-heads because the energy consumption increases as the n-th power of distance, where n is the path loss exponent. In order to balance the energy load of the basic sensor nodes, our model employs the mixed communication modes which consist of the mixed single-hop and multi-hop mode. The basic sensor nodes use single-hop communication mode in some rounds but multi-hop communication mode in the other rounds. This kind of mixed communication modes is easy to implement. For example, the cluster-head can broadcast a notifying message periodically to all member nodes to inform of which communication mode should be used for next data-gathering round. We use parameter to measure how often the single-hop mode is used. Suppose T is the total rounds that the network can perform, T s is the number of rounds that single-hop communication mode is used and T m is the number of rounds that multi-hop communication mode is employed. Let = T s /T = 1-T m /T be the frequency with which the single-hop communication mode is used. Note that = 0 means that the pure multi-hop communication mode is employed, while = 1 represents that only the single-hop communication mode is used. C. Deployment Models The sensor nodes and the cluster-heads are randomly distributed on a 2-D circular area, whose radius is A unit. We can model such random deployment (e.g., deployed by the aircraft in a large-scale mode) as a spatial Poisson point process. Specifically, the cluster-heads and basic sensor nodes in the wireless sensor network are distributed according to two independent spatial Poisson processes PP1 and PP2 with densities equal to 1 and 2 , respectively. The basic sensor nodes will join the clusters in which the cluster-heads are the closest to the sensor nodes to form Voronoi cells [14]. The 2-D plane is thus partitioned into a number of Voronoi cells which correspond to a PP1 process point. The authors of [14] have studied the moments and tail of the distributions of the number of PP2 nodes (i.e., the basic sensor nodes) connected to a particular PP1 node (i.e, a cluster-head ). Because PP1 and PP2 are homogeneous Poisson point processes, we can shift the origin to one of the PP1 nodes. Let V be the set of nodes which belong to the Voronoi cell corresponding to a PP1 node located at the origin and S (r,) be a PP2 node whose polar coordinate is ( r, ). By using the results of [14], we can get the probability that S (r,) belongs to V as follows: P r S (r,) V = e 1 r 2 (3) Let N V be the number of PP2 nodes belonging to the V. Then, the average N V can be given by E[N V ] = 2 0 0 P r S (r,) V 2 rdrd = 2 0 0 e 1 r 2 2 rdrd = 2 1 (4) where 2 rdrd denotes the number of PP2 nodes in a small area of rdrd. If we confine the cluster size to X hops (i.e., maximum number of X hops is allowed from the basic sensor node to its cluster-head), the average number of sensors which do not belong to any cluster-head, denoted by E[N O ], can be expressed as follows: E[N O ] = 1 A 2 2 0 XR P r S (r,) V 2 rdrd = 2 A 2 e 1 (XR) 2 (5) Clearly, we want a small E[N O ] with an appropriate cluster size X. E [N O ] 2 A 2 (6) where ( &gt; 0) is the percentage of sensor nodes that will not join any cluster-head. Solving Eq. (5) and Eq. (6) together, we obtain the following inequality. X - log 1 R 2 (7) The average number of sensor nodes which do not join any cluster-head is less or equal than 2 A 2 if X is given by X = - log 1 R 2 (8) Thus, we know that every cluster can be divided into X ring zones with the ring width equal to R when multi-hop communication mode is used. The average number of sensor nodes in the i-th zone of the cluster can be written as follows: E[N i ] = 2 0 iR (i-1)R P r S (r,) V 2 rdrd = 2 1 e 1 [(i-1)R] 2 - e 1 (iR) 2 (9) where i is an positive integer ranging from 1 to X. D. The Wireless Sensor Network Lifetime In our model, the nodes at the same zone will die at almost the same time. Therefore, the network lifetime is equivalent to the period from the time when the WSN begins working to the time when all nodes of a zone die. i) The basic sensor nodes The sensor nodes have the responsibility to relay the traffic from the peers laid in the outer zone in the multi-hop communication mode. We define the average number of packets Proceedings of the 2006 International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM'06) 0-7695-2593-8/06 $20.00 2006 IEEE by Y(i), which a sensor node placed in the i-th zone need to relay. Because each basic sensor node sends out a packet of sensed information per round, Y(i) is determined by the average number of nodes for which node i needs to forward messages and can be written as follows: Y(i) = X j=i+1 E[N j ] E[N i ] . (10) The energy consumption of the basic sensor nodes in the i-th zone can be written as the summation of the following 3 terms: E sensor (i, ) = (1 - ) E tx (R, l) + Y(i) E tx (R, l) +E rx (l) + E tx (iR, l) + E sense (l) (11) where (0 1) is a parameter measuring the frequency with which the single-hop communication mode will be employed. The first term of Eq. (11) represents the energy spent in the relaying the traffic for the sensor nodes in the outer zone and transmitting its own traffic when the multihop communication mode is employed. The second term of Eq. (11) is energy consumption for transmitting a packet by single-hop. The third term of Eq. (11) is the energy dissipation for sensing. In our model, because the basic sensor nodes in the same zone will consume almost the same energy in each round (i.e., sharing the same relaying traffic load when multi-hop communication mode is used), their lifetimes are equal if their initial energies are the same. Therefore, if the initial energy is the same for each basic sensor node, the sensor nodes which belong to the (arg max 1iX {E sensor (i, )})-th zone will cost the most energy and die first which decides the network lifetime. Given the initial energy, denoted by E init2 , which is carried by the basic sensor node, the network lifetime in rounds ( T ) can be written as follows: T = E init2 max 1iX E sensor (i, ) = E init2 E max () (12) where E max () max 1iX E sensor (i, ). One of our objectives is to find the optimal which is determined by = arg min 01 {E max ()} (13) ii) The cluster-heads Because the main functions of cluster-heads include (1) sensing, (2) collecting data from the basic sensor nodes, (3) aggregating the raw data, and (4) transferring the processed data to the base station, the energy consumption of cluster-heads is the sum of the energy dissipation of these four parts for the above four parts. Therefore, the energy consumption for a cluster-head, denoted by E CH , in each round can be expressed as summation of the following 4 terms: E CH = E [N ] E rx (l) + E tx (H, l ) + E sense (l) + E aggr (E [N ] + 1, l) (14) where E[N ] is the average number of basic sensor nodes in a cluster, and H is the distance between the cluster-head and the mobile base-station. The first term of Eq. (14) represents the energy consumed for receiving the packets from the basic sensor nodes. The second term of Eq. (14) is the energy spent in transmitting the aggregated information to the mobile base station. The third term of Eq. (14) denotes the energy dissipation for sensing and the fourth term of Eq. (14) means the energy consumption for aggregating ( E [N ] + 1) packets, each with l bits, into one packet of l bits. Because E[N ] = 2 / 1 , E CH depends on the ratio between 2 and 1 . The energy consumption of cluster-heads is reversely proportional to the density of cluster-heads. The larger the density of cluster-heads, the smaller the value of E CH . Thus, in order to ensure T 0 rounds network lifetime, the initial energy for the cluster-heads can be written as follows: E init1 T 0 E CH (15) E. Connectivity Because the mixed communication modes contains multihop communication mode, the communication range ( R) is required to be large enough to ensure the connectivity of the network. When the nodes are assumed to be distributed with Poisson density in a disc of a unit area, the authors of [18] derived a lower bound on the communication range ( R) to ensure the network connectivity with probability P r{conn}, which is determined by P r{conn} 1 - e -R 2 (16) Hence, the following inequality need to hold to ensure the connectivity. e -R 2 (17) where &gt; 0. By resolving Eq. (17), we obtain the minimum transmission range, denoted by R min , to ensure that the probability of connectivity is greater than (1 - ). R min = - 1 log = 1 ( 1 + 2 ) log ( 1 + 2 ) (18) F. The Optimization Problem Formulation Our objective is to find the optimal transmission range ( R) and the parameter for the mixed communication modes ( ) to maximize the network lifetime. Objective: max{T } (19) Subject to R R min (20) 0 1 (21) 0 E init2 E 0 (22) where the first constraint given by Eq. (20) is to ensure the connectivity of the network. The expression of R min depends Proceedings of the 2006 International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM'06) 0-7695-2593-8/06 $20.00 2006 IEEE on the types of deployment. The second constraint given by Eq. (21) indicates that we can use mixed communication modes. The third constraint given by Eq. (22) is due to the very limited energy carried by the basic sensor nodes. By observing Eq. (12), we can simplify the objective function as "min {E max }" and remove the constraint given by Eq. (22) by letting E init2 = E 0 . Then, the simplified optimization problem can be written as follows: Objective: min{E max } (23) Subject to R R min 0 1 SOLUTIONS FOR THE OPTIMIZATION PROBLEM First, we show that given a specified transmission range ( R), the function of E max () is convex in and the optimal = arg min 01 {E max ()} is a function of R. Because the energy consumption E sensor (i, ) of the sensor nodes in the i-th zone is a convex function of i (the proof is omitted for lack of space), the value of E max () will be achieved by the sensor nodes laid in either the 1st zone or the X-th zone, i.e., E max () = max {E sensor (1, ), E sensor (X, )} (24) Note that E sensor (1, ) is a monotonically decreasing linear function of while E sensor (X, ) is a monotonically increasing linear function of , and E sensor (1, 1) = E sensor (X, 0). Hence, we can find that the two lines corresponding to these two linear functions will intersect at the point where is within the range between 0 to 1. Clearly, the intersecting point, denoted by , yields the minimum value of E max (). Therefore, = is the optimal value when the following equation is satisfied. E sensor (1, ) = E sensor (X, ) (25) By resolving Eq. (25), we obtain the solution for Eq. (13). = arg min 01 {E max ()} = Y(1)[E tx (R) + E rx ] Y(1)[E tx (R) + E rx ] + E tx (XR) - E tx (R) = Y(1)(R n + 2) Y(1)(R n + 2) + R n (X n - 1) = Y(1)(R n + 2) Y(1)(R n + 2) + R n (X n - 1) (26) where = /, and we use E tx (R), E tx (XR), and E rx instead of E tx (R, l), E tx (XR, l), and E rx (l) since the value of l is a constant for a specific application. The factor measures to what extent R has the impact on the transmission energy consumption. For example, The transmission energy consumption is more sensitive to R when is larger. The factor also determines the cost of relaying traffic. The cost of relaying traffic increases with the increment of because receiving packets consumes more energy with a greater . 10 -5 10 0 10 5 10 4 10 3 10 2 10 1 10 -1 10 -2 10 -3 10 -4 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 * R=0.1, n=2 R=1, n=2 R=0.1, n=4 R=1, n=4 Fig. 2. The optimal against . If R n , Eq. (26) reduces to its approximated expression as follows: Y(1) Y(1) + (X n - 1) (27) By substituting Eq. (9) and (10) into Eq. (26), we have = 1-e -1(X2-1)R2 e 1R2 -1 (R n + 2) 1-e -1(X2-1)R2 e 1R2 -1 (R n + 2) + R n (X n - 1) (28) Notice that is a function of R if we substitute Eq. (8) into Eq. (28). Let 1 = 0.001, 2 = 3, and = 10 -12 . By using Eq. (28), we plot the optimal against as shown in Fig. 2. We observe from Fig. 2 that the optimal is almost 0 (i.e., pure multi-hop communication mode) if n = 4 and is small. The reason is because the energy consumption for transmission is proportional to R 4 and the term of R 4 in the first part of Eq. (1) dominates the transmission energy consumption if is small. Thus, the energy consumption in single-hop mode is much more than that in multi-hop mode. In contrast, if is large, the multi-hop mode loses its advantage over the single-hop mode because the transmission energy consumption is dominated by the constant term of in the first part of Eq. (1) and it is not sensitive to the transmission range. So far, given the communication range ( R), we obtain the minimum E max ( ) for the basic sensor nodes in order to derive the solutions for objective function of Eq. (24) as follows: E max ( ) = [R n (1 + X n ) + + ]l (29) Next, we want to identify the optimal R to minimize E max ( ) when constraint given by Eq. (20) applies. Because it is difficult to find the closed form for the optimal R , we use a numerical solutions to determine the optimal R which is detailed for some scenarios in Section IV. THE NUMERICAL AND SIMULATION EVALUATIONS For the following discussions, we set = 10pJ/bit/m 2 , = 50nJ/bit, = 5nJ/bit/stream, the packet length Proceedings of the 2006 International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM'06) 0-7695-2593-8/06 $20.00 2006 IEEE 10 -3 10 -2 10 -1 10 0 10 1 10 2 0 5 10 15 20 25 30 The density of the cluster-heads ( 1 ) The optimal transmission range (R * ) 2 / 1 =100 2 / 1 =5000 =1000 =50 =0 10 -3 10 -2 10 -1 10 0 10 1 10 2 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 The density of the cluster-heads ( 1 ) * 2 / 1 =100 2 / 1 =5000 =1000 =0 =50 (a) (b) Fig. 3. The optimal parameters versus the density of cluster-heads under different . (a) The optimal communication range R . (b) The optimal . 0.5 1 1.5 2 2.5 3 3.5 4 4.5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Transmission range in meters (R) The normalized network lifetime =0 =0.2 =0.5 =0.8 =1 Fig. 4. The normalized network lifetime versus the transmission range ( R) with = 50, 1 = 0.1, 2 / 1 = 5000, and = 0, 0.2, 0.5, 0.8, 1. l = 120 bits and the path loss exponent n = 2. We consider various scenarios with three different values of , two average numbers of basic sensor nodes in a cluster ( 2 / 1 ) and various densities of cluster-heads ( 1 ). According to the discussion in Section III, we get the numerical solutions for the optimal R , and then the value of optimal can be calculated by using Eq. (28). The numerical results of optimal R and with different and ( 2 / 1 ) are shown in Fig. 3(a) and Fig. 3(b), respectively. The optimal R decreases as 1 increases. With the same 1 , the larger , the larger the value of R . When is small (e.g., = 0 and = 50), the value of is small (e.g., &lt; 0.1). This implies that the multi-hop communication mode dominates the single-hop mode. The reason for these observations includes the following two. First, the cost of relaying traffics is small since the receiving energy is small. Second, the transmission energy is sensitive to the transmission range. We also conduct the simulation experiments to verify our analytical results. In our simulations, Minimum Transmission Energy (MTE) routing algorithm [19], which minimizes the total energy consumption for sending a packet, is used as the relaying scheme for the multi-hop communication mode. Let = 50, 1 = 0.1, and 2 / 1 = 5000. We set the initial energy of cluster-heads ( E init2 ) high enough to guarantee that the cluster-heads can have longer lifetime than the basic sensor nodes. Fig. 4 shows the simulation results of network lifetime. The plots in Fig. 4 is the average results of 1000 experiments. It shows that in most cases (i.e., = 0, 0.2, 0.5, and 0.8) the network lifetime is maximized when R = 3.5, which agrees with the numerical results shown in Fig. 3. In the following simulations, the parameters are set as follows: the initial energy of basic sensor nodes E init2 = 0.01J, the distance between the cluster-heads and the mobile base station H = 100m, l = l = 120bits and 2 = 1000. Fig. 5 shows the network lifetime changes with the average number of the basic sensor nodes in a cluster by using the optimal R and when is equal to 0, 100 and 1000. We observe that increasing the density of cluster-heads (i.e., 2 / 1 decreases given constant 2 ) does not always help to extend the network lifetime. For example, the network lifetime can be increased by 51% when the density of cluster-heads changes from 1 = 0.01 or 2 / 1 = 10 6 to 1 = 0.1 or 2 / 1 = 10 5 , while the network lifetime is almost the same when the density of cluster-heads is greater than 1 = 1 (i.e., the average of basic sensor nodes 2 / 1 10 4 ). On the other hand, Fig. 6 shows the energy consumption of a cluster-head against the average number of the basic sensor nodes. We find that the average energy consumption of a cluster-head is proportional to ( 2 / 1 ) from Fig. 6. The required initial energy of cluster-heads increases with the decrease of the density of cluster-heads ( 1 ). Thus, there is a trade-off between the network lifetime and the initial energy of the cluster-heads. 3-D WSN EXTENSION Our work can be easily extended to a 3-D space model. The differences between the 3-D space model and the 2-D space model lie in the deployment models and the connectivity models. The probability that a basic sensor node with spherical coordinate ( r, , ) belongs to a cluster-head located in the origin is P r S (r,,) V = e 1 4 3 r 3 (30) Proceedings of the 2006 International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM'06) 0-7695-2593-8/06 $20.00 2006 IEEE 10 1 10 2 10 3 10 4 10 5 10 6 600 700 800 900 1000 1100 1200 1300 The average number of basic sensor nodes in a cluster ( 2 / 1 ) Network lifetime in rounds =0 =100 =1000 Fig. 5. Optimal network lifetime under various situations where 2 = 1000. 10 1 10 2 10 3 10 4 10 5 10 6 10 -1 10 0 10 1 10 2 10 3 The average number of basic sensor nodes in a cluster ( 2 / 1 ) Energy consumption of a cluster-head in Joules =0 =100 =1000 Fig. 6. The energy consumption of cluster-heads against the number of the basic sensor nodes. The average number of basic sensor nodes in a cluster is determined by E[N V ] = 0 2 0 0 P r S (r,,) V 2 r 2 sin drdd = 0 2 0 0 e 1 4 3 r 3 2 r 2 sin drdd = 2 1 (31) In the similar way, we can obtain the number of basic sensor nodes in the i-th zone as follows: E [N V ] = 0 2 0 iR (i-1)R P r S (r,,) V 2 r 2 sin drdd = 2 1 e 1 4 3 [(i-1)R] 3 - e 1 4 3 (iR) 3 (32) To satisfy the requirement of connectivity, the minimum communication range R min can be written as follows: R min = 3 3 4( 1 + 2 ) log ( 1 + 2 ) (33) Again, we can obtain the optimal and R for the 3-D space model along the same manner as a class of the case of 2-D space WSN model. CONCLUSION We investigated the optimal transmission range for a heterogeneous cluster-based sensor network which consists of two types of nodes, the super cluster-heads and the basic sensor nodes. To balance the energy load of the basic sensor nodes, the mixed communication modes are employed. By developing the analytical models, we numerically derive the optimal transmission range R and the frequency of single-hop mode to achieve the longest network lifetime. Our analyses also showed that our proposed model can be easily extended from 2-D to 3-D. The simulation results validated our proposed analytical models. Our simulation results with the optimal R and indicated that the high density of cluster-heads is not very helpful for prolonging the network lifetime. REFERENCES [1] I. F. Akyildiz, W. Su, Y. Sankarsubramaniam, and E. Cayirci, "Wireless sensor networks: a survey," Computer Networks, 38 (2002) pp. 393-422. [2] G. J. Pottie and W. J. Kaiser, "Wireless integrated network sensors," Communications of the ACM, Vol. 43, No 5, pp 51-58, May 2000. [3] A. Cerpa, J. Elson, D. Estrin, L. Girod, M. Hamilton, and J. Zhao, "Habitat monitoring: application driver for wireless communications technology," in Proc. of ACM SIGCOMM Workshop on Data Communications in Latin America and the Caribbean, Costa Rica, April 2001. [4] J. Lundquist, D. Cayan, and M. Dettinger, "Meteorology and hydrology in yosemite national park: a sensor network application," in Proc. of Information Processing in Sensor Networks (IPSN), April, 2003. [5] A. Mainwaring, J. Polastre, R. Szewczyk, D. Culler, and J. Anderson, "Wireless sensor networks for habitat monitoring," in Proc. of WSNA'02, Atlanta, Georgia, September 28, 2002. [6] G. Tolle, D. Gay, W. Hong, J. Polastre, R. Szewczyk, D. Culler, N. Turner, K. Tu, S. Burgess, T. Dawson, and P. Buonadonna, "A macroscope in the redwoods," in 3rd ACM Conference on Embedded Networked Sensor Systems (SenSys), San Diego, November 2-4, 2005 [7] O. Younis and S. Fahmy, "Distributed clustering in ad-hoc sensor networks: a hybrid, energy-efficient approach", in Proc. of INFOCOM'04, March 2004 [8] A. Boulis, S. Ganeriwal, and M. B. Srivastava, "Aggregation in sensor networks: an energy-accuracy trade-off," Elsevier Ad Hoc Networks Journal, Vol. 1, 2003, pp. 317-331 [9] W. R. Heinzelman, A. Chandrakasan and H. Balakrishman, "Energy-efficient communication protocol for wireless microsensor networks," in Proc. Of IEEE HICSS, January 2000. [10] H. Su and X. Zhang, "Energy-efficient clustering system model and reconfiguration schemes for wireless sensor networks," in proc. of the 40th Conference on Information Sciences and Systems (CISS 2006), March 2006. [11] V. P. Mhatre, C. Rosenberg, D. Kofman, R. Mazumdar and N. Shroff, "A minmum cost heterogeneous sensor network with a lifetime constraint," IEEE Trans. on Mobile Computing, Vol.4, No.1, Jan./Feb. 2005, pp.4-15. [12] M. Bhardwaj and A. P. Chanrakasan, "Bouding the lifetime of sensor networks via optimal role assignments," in Proc. of INFOCOM'02, pp.1587-1596, 2002. [13] V. Mhatre and C. Rosenberg, "Design guidelines for wireless sensor networks: communication, clustering and aggregation", Elsevier Ad Hoc Networks Journal, Vol. 2, 2004, pp. 45-63. [14] S. G. Foss and S. A. Zuyev, "On a voronoi aggregative process related to a bivariate poisson process," Advances in Applied Probability, vol. 28, no. 4, 1996, pp. 965-981 . [15] J. H. Chang and L. Tassiulas, "Energy conserving routing in wireless ad hoc networks," in Proc. INFOCOM, Tel Aviv, Israel, March 2000, pp. 22-31. [16] T. Rappaport, Wireless Communication Priciples and Practice (2nd Edition). Upper Saddle River, N.J. Prentice Hall PTR, 2002. [17] A. Wang, W. Heinzelman, and A. Chandrakasan, "Energy-scalable protocols for battery-operated microsensor networks," in Proc. of the IEEE workshop on Signal Processing Systems (SiPS'99), pp. 483-492, 1999. [18] P. Gupta and P. R. Kumar, Critical power for asymptotic connectivity in wireless networks, in W. M. McEneany, G. Yin, Q. Zhang (Editors), Stochastic Analysis, Control, Optimization and Applications: A Volume in Honor of W. H. Fleming, Birkhauser, Boston, MA, 1998, pp. 547-566. [19] T. Shepard, "Decentralized channel management in scalable multihop spread spectrum packet radio networks," Massachusetts Inst. of Technol., Lab. for Comput. Sci., Cambridge, Tech. Rep. MIT/LCS/ TR-670, July 1995. Proceedings of the 2006 International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM'06) 0-7695-2593-8/06 $20.00 2006 IEEE

Wireless sensor networks;heterogeneous cluster-based sensor network;optimization;network lifetime;optimal transmission range;energy optimization;Voronoi cell;numerical model;clustering

146

Parallel Crawlers

In this paper we study how we can design an effective parallel crawler. As the size of the Web grows, it becomes imperative to parallelize a crawling process, in order to finish downloading pages in a reasonable amount of time. We first propose multiple architectures for a parallel crawler and identify fundamental issues related to parallel crawling. Based on this understanding, we then propose metrics to evaluate a parallel crawler, and compare the proposed architectures using 40 million pages collected from the Web. Our results clarify the relative merits of each architecture and provide a good guideline on when to adopt which architecture.

INTRODUCTION A crawler is a program that downloads and stores Web pages, often for a Web search engine. Roughly, a crawler starts off by placing an initial set of URLs, S , in a queue, where all URLs to be retrieved are kept and prioritized. From this queue, the crawler gets a URL (in some order), downloads the page, extracts any URLs in the downloaded page, and puts the new URLs in the queue. This process is repeated until the crawler decides to stop. Collected pages are later used for other applications, such as a Web search engine or a Web cache. As the size of the Web grows, it becomes more difficult to retrieve the whole or a significant portion of the Web using a single process. Therefore, many search engines often run multiple processes in parallel to perform the above task, so that download rate is maximized. We refer to this type of crawler as a parallel crawler. In this paper we study how we should design a parallel crawler, so that we can maximize its performance (e.g., download rate) while minimizing the overhead from parallelization . We believe many existing search engines already use some sort of parallelization, but there has been little scientific research conducted on this topic. Thus, little has been known on the tradeoffs among various design choices for a parallel crawler. In particular, we believe the following issues make the study of a parallel crawler challenging and interesting: Overlap: When multiple processes run in parallel to download pages, it is possible that different processes download the same page multiple times. One process may not be aware that another process has already downloaded the page. Clearly, such multiple downloads should be minimized to save network bandwidth and increase the crawler's effectiveness. Then how can we coordinate the processes to prevent overlap? Quality: Often, a crawler wants to download "important" pages first, in order to maximize the "quality" of the downloaded collection. However, in a parallel crawler, each process may not be aware of the whole image of the Web that they have collectively downloaded so far. For this reason, each process may make a crawling decision solely based on its own image of the Web (that itself has downloaded) and thus make a poor crawling decision. Then how can we make sure that the quality of the downloaded pages is as good for a parallel crawler as for a centralized one? Communication bandwidth: In order to prevent overlap , or to improve the quality of the downloaded pages, crawling processes need to periodically communicate to coordinate with each other. However, this communication may grow significantly as the number of crawling processes increases. Exactly what do they need to communicate and how significant would this overhead be? Can we minimize this communication overhead while maintaining the effectiveness of the crawler? While challenging to implement, we believe that a parallel crawler has many important advantages, compared to a single-process crawler: Scalability: Due to enormous size of the Web, it is often imperative to run a parallel crawler. A single-process crawler simply cannot achieve the required download rate in certain cases. Network-load dispersion: Multiple crawling processes of a parallel crawler may run at geographically distant locations, each downloading "geographically-adjacent" pages. For example, a process in Germany may download all European pages, while another one in Japan crawls all Asian pages. In this way, we can disperse the network load to multiple regions. In particular, this dispersion might be necessary when a single network cannot handle the heavy load from a large-scale crawl. Network-load reduction: In addition to the dispersing load, a parallel crawler may actually reduce the network load. For example, assume that a crawler in North America retrieves a page from Europe. To be downloaded by the crawler, the page first has to go through the network in Europe, then the Europe-to-North America inter-continental network and finally the network in North America. Instead, if a crawling process in Europe collects all European pages, and if another process in North America crawls all North American pages, the overall network load will be reduced , because pages go through only "local" networks. Note that the downloaded pages may need to be transferred later to a central location, so that a central index can be built. However, even in that case, we believe that the transfer can be significantly smaller than the original page download traffic, by using some of the following methods: Compression: Once the pages are collected and stored, it is easy to compress the data before sending them to a central location. Difference: Instead of sending the entire image with all downloaded pages, we may first take difference between previous image and the current one and send only this difference. Since many pages are static and do not change very often, this scheme can significantly reduce the network traffic. Summarization: In certain cases, we may need only a central index, not the original pages themselves . In this case, we may extract the necessary information for the index construction (e.g., postings list) and transfer this data only. To build an effective web crawler, we clearly need to address many more challenges than just parallelization. For example, a crawler needs to figure out how often a page changes and how often it would revisit the page in order to maintain the page up to date [7, 10]. Also, it has to make sure that a particular Web site is not flooded with its HTTP requests during a crawl [17, 12, 24]. In addition, it has to carefully select what page to download and store in its limited storage space in order to make the best use of its stored collection of pages [9, 5, 11]. While all of these issues are important, we focus on the crawler parallelization in this paper, because this problem has been paid significantly less attention than the others. In summary, we believe a parallel crawler has many advantages and poses interesting challenges. In particular, we believe our paper makes the following contributions: We identify major issues and problems related to a parallel crawler and discuss how we can solve these problems. We present multiple techniques for a parallel crawler and discuss their advantages and disadvantages. As far as we know most of these techniques have not been described in open literature. (Very little is known about the internals of crawlers, as they are closely guarded secrets.) Using a large dataset (40M web pages) collected from the Web, we experimentally compare the design choices and study their tradeoffs quantitatively. We propose various optimization techniques that can minimize the coordination effort between crawling processes , so that they can operate more independently while maximizing their effectiveness. 1.1 Related work Web crawlers have been studied since the advent of the Web [18, 23, 4, 22, 14, 6, 19, 12, 9, 5, 11, 10, 7]. These studies can be roughly categorized into one of the following topics: General architecture [22, 14, 6, 19, 12]: The work in this category describes the general architecture of a Web crawler and studies how a crawler works. For example , Reference [14] describes the architecture of the Compaq SRC crawler and its major design goals. Some of these studies briefly describe how the crawling task is parallelized. For instance, Reference [22] describes a crawler that distributes individual URLs to multiple machines, which download Web pages in parallel. The downloaded pages are then sent to a central machine , on which links are extracted and sent back to the crawling machines. However, these studies do not carefully compare various issues related to a parallel crawler and how design choices affect performance. In this paper, we first identify multiple techniques for a parallel crawler and compare their relative merits using real Web data. Page selection [9, 5, 11]: Since many crawlers can download only a small subset of the Web, crawlers need to carefully decide what page to download. By retrieving "important" or "relevant" pages early, a crawler may improve the "quality" of the downloaded pages. The studies in this category explore how a crawler can discover and identify "important" pages early, and propose various algorithms to achieve this goal. In our paper , we study how parallelization affects some of these techniques and explain how we can fix the problems introduced by parallelization. Page update [10, 7]: Web crawlers need to update the downloaded pages periodically, in order to maintain the pages up to date. The studies in this category discuss various page revisit policies to maximize the "freshness" of the downloaded pages. For example, Reference [7] studies how a crawler should adjust revisit frequencies for pages when the pages change at different rates. We believe these studies are orthogonal to what we discuss in this paper. There also exists a significant body of literature studying the general problem of parallel and distributed computing [21, 25]. Some of these studies focus on the design of efficient parallel algorithms. For example, References [20, 16] 125 present various architectures for parallel computing, propose algorithms that solve various problems (e.g., finding maximum cliques) under the architecture, and study the complexity of the proposed algorithms. While the general principles described are being used in our work, 1 none of the existing solutions can be directly applied to the crawling problem. Another body of literature designs and implements distributed operating systems, where a process can use distributed resources transparently (e.g., distributed memory, distributed file systems) [25, 1]. Clearly, such OS-level support makes it easy to build a general distributed application , but we believe that we cannot simply run a centralized crawler on a distributed OS to achieve parallelism. A web crawler contacts millions of web sites in a short period of time and consumes extremely large network, storage and memory resources. Since these loads push the limit of existing hardwares, the task should be carefully partitioned among processes and they should be carefully coordinated. Therefore, a general-purpose distributed operating system that does not understand the semantics of web crawling will lead to unacceptably poor performance. ARCHITECTURE OF A PARALLEL CRAWLER In Figure 1 we illustrate the general architecture of a parallel crawler. A parallel crawler consists of multiple crawling processes, which we refer to as C-proc's. Each C-proc performs the basic tasks that a single-process crawler conducts. It downloads pages from the Web, stores the pages locally, extracts URLs from the downloaded pages and follows links. Depending on how the C-proc's split the download task, some of the extracted links may be sent to other C-proc's. The C-proc's performing these tasks may be distributed either on the same local network or at geographically distant locations. Intra-site parallel crawler: When all C-proc's run on the same local network and communicate through a high speed interconnect (such as LAN), we call it an intra-site parallel crawler. In Figure 1, this scenario corresponds to the case where all C-proc's run only on the local network on the top. In this case, all C-proc's use the same local network when they download pages from remote Web sites. Therefore, the network load from C-proc's is centralized at a single location where they operate. Distributed crawler: When C-proc's run at geographically distant locations connected by the Internet (or a wide area network), we call it a distributed crawler. For example, one C-proc may run in the US, crawling all US pages, and another C-proc may run in France, crawling all European pages. As we discussed in the introduction, a distributed crawler can disperse and even reduce the load on the overall network. When C-proc's run at distant locations and communicate through the Internet, it becomes important how often and how much C-proc's need to communicate. The bandwidth between C-proc's may be limited and 1 For example, we may consider that our proposed solution is a variation of "divide and conquer" approach, since we partition and assign the Web to multiple processes. C-proc . . . C-proc Local connect C-proc collected pages queues of URLs to visit . . . C-proc Local connect NET INTER Figure 1: General architecture of a parallel crawler 1 S 2 S 1 2 (C ) (C ) b a c d e f g h i Figure 2: Site S 1 is crawled by C 1 and site S 2 is crawled by C 2 sometimes unavailable, as is often the case with the Internet. When multiple C-proc's download pages in parallel, different C-proc's may download the same page multiple times. In order to avoid this overlap, C-proc's need to coordinate with each other on what pages to download. This coordination can be done in one of the following ways: Independent: At one extreme, C-proc's may download pages totally independently without any coordination. That is, each C-proc starts with its own set of seed URLs and follows links without consulting with other C-proc's. In this scenario, downloaded pages may overlap , but we may hope that this overlap will not be significant , if all C-proc's start from different seed URLs. While this scheme has minimal coordination overhead and can be very scalable, we do not directly study this option due to its overlap problem. Later we will consider an improved version of this option, which significantly reduces overlap. Dynamic assignment: When there exists a central coordinator that logically divides the Web into small partitions (using a certain partitioning function) and dy-namically assigns each partition to a C-proc for download , we call it dynamic assignment. For example, assume that a central coordinator partitions the Web by the site name of a URL. That is, pages in the same site (e.g., http://cnn.com/top. html and http://cnn.com/content.html) belong to 126 the same partition, while pages in different sites belong to different partitions. Then during a crawl, the central coordinator constantly decides on what partition to crawl next (e.g., the site cnn.com) and sends URLs within this partition (that have been discovered so far) to a C-proc as seed URLs. Given this request, the C-proc downloads the pages and extracts links from them. When the extracted links point to pages in the same partition (e.g., http://cnn.com/article.html), the C-proc follows the links, but if a link points to a page in another partition (e.g., http://nytimes.com/ index.html), the C-proc reports the link to the central coordinator. The central coordinator later uses this link as a seed URL for the appropriate partition. Note that the Web can be partitioned at various gran-ularities . At one extreme, the central coordinator may consider every page as a separate partition and assign individual URLs to C-proc's for download. In this case, a C-proc does not follow links, because different pages belong to separate partitions. It simply reports all extracted URLs back to the coordinator. Therefore, the communication between a C-proc and the central coordinator may vary dramatically, depending on the granularity of the partitioning function. Static assignment: When the Web is partitioned and assigned to each C-proc before they start to crawl, we call it static assignment. In this case, every C-proc knows which C-proc is responsible for which page during a crawl, and the crawler does not need a central coordinator. We will shortly discuss in more detail how C-proc's operate under this scheme. In this paper, we mainly focus on static assignment because of its simplicity and scalability, and defer the study of dynamic assignment to future work. Note that in dynamic assignment, the central coordinator may become the major bottleneck, because it has to maintain a large number of URLs reported from all C-proc's and has to constantly coordinate all C-proc's. Thus the coordinator itself may also need to be parallelized. CRAWLING MODES FOR STATIC ASSIGNMENT Under static assignment, each C-proc is responsible for a certain partition of the Web and has to download pages within the partition. However, some pages in the partition may have links to pages in another partition. We refer to this type of link as an inter-partition link. To illustrate how a C-proc may handle inter-partition links, we use Figure 2 as our example. In the figure, we assume two C-proc's, C 1 and C 2 , are responsible for sites S 1 and S 2 , respectively. For now, we assume that the Web is partitioned by sites and that the Web has only S 1 and S 2 . Also, we assume that each C-proc starts its crawl from the root page of each site, a and f. 1. Firewall mode: In this mode, each C-proc downloads only the pages within its partition and does not follow any inter-partition link. All inter-partition links are ignored and thrown away. For example, the links a g, c g and h d in Figure 2 are ignored and thrown away by C 1 and C 2 . In this mode, the overall crawler does not have any overlap in the downloaded pages, because a page can be downloaded by only one C-proc, if ever. However, the overall crawler may not download all pages that it has to download, because some pages may be reachable only through inter-partition links. For example, in Figure 2, C 1 can download a, b and c, but not d and e, because they can be reached only through h d link. However, C-proc's can run quite independently in this mode, because they do not conduct any run-time coordination or URL exchanges. 2. Cross-over mode: Primarily, each C-proc downloads pages within its partition, but when it runs out of pages in its partition, it also follows inter-partition links. For example, consider C 1 in Figure 2. Process C 1 first downloads pages a, b and c by following links from a. At this point, C 1 runs out of pages in S 1 , so it follows a link to g and starts exploring S 2 . After downloading g and h, it discovers a link to d in S 1 , so it comes back to S 1 and downloads pages d and e. In this mode, downloaded pages may clearly overlap (pages g and h are downloaded twice), but the overall crawler can download more pages than the firewall mode (C 1 downloads d and e in this mode). Also, as in the firewall mode, C-proc's do not need to communicate with each other, because they follow only the links discovered by themselves. 3. Exchange mode: When C-proc's periodically and incrementally exchange inter-partition URLs, we say that they operate in an exchange mode. Processes do not follow inter-partition links. For example, C 1 in Figure 2 informs C 2 of page g after it downloads page a (and c) and C 2 transfers the URL of page d to C 1 after it downloads page h. Note that C 1 does not follow links to page g. It only transfers the links to C 2 , so that C 2 can download the page. In this way, the overall crawler can avoid overlap, while maximizing coverage. Note that the firewall and the cross-over modes give C-proc's much independence (C-proc's do not need to communicate with each other), but they may download the same page multiple times, or may not download some pages. In contrast, the exchange mode avoids these problems but requires constant URL exchange between C-proc's. 3.1 URL exchange minimization To reduce URL exchange, a crawler based on the exchange mode may use some of the following techniques: 1. Batch communication: Instead of transferring an inter-partition URL immediately after it is discovered, a C-proc may wait for a while, to collect a set of URLs and send them in a batch. That is, with batching, a C-proc collects all inter-partition URLs until it downloads k pages. Then it partitions the collected URLs and sends them to an appropriate C-proc. Once these URLs are transferred, the C-proc then purges them and starts to collect a new set of URLs from the next downloaded pages. Note that a C-proc does not maintain the list of all inter-partition URLs discovered so far. It only maintains the list of inter-partition links 127 in the current batch, in order to minimize the memory overhead for URL storage. This batch communication has various advantages over incremental communication. First, it incurs less communication overhead, because a set of URLs can be sent in a batch, instead of sending one URL per message . Second, the absolute number of exchanged URLs will also decrease. For example, consider C 1 in Figure 2. The link to page g appears twice, in page a and in page c. Therefore, if C 1 transfers the link to g after downloading page a, it needs to send the same URL again after downloading page c. 2 In contrast, if C 1 waits until page c and sends URLs in batch, it needs to send the URL for g only once. 2. Replication: It is known that the number of incoming links to pages on the Web follows a Zipfian distribution [3, 2, 26]. That is, a small number of Web pages have an extremely large number of links pointing to them, while a majority of pages have only a small number of incoming links. Thus, we may significantly reduce URL exchanges, if we replicate the most "popular" URLs at each C-proc (by most popular, we mean the URLs with most incoming links) and stop transferring them between C-proc's . That is, before we start crawling pages, we identify the most popular k URLs based on the image of the Web collected in a previous crawl. Then we replicate these URLs at each C-proc, so that the C-proc's do not exchange them during a crawl. Since a small number of Web pages have a large number of incoming links, this scheme may significantly reduce URL exchanges between C-proc's, even if we replicate a small number of URLs. Note that some of the replicated URLs may be used as the seed URLs for a C-proc. That is, if some URLs in the replicated set belong to the same partition that a C-proc is responsible for, the C-proc may use those URLs as its seeds rather than starting from other pages. Also note that it is possible that each C-proc tries to discover popular URLs on the fly during a crawl, instead of identifying them based on the previous image . For example, each C-proc may keep a "cache" of recently seen URL entries. This cache may pick up "popular" URLs automatically, because the popular URLs show up repeatedly. However, we believe that the popular URLs from a previous crawl will be a good approximation for the popular URLs in the current Web; Most popular Web pages (such as Yahoo!) maintain their popularity for a relatively long period of time, even if their exact popularity may change slightly. 3.2 Partitioning function So far, we have mainly assumed that the Web pages are partitioned by Web sites. Clearly, there exists a multitude of ways to partition the Web, including the following: 1. URL-hash based: Based on the hash value of the URL of a page, we assign the page to a C-proc. In 2 When it downloads page c, it does not remember whether the link to g has been already sent. this scheme, pages in the same site can be assigned to different C-proc's. Therefore, the locality of link structure 3 is not reflected in the partition, and there will be many inter-partition links. 2. Site-hash based: Instead of computing the hash value on an entire URL, we compute the hash value only on the site name of a URL (e.g., cnn.com in http: //cnn.com/index.html) and assign the page to a C-proc . In this scheme, note that the pages in the same site will be allocated to the same partition. Therefore, only some of the inter-site links will be inter-partition links, and thus we can reduce the number of inter-partition links quite significantly compared to the URL-hash based scheme. 3. Hierarchical: Instead of using a hash-value, we may partition the Web hierarchically based on the URLs of pages. For example, we may divide the Web into three partitions (the pages in the .com domain, .net domain and all other pages) and allocate them to three C-proc's. Even further, we may decompose the Web by language or country (e.g., .mx for Mexico). Because pages hosted in the same domain or country may be more likely to link to pages in the same domain, scheme may have even fewer inter-partition links than the site-hash based scheme. In this paper, we do not consider the URL-hash based scheme, because it generates a large number of inter-partition links. When the crawler uses URL-hash based scheme, C-proc's need to exchange much larger number of URLs (exchange mode), and the coverage of the overall crawler can be much lower (firewall mode). In addition, in our later experiments, we will mainly use the site-hash based scheme as our partitioning function. We chose this option because it is much simpler to implement, and because it captures the core issues that we want to study. For example, under the hierarchical scheme, it is not easy to divide the Web into equal size partitions, while it is relatively straightforward under the site-hash based scheme. 4 Also, we believe we can interpret the results from the site-hash based scheme as the upper/lower bound for the hierarchical scheme. For instance, assuming Web pages link to more pages in the same domain, the number of inter-partition links will be lower in the hierarchical scheme than in the site-hash based scheme (although we could not confirm this trend in our experiments). In Figure 3, we summarize the options that we have discussed so far. The right-hand table in the figure shows more detailed view on the static coordination scheme. In the diagram, we highlight the main focus of our paper with dark grey. That is, we mainly study the static coordination scheme (the third column in the left-hand table) and we use the site-hash based partitioning for our experiments (the second row in the second table). However, during our discussion , we will also briefly explore the implications of other 3 According to our experiments, about 90% of the links in a page point to pages in the same site on average. 4 While the sizes of individual Web sites vary, the sizes of partitions are similar, because each partition contains many Web sites and their average sizes are similar among partitions . 128 Type URL-hash Site-hash Hierarchical Distributed Intra-site Independent Dynamic Static Batch Replication None Main focus Also discussed Partitioning Exchange Coordination Firewall Cross-over Mode Not discussed further Figure 3: Summary of the options discussed options. For instance, the firewall mode is an "improved" version of the independent coordination scheme (in the first table), so our study on the firewall mode will show the implications of the independent coordination scheme. Also, we roughly estimate the performance of the URL-hash based scheme (first row in the second table) when we discuss the results from the site-hash based scheme. Given our table of crawler design space, it would be very interesting to see what options existing search engines selected for their own crawlers. Unfortunately, this information is impossible to obtain in most cases because companies consider their technologies proprietary and want to keep them secret. The only two crawler designs that we know of are the prototype Google crawler [22] (when it was developed at Stanford) and the Mercator crawler [15] at Compaq. The prototype google crawler used the intra-site, static and site-hash based scheme and ran in exchange mode [22]. The Mercator crawler uses the site-based hashing scheme. EVALUATION MODELS In this section, we define metrics that will let us quantify the advantages or disadvantages of different parallel crawling schemes. These metrics will be used later in our experiments . 1. Overlap: When multiple C-proc's are downloading Web pages simultaneously, it is possible that different C-proc's download the same page multiple times. Multiple downloads of the same page are clearly undesirable . More precisely, we define the overlap of downloaded pages as N-I I . Here, N represents the total number of pages downloaded by the overall crawler, and I represents the number of unique pages downloaded, again, by the overall crawler. Thus, the goal of a parallel crawler is to minimize the overlap. Note that a parallel crawler does not have an overlap problem, if it is based on the firewall mode (Section 3, Item 1) or the exchange mode (Section 3, Item 3). In these modes, a C-proc downloads pages only within its own partition, so the overlap is always zero. 2. Coverage: When multiple C-proc's run independently, it is possible that they may not download all pages that they have to. In particular, a crawler based on the firewall mode (Section 3, Item 1) may have this problem, because its C-proc's do not follow inter-partition links nor exchange the links with others. To formalize this notion, we define the coverage of downloaded pages as I U , where U represents the total number of pages that the overall crawler has to download , and I is the number of unique pages downloaded by the overall crawler. For example, in Figure 2, if C 1 downloaded pages a, b and c, and if C 2 downloaded pages f through i, the coverage of the overall crawler is 7 9 = 0.77, because it downloaded 7 pages out of 9. 3. Quality: Often, crawlers cannot download the whole Web, and thus they try to download an "important" or "relevant" section of the Web. For example, a crawler may have storage space only for 1 million pages and may want to download the "most important" 1 million pages. To implement this policy, a crawler needs a notion of "importance" of pages, often called an importance metric [9]. For example, let us assume that the crawler uses backlink count as its importance metric. That is, the crawler considers a page p important when a lot of other pages point to it. Then the goal of the crawler is to download the most highly-linked 1 million pages. To achieve this goal, a single-process crawler may use the following method [9]: The crawler constantly keeps track of how many backlinks each page has from the pages that it has already downloaded, and first visits the page with the highest backlink count. Clearly, the pages downloaded in this way may not be the top 1 million pages, because the page selection is not based on the entire Web, only on what has been seen so far. Thus, we may formalize the notion of "quality" of downloaded pages as follows [9]: First, we assume a hypothetical oracle crawler, which knows the exact importance of every page under a certain importance metric. We assume that the oracle crawler downloads the most important N pages in total , and use P N to represent that set of N pages. We also use A N to represent the set of N pages that an actual crawler would download, which would not be necessarily the same as P N . Then we define |A N P N | |P N | as the quality of downloaded pages by the actual crawler. Under this definition, the quality represents the fraction of the true top N pages that are downloaded by 129 the crawler. Note that the quality of a parallel crawler may be worse than that of a single-process crawler, because many importance metrics depend on the global structure of the Web (e.g., backlink count). That is, each C-proc in a parallel crawler may know only the pages that are downloaded by itself, and thus have less information on page importance than a single-process crawler does. On the other hand, a single-process crawler knows all pages it has downloaded so far. Therefore, a C-proc in a parallel crawler may make a worse crawling decision than a single-process crawler. In order to avoid this quality problem, C-proc's need to periodically exchange information on page importance. For example, if the backlink count is the importance metric, a C-proc may periodically notify other C-proc's of how many pages in its partition have links to pages in other partitions. Note that this backlink exchange can be naturally incorporated in an exchange mode crawler (Section 3, Item 3). In this mode, crawling processes exchange inter-partition URLs periodically, so a C-proc can simply count how many inter-partition links it receives from other C-proc's, to count backlinks originating in other partitions. More precisely, if the crawler uses the batch communication technique (Section 3.1, Item 1), process C 1 would send a message like [http://cnn. com/index.html, 3] to C 2 , to notify that C 1 has seen 3 links to the page in the current batch. 5 On receipt of this message, C 2 then increases the backlink count for the page by 3 to reflect the inter-partition links. By incorporating this scheme, we believe that the quality of the exchange mode will be better than that of the firewall mode or the cross-over mode. However, note that the quality of an exchange mode crawler may vary depending on how often it exchanges backlink messages. For instance, if crawling processes exchange backlink messages after every page download , they will have essentially the same backlink information as a single-process crawler does. (They know backlink counts from all pages that have been downloaded .) Therefore, the quality of the downloaded pages would be virtually the same as that of a single-process crawler. In contrast, if C-proc's rarely exchange backlink messages, they do not have "accurate" backlink counts from downloaded pages, so they may make poor crawling decisions, resulting in poor quality . Later, we will study how often C-proc's should exchange backlink messages in order to maximize the quality. 4. Communication overhead: The C-proc's in a parallel crawler need to exchange messages to coordinate their work. In particular, C-proc's based on the exchange mode (Section 3, Item 3) swap their inter-partition URLs periodically. To quantify how much communication is required for this exchange, we define communication overhead as the average number of inter-partition URLs exchanged per downloaded page. 5 If the C-proc's send inter-partition URLs incrementally after every page, the C-proc's can send the URL only, and other C-proc's can simply count these URLs. Mode Coverage Overlap Quality Communication Firewall Bad Good Bad Good Cross-over Good Bad Bad Good Exchange Good Good Good Bad Table 1: Comparison of three crawling modes For example, if a parallel crawler has downloaded 1,000 pages in total and if its C-proc's have exchanged 3,000 inter-partition URLs, its communication overhead is 3,000/1,000 = 3. Note that a crawler based on the the firewall and the cross-over mode do not have any communication overhead, because they do not exchange any inter-partition URLs. In Table 1, we compare the relative merits of the three crawling modes (Section 3, Items 13). In the table, "Good" means that the mode is expected to perform relatively well for that metric, and "Bad" means that it may perform worse compared to other modes. For instance, the firewall mode does not exchange any inter-partition URLs (Communication : Good) and downloads pages only once (Overlap: Good), but it may not download every page (Coverage: Bad). Also, because C-proc's do not exchange inter-partition URLs, the downloaded pages may be of lower quality than those of an exchange mode crawler. Later, we will examine these issues more quantitatively through experiments based on real Web data. DESCRIPTION OF DATASET We have discussed various issues related to a parallel crawler and identified multiple alternatives for its architecture . In the remainder of this paper, we quantitatively study these issues through experiments conducted on real Web data. In all of the following experiments, we used 40 million Web pages in our Stanford WebBase repository. Because the property of this dataset may significantly impact the result of our experiments, readers might be interested in how we collected these pages. We downloaded the pages using our Stanford WebBase crawler in December 1999 in the period of 2 weeks. In downloading the pages, the WebBase crawler started with the URLs listed in Open Directory (http://www.dmoz.org), and followed links. We decided to use the Open Directory URLs as seed URLs, because these pages are the ones that are considered "important" by its maintainers. In addition, some of our local WebBase users were keenly interested in the Open Directory pages and explicitly requested that we cover them. The total number of URLs in the Open Directory was around 1 million at that time. Then conceptually, the WebBase crawler downloaded all these pages, extracted URLs within the downloaded pages, and followed links in a breadth-first manner. (The WebBase crawler uses various techniques to expedite and prioritize crawling process, but we believe these optimizations do not affect the final dataset significantly.) Our dataset is relatively "small" (40 million pages) compared to the full Web, but keep in mind that using a significantly larger dataset would have made many of our experiments prohibitively expensive. As we will see, each of the graphs we present study multiple configurations, and for each configuration, multiple crawler runs were made to 130 2 4 8 16 32 64 0.2 0.4 0.6 0.8 1 Number of C-proc 's Coverage n Figure 4: Number of processes vs. Coverage 64 4096 10000 20000 30000 0.2 0.4 0.6 0.8 1 64 32 8 2 processes processes processes processes Coverage Number of Seeds s Figure 5: Number of seed URLs vs. Coverage obtain statistically valid data points. Each run involves simulating how one or more C-proc's would visit the 40 million pages. Such detailed simulations are inherently very time consuming. It is clearly difficult to predict what would happen for a larger dataset. In the extended version of this paper [8], we examine this data size issue a bit more carefully and discuss whether a larger dataset would have changed our conclusions. FIREWALL MODE AND COVERAGE A firewall mode crawler (Section 3, Item 1) has minimal communication overhead, but it may have coverage and quality problems (Section 4). In this section, we quantitatively study the effectiveness of a firewall mode crawler using the 40 million pages in our repository. In particular, we estimate the coverage (Section 4, Item 2) of a firewall mode crawler when it employs n C-proc's in parallel. (We discuss the quality issue of a parallel crawler later.) In our experiments, we considered the 40 million pages within our WebBase repository as the entire Web, and we used site-hash based partitioning (Section 3.2, Item 2). As the seed URLs, each C-proc was given 5 random URLs from its own partition, so 5n seed URLs were used in total by the overall crawler. (We discuss the effect of the number of seed URLs shortly.) Since the crawler ran in firewall mode, C-proc's followed only intra-partition links, not inter-partition links. Under these settings, we let the C-proc's run until they ran out of URLs. After this simulated crawling, we measured the overall coverage of the crawler. We performed these experiments with 5n random seed URLS and repeated the experiments multiple times with different seed URLs. In all of the runs, the results were essentially the same. In Figure 4, we summarize the results from the experiments . The horizontal axis represents n, the number of parallel C-proc's, and the vertical axis shows the coverage of the overall crawler for the given experiment. Note that the coverage is only 0.9 even when n = 1 (a single-process). This result is because the crawler in our experiment started with only 5 URLs, while the actual dataset was collected with 1 million seed URLs. Thus, some of the 40 million pages were unreachable from the 5 seed URLs. From the figure it is clear that the coverage decreases as the number of processes increases. This trend is because the number of inter-partition links increases as the Web is split into smaller partitions, and thus more pages are reachable only through inter-partition links. From this result we can see that we may run a crawler in a firewall mode without much decrease in coverage with fewer than 4 C-proc's. For example, for the 4-process case, the coverage decreases only 10% from the single-process case. At the same time, we can also see that the firewall mode crawler becomes quite ineffective with a large number of C-proc's . Less than 10% of the Web can be downloaded when 64 C-proc's run together, each starting with 5 seed URLs. Clearly, coverage may depend on the number of seed URLs that each C-proc starts with. To study this issue, we also ran experiments varying the number of seed URLs, s, and we show the results in Figure 5. The horizontal axis in the graph represents s, the total number of seed URLs that the overall crawler used, and the vertical axis shows the coverage for that experiment. For example, when s = 128, the overall crawler used 128 total seed URLs, each C-proc starting with 2 seed URLs when 64 C-proc's ran in parallel. We performed the experiments for 2, 8, 32, 64 C-proc cases and plotted their coverage values. From this figure, we can observe the following trends: When a large number of C-proc's run in parallel (e.g., 32 or 64), the total number of seed URLs affects the coverage very significantly. For example, when 64 processes run in parallel the coverage value jumps from 0.4% to 10% if the number of seed URLs increases from 64 to 1024. When only a small number of processes run in parallel (e.g., 2 or 8), coverage is not significantly affected by the number of seed URLs. While coverage increases slightly as s increases, the improvement is marginal. Based on these results, we draw the following conclusions: 1. When a relatively small number of C-proc's are running in parallel, a crawler using the firewall mode provides good coverage. In this case, the crawler may start with only a small number of seed URLs, because coverage is not much affected by the number of seed URLs. 2. The firewall mode is not a good choice if the crawler wants to download every single page on the Web. The crawler may miss some portion of the Web, particularly when it runs many C-proc's in parallel. Example 1. (Generic search engine) To illustrate how our results could guide the design of a parallel crawler, consider the following example. Assume that to operate a Web search engine, we need to download 1 billion pages 6 in one 6 Currently the Web is estimated to have around 1 billion pages. 131 0.2 0.4 0.6 0.8 1 0.5 1 1.5 2 2.5 3 overlap coverage n=64 n=32 : number of n 's C-proc n=4 n=8 n=2 n=16 Figure 6: Coverage vs. Overlap for a cross-over mode crawler month. Each machine that we plan to run our C-proc's on has 10 Mbps link to the Internet, and we can use as many machines as we want. Given that the average size of a web page is around 10K bytes, we roughly need to download 10 4 10 9 = 10 13 bytes in one month. This download rate corresponds to 34 Mbps, and we need 4 machines (thus 4 C-proc's) to obtain the rate. Given the results of our experiment (Figure 4), we may estimate that the coverage will be at least 0.8 with 4 C-proc's. Therefore, in this scenario, the firewall mode may be good enough, unless it is very important to download the "entire" Web. Example 2. (High freshness) As a second example, let us now assume that we have strong "freshness" requirement on the 1 billion pages and need to revisit every page once every week, not once every month. This new scenario requires approximately 140 Mbps for page download, and we need to run 14 C-proc's. In this case, the coverage of the overall crawler decreases to less than 0.5 according to Figure 4. Of course, the coverage could be larger than our conservative estimate, but to be safe one would probably want to consider using a crawler mode different than the firewall mode. CROSS-OVER MODE AND OVERLAP In this section, we study the effectiveness of a cross-over mode crawler (Section 3, Item 2). A cross-over crawler may yield improved coverage of the Web, since it follows inter-partition links when a C-proc runs out of URLs in its own partition. However, this mode incurs overlap in downloaded pages (Section 4, Item 1), because a page can be downloaded by multiple C-proc's. Therefore, the crawler increases its coverage at the expense of overlap in the downloaded pages. In Figure 6, we show the relationship between the coverage and the overlap of a cross-over mode crawler obtained from the following experiments. We partitioned the 40M pages using site-hash partitioning and assigned them to n C-proc's. Each of the n C-proc's then was given 5 random seed URLs from its partition and followed links in the cross-over mode. During this experiment, we measured how much overlap the overall crawler incurred when its coverage reached various points. The horizontal axis in the graph shows the coverage at a particular time and the vertical axis shows the overlap at the given coverage. We performed the experiments for n = 2, 4, . . . , 64. Note that in most cases the overlap stays at zero until the coverage becomes relatively large. For example, when n = 2 4 8 16 32 64 0.5 1 1.5 2 2.5 3 URL Hash Site Hash Communication overhead n Number of C-proc's Figure 7: Number of crawling processes vs. Number of URLs exchanged per page 16, the overlap is zero until coverage reaches 0.5. We can understand this result by looking at the graph in Figure 4. According to that graph, a crawler with 16 C-proc's can cover around 50% of the Web by following only intra-partition links. Therefore, even a cross-over mode crawler will follow only intra-partition links until its coverage reaches that point. Only after that, each C-proc starts to follow inter-partition links, thus increasing the overlap. For this reason, we believe that the overlap would have been much worse in the beginning of the crawl, if we adopted the independent model (Section 2). By applying the partitioning scheme to C-proc's, we make each C-proc stay in its own partition in the beginning and suppress the overlap as long as possible. While the crawler in the cross-over mode is much better than one based on the independent model, it is clear that the cross-over crawler still incurs quite significant overlap. For example, when 4 C-proc's run in parallel in the cross-over mode, the overlap becomes almost 2.5 to obtain coverage close to 1. For this reason, we do not recommend the crossover mode, unless it is absolutely necessary to download every page without any communication between C-proc's. EXCHANGE MODE AND COMMUNICATION To avoid the overlap and coverage problems, an exchange mode crawler (Section 3, Item 3) constantly exchanges inter-partition URLs between C-proc's. In this section, we study the communication overhead (Section 4, Item 4) of an exchange mode crawler and how much we can reduce it by replicating the most popular k URLs. For now, let us assume that a C-proc immediately transfers inter-partition URLs. (We will discuss batch communication later when we discuss the quality of a parallel crawler.) In the experiments, again, we split the 40 million pages into n partitions based on site-hash values and ran n C-proc's in the exchange mode. At the end of the crawl, we measured how many URLs had been exchanged during the crawl. We show the results in Figure 7. In the figure, the horizontal axis represents the number of parallel C-proc's, n, and the vertical axis shows the communication overhead (the average number of URLs transferred per page). For comparison purposes, the figure also shows the overhead for a URL-hash based scheme, although the curve is clipped at the top because of its large overhead values. To explain the graph, we first note that an average page has 10 out-links, and about 9 of them point to pages in 132 the same site. Therefore, the 9 links are internally followed by a C-proc under site-hash partitioning. Only the remaining 1 link points to a page in a different site and may be exchanged between processes. Figure 7 indicates that this URL exchange increases with the number of processes. For example, the C-proc's exchanged 0.4 URLs per page when 2 processes ran, while they exchanged 0.8 URLs per page when 16 processes ran. Based on the graph, we draw the following conclusions: The site-hash based partitioning scheme significantly reduces communication overhead, compared to the URL-hash based scheme. We need to transfer only up to one link per page (or 10% of the links), which is significantly smaller than the URL-hash based scheme. For example, when we ran 2 C-proc's using the URL-hash based scheme the crawler exchanged 5 links per page under the URL-hash based scheme, which was significantly larger than 0.5 links per page under the site-hash based scheme. The network bandwidth used for the URL exchange is relatively small, compared to the actual page download bandwidth. Under the site-hash based scheme, at most 1 URL will be exchanged per page, which is about 40 bytes. 7 Given that the average size of a Web page is 10 KB, the URL exchange consumes less than 40/10K = 0.4% of the total network bandwidth. However, the overhead of the URL exchange on the overall system can be quite significant. The processes need to exchange up to one message per page, and the message has to go through the TCP/IP network stack at the sender and the receiver. Thus it is copied to and from kernel space twice, incurring two context switches between the kernel and the user mode. Since these operations pose significant overhead even if the message size is small, the overall overhead can be important if the processes exchange one message per every downloaded page. In the extended version of this paper [8], we also study how much we can reduce this overhead by replication. In short, our results indicate that we can get significant reduction in communication cost (between 40% 50% reduction) when we replication the most popular 10,000 100,000 URLs in each C-proc. When we replicated more URLs, the cost reduction was not as dramatic as the first 100,000 URLs. Thus, we recommend replicating 10,000 100,000 URLs. QUALITY AND BATCH COMMUNICATION As we discussed, the quality (Section 4, Item 3) of a parallel crawler can be worse than that of a single-process crawler, because each C-proc may make crawling decisions solely based on the information collected within its own partition . We now study this quality issue. In the discussion we also study the impact of the batch communication technique (Section 3.1, Item 1) on quality. Throughout the experiments in this section, we assume that the crawler uses the number of backlinks to page p as the importance of p, or I(p). That is, if 1000 pages on the 7 In our estimation, an average URL was about 40 bytes long. 0 1 2 4 10 20 50 100 500 1000 0.025 0.05 0.075 0.1 0.125 0.15 0.175 0.2 Quality Number of URL exchanges 2 8 64 Processes x (a) URL exchange vs. Quality 0 1 2 4 10 20 50 0.2 0.4 0.6 0.8 1.0 Number of URL exchanges 2 8 64 Processes x Communication overhead (b) URL exchange vs. Communication Figure 8: Crawlers downloaded 500K pages (1.2% of 40M) Web have links to page p, the importance of p is I(p) = 1000. Clearly, there exist many other ways to define the importance of a page, but we use this metric because it (or its variations) is being used by some existing search engines [22, 13]. Also, note that this metric depends on the global structure of the Web. If we use an importance metric that solely depends on a page itself, not on the global structure of the Web, the quality of a parallel crawler will be essentially the same as that of a single crawler, because each C-proc in a parallel crawler can make good decisions based on the pages that it has downloaded. Under the backlink metric, each C-proc in our experiments counted how many backlinks a page has from the downloaded pages and visited the page with the most backlinks first. Remember that the C-proc's need to periodically exchange messages to inform others of the inter-partition backlinks . Depending on how often they exchange messages, the quality of the downloaded pages will differ. For example, if C-proc's never exchange messages, the quality will be the same as that of a firewall mode crawler, and if they exchange messages after every downloaded page, the quality will be similar to that of a single-process crawler. To study these issues, we compared the quality of the downloaded pages when C-proc's exchanged backlink messages at various intervals and we show the results in Figures 8(a), 9(a) and 10(a). Each graph shows the quality achieved by the overall crawler when it downloaded a total of 500K, 2M, and 8M pages, respectively. The horizontal axis in the graphs represents the total number of URL exchanges during a crawl, x, and the vertical axis shows the quality for the given experiment. For example, when x = 1, the C-proc's exchanged backlink count information only once in the middle of the crawl. Therefore, the case when x = 0 represents the quality of a firewall mode crawler, and the case 133 0 1 2 4 10 20 50 100 500 1000 0.05 0.1 0.15 0.2 0.25 Number of URL exchanges Quality 2 8 64 Processes x (a) URL exchange vs. Quality 0 1 2 4 10 20 50 100 5001000 0.25 0.5 0.75 1 2 8 64 Communication overhead x Number of URL exchanges Processes (b) URL exchange vs. Communication Figure 9: Crawlers downloaded 2M pages (5% of 40M) when x shows the quality of a single-process crawler. In Figures 8(b), 9(b) and 10(b), we also show the communication overhead (Section 4, Item 4); that is, the average number of [URL, backlink count] pairs exchanged per a downloaded page. From these figures, we can observe the following trends: As the number of crawling processes increases, the quality of downloaded pages becomes worse, unless they exchange backlink messages often. For example, in Figure 8(a), the quality achieved by a 2-process crawler (0.12) is significantly higher than that of a 64-process crawler (0.025) in the firewall mode (x = 0). Again, this result is because each C-proc learns less about the global backlink counts when the Web is split into smaller parts. The quality of the firewall mode crawler (x = 0 ) is significantly worse than that of the single-process crawler (x ) when the crawler downloads a relatively small fraction of the pages (Figures 8(a) and 9(a)). However , the difference is not very significant when the crawler downloads a relatively large fraction (Figure 10(a)). In other experiments, when the crawler downloaded more than 50% of the pages, the difference was almost negligible in any case. (Due to space limitations , we do not show the graphs.) Intuitively, this result makes sense because quality is an important issue only when the crawler downloads a small portion of the Web. (If the crawler will visit all pages anyway, quality is not relevant.) The communication overhead does not increase lin-early as the number of URL exchange increases. The graphs in Figures 8(b), 9(b) and 10(b) are not straight lines. This result is because a popular URL appears 0 1 2 4 10 20 50 100 500 1000 0.1 0.2 0.3 0.4 0.5 0.6 Number of URL exchanges Quality 2 8 64 Processes x (a) URL exchange vs. Quality 0 1 2 4 10 20 50 100 500 1000 0.5 0.4 0.3 0.2 0.1 2 8 64 Communication overhead Processes x Number of URL exchanges (b) URL exchange vs. Communication Figure 10: Crawlers downloaded 8M pages (20% of 40M) multiple times between backlink exchanges. Therefore , a popular URL can be transferred as one entry (URL and its backlink count) in the exchange, even if it appeared multiple times. This reduction increases as C-proc's exchange backlink messages less frequently. One does not need a large number of URL exchanges to achieve high quality. Through multiple experiments, we tried to identify how often C-proc's should exchange backlink messages to achieve the highest quality value. From these experiments, we found that a parallel crawler can get the highest quality values even if the processes communicate less than 100 times during a crawl. We use the following example to illustrate how one can use the results of our experiments. Example 3. (Medium-Scale Search Engine) Say we plan to operate a medium-scale search engine, and we want to maintain about 20% of the Web (200 M pages) in our index. Our plan is to refresh the index once a month. The machines that we can use have individual T1 links (1.5 Mbps) to the Internet. In order to update the index once a month, we need about 6.2 Mbps download bandwidth, so we have to run at least 5 C-proc's on 5 machines. According to Figure 10(a) (20% download case), we can achieve the highest quality if the C-proc's exchange backlink messages 10 times during a crawl when 8 processes run in parallel. (We use the 8 process case because it is the closest number to 5.) Also, from Figure 10(b), we can see that when C-proc's exchange messages 10 times during a crawl they need to exchange fewer than 0 .17 200M = 34M [URL, backlink count] pairs in total . Therefore, the total network bandwidth used by the back-134 link exchange is only (34M 40 )/(200M 10K ) 0 .06 % of the bandwidth used by actual page downloads. Also, since the exchange happens only 10 times during a crawl, the context-switch overhead for message transfers (discussed in Section 8) is minimal. Note that in this scenario we need to exchange 10 backlink messages in one month or one message every three days. Therefore, even if the connection between C-proc's is unreliable or sporadic, we can still use the exchange mode without any problem. CONCLUSION Crawlers are being used more and more often to collect Web data for search engine, caches, and data mining. As the size of the Web grows, it becomes increasingly important to use parallel crawlers. Unfortunately, almost nothing is known (at least in the open literature) about options for parallelizing crawlers and their performance. Our paper addresses this shortcoming by presenting several architectures and strategies for parallel crawlers, and by studying their performance. We believe that our paper offers some useful guidelines for crawler designers, helping them, for example, select the right number of crawling processes, or select the proper inter-process coordination scheme. In summary, the main conclusions of our study were the following: When a small number of crawling processes run in parallel (in our experiment, 4 or fewer), the firewall mode provides good coverage. Given that firewall mode crawlers can run totally independently and are easy to implement, we believe that it is a good option to consider. The cases when the firewall mode might not be appropriate are: 1. when we need to run more than 4 crawling processes or 2. when we download only a small subset of the Web and the quality of the downloaded pages is important . A crawler based on the exchange mode consumes small network bandwidth for URL exchanges (less than 1% of the network bandwidth). It can also minimize other overheads by adopting the batch communication technique . In our experiments, the crawler could maximize the quality of the downloaded pages, even if it exchanged backlink messages fewer than 100 times during a crawl. By replicating between 10,000 and 100,000 popular URLs, we can reduce the communication overhead by roughly 40%. Replicating more URLs does not significantly reduce the overhead. REFERENCES [1] T. E. Anderson, M. D. Dahlin, J. M. Neefe, D. A. Patterson, D. S. Roselli, and R. Y. Wang. Serverless network file systems. In Proc. of SOSP, 1995. [2] A. Barabasi and R. Albert. Emergence of scaling in random networks. Science, 286(509), 1999. [3] A. Z. Broder, S. R. Kumar, F. Maghoul, P. Raghavan, S. Rajagopalan, R. Stata, A. Tomkins, and J. L. Wiener. Graph structure in the web. In Proc. of WWW Conf., 2000. [4] M. Burner. Crawling towards eterneity: Building an archive of the world wide web. Web Techniques Magazine, 2(5), May 1998. [5] S. Chakrabarti, M. van den Berg, and B. Dom. Focused crawling: A new approach to topic-specific web resource discovery. In Proc. of WWW Conf., 1999. [6] J. Cho and H. Garcia-Molina. The evolution of the web and implications for an incremental crawler. In Proc. of VLDB Conf., 2000. [7] J. Cho and H. Garcia-Molina. Synchronizing a database to improve freshness. In Proc. of SIGMOD Conf., 2000. [8] J. Cho and H. Garcia-Molina. Parallel crawlers. Technical report, UCLA Computer Science, 2002. [9] J. Cho, H. Garcia-Molina, and L. Page. Efficient crawling through URL ordering. In Proc. of WWW Conf., 1998. [10] E. Coffman, Jr., Z. Liu, and R. R. Weber. Optimal robot scheduling for web search engines. Technical report, INRIA, 1997. [11] M. Diligenti, F. M. Coetzee, S. Lawrence, C. L. Giles, and M. Gori. Focused crawling using context graphs. In Proc. of VLDB Conf., 2000. [12] D. Eichmann. The RBSE spider: Balancing effective search against web load. In Proc. of WWW Conf., 1994. [13] Google Inc. http://www.google.com. [14] A. Heydon and M. Najork. Mercator: A scalable, extensible web crawler. Word Wide Web, 2(4):219229, December 1999. [15] A. Heydon and M. Najork. High-performance web crawling. Technical report, SRC Research Report, 173, Compaq Systems Research Center, September 2001. [16] D. Hirschberg. Parallel algorithms for the transitive closure and the connected component problem. In Proc. of STOC Conf., 1976. [17] M. Koster. Robots in the web: threat or treat? ConneXions, 4(4), April 1995. [18] O. A. McBryan. GENVL and WWWW: Tools for taming the web. In Proc. of WWW Conf., 1994. [19] R. C. Miller and K. Bharat. SPHINX: a framework for creating personal, site-specific web crawlers. In Proc. of WWW Conf., 1998. [20] D. Nassimi and S. Sahni. Parallel permutation and sorting algorithms and a new generalized connection network. Journal of ACM, 29:642667, July 1982. [21] M. T. Ozsu and P. Valduriez. Principles of Distributed Database Systems. Prentice Hall, 1999. [22] L. Page and S. Brin. The anatomy of a large-scale hypertextual web search engine. In Proc. of WWW Conf., 1998. [23] B. Pinkerton. Finding what people want: Experiences with the web crawler. In Proc. of WWW Conf., 1994. [24] Robots exclusion protocol. http://info.webcrawler. com/mak/projects/robots/exclusion.html. [25] A. S. Tanenbaum and R. V. Renesse. Distributed operating systems. ACM Computing Surveys, 17(4), December 1985. [26] G. K. Zipf. Human Behaviour and the Principle of Least Effort: an Introduction to Human Ecology. Addison-Wesley, 1949. 135

guideline;architecture;Parallelization;Web Spider;parallel crawler;Web Crawler;model evaluation

147

Performance Enhancing Proxy for Interactive 3G Network Gaming

Unlike non-time-critical applications like email and file transfer , network games demand timely data delivery to maintain the seemingly interactive presence of players in the virtual game world. Yet the inherently large transmission delay mean and variance of 3G cellular links make on-time game data delivery difficult. Further complicating the timely game data delivery problem is the frequent packet drops at these links due to inter-symbol interference, fading and shadowing at the physical layer. In this paper, we propose a proxy architecture that enhances the timeliness and reliability of data delivery of interactive games over 3G wireless networks. In particular, a performance enhancing proxy is designed to optimize a new time-critical data type -- variable-deadline data, where the utility of a datum is inversely proportional to the time required to deliver it. We show how a carefully designed and configured proxy can noticeably improve the delivery of network game data.

INTRODUCTION While network gaming has long been projected to be an application of massive economic growth, as seen in the recent explosive development on the wired Internet in South Korea and Japan, deployment of similar network games on 3G wireless networks continues to be slow and difficult. One reason is that unlike their wired counterparts, wireless links are notoriously prone to errors due to channel fading, shadowing and inter-symbol interference. While 3G wireless networks, such as High Speed Downlink Packet Access (HSDPA) of 3rd Generation Partnership Project (3GPP) Release 5 (R5) [1] and CDMA 1x EvDO of 3GPP2 [5], combat wireless link failures at the MAC and physical layer with an elaborate system of channel coding, retransmission, modulation and spreading, with resulting packet loss rate being reduced to negligible 1 to 2%, the detrimental side-effect to network gaming is the large and often unpredictable transmission delay mean and variance [15]. Such large and variable delays greatly reduce the necessary interactivity of network game players and deteriorate the overall gaming experience. In a separate development, a new 3G network element called IP Multimedia Subsystem (IMS) [3] has been intro-duced in 3GPP specifications R5 and later, as shown in Figure 1. The Session Initiation Protocol (SIP)-based IMS provides a multitude of multimedia services: from establishing connections from the legacy telephone networks to the new IP core network using Voice over IP (VoIP), to delivering streaming services such as video as a value-added service to mobile users (UE). Strategically located as a pseudo-gateway to the private and heavily provisioned 3G networks, it is foreseeable that IMS will continue to enlarge and enrich its set of multimedia services in future wireless networks. In this paper, we propose a performance enhancing proxy (PEP) called (W)ireless (I)nteractive (N)etwork (G)aming Proxy (WING) to improve the timely delivery of network game data in 3G wireless networks. WING is located inside IMS as an application service on top of the myriad of 207 services that IMS already provides. In a nutshell, WING improves the delivery of game data from the game server to 3G wireless game players 1 using the following three techniques. First, by virtue of locating at the intersection of the private wireless network and the open Internet, connection from the game server to the wireless game player can be strategi-cally split; for the server-WING connection, only the statis-tically stable and fast round-trip time (RTT) and low wired-network -only packet loss rate (PLR) are used for congestion control, resulting in a steady yet TCP-friendly server-WING connection. Second, by configuring parameters in the radio link layer (RLC) specifically for gaming during session setup, excessive RLC retransmissions are avoided, and timeliness of game data is improved at the controlled expense of in-creased packet losses. Finally, by constructing small but error-resilient packets that contain location data, packets can be transmitted in fewer MAC-layer protocol data units (PDU) and hence further reduces delay. The paper is organized as follows. Related work is presented in Section 2. We overview the 3G wireless system in focus, HSDPA of 3GPP R5, in Section 3. Note that because similar link and MAC layer transport optimizations that chiefly affect delay mean and variance are also employed in other 3G networks, our proposed WING can conceivably be applied to other wireless networks such as CDMA 1x EvDO of 3GPP2. We discuss the design of WING in details in Section 4. Finally, experimental results and conclusion are provided in Section 5 and 6, respectively. RELATED WORK We divide the discussion on the large volume of related work into two section. Section 2.1 discusses related research on wireless transport optimization. Section 2.2 discusses related research in transport of network game data. 2.1 Wireless Transport Optimization We note that proxy-based transport optimization for last-hop wireless networks has a long history, with the majority of the research [4, 15] focusing on optimization of TCP over last-hop wireless networks. In particular, [15] showed that while 3G network packet losses can indeed be successfully overcome by using ample link layer retransmissions, the resulting large RTT mean and variance may severely affect the performance of a TCP-like congestion avoidance rate control that is based on end-to-end observable statistics of RTT and PLR. The limiting rate constraint and undesirable fluctuations can be alleviated using a proxy with split-connection -- a theme we develop in Section 4.2. Recently, efforts on proxy design have shifted to delay-sensitive multimedia transport [18, 13, 8, 9], though all of them focused exclusively on streaming media, while we focus on network gaming. Note that due to cited complexity reason, a competing end-to-end approach for rate control that does not rely on an intermediate proxy is popular as well [17, 6]. However, we chose the proxy-based approach and will juxtapose its advantages in Section 4.2. 1 While peer-to-peer model for interactive network games is also possible, we assume the more common server-client model where the game server maintains and disseminates all game states in this paper. 2.2 Transport of Network Game Data In [3], a general gaming platform for IMS that provides network services needed for network gaming such as session setup and registration is proposed to ease deployment over 3G networks. Our work is orthogonal to [3] since we focus only on the efficient transport of game data. An early work on gaming protocol is [10], which defined a Game Transport Protocol (GTP) for massive multi-player on-line games (MMPOGs) over the Internet. Our proposed gaming proxy WING differs in the following respects: i) we design WING specifically for lossy, bandwidth-limited networks , hence focusing on design of network-optimized differential coding to produce small but loss-resilient packets; and, ii) we tailor WING for HSDPA of 3G wireless networks, optimizing performance by intelligently configuring parameters of the RLC layer. The most similar related work is [12], which proposed an end-to-end adaptive FEC and dynamic packetization algorithm to combat packet losses due to wireless link failures and reduce packet sizes. Unlike [12], our approach is proxy-based , and we tailor our gaming optimization exclusively for 3G networks. OVERVIEW OF UMTS RELEASE 5 HSDPA of UMTS Release 5, also known as 3.5G, improves upon Release 4 with numerous lower-layer optimizations. First, a shared channel is periodically scheduled to users in the cell with good observable network conditions to take advantage of user diversity during fading without sacrificing fairness. Second, an elaborate MAC-layer scheme chooses an appropriate combination of FEC, hybrid ARQ, modulation and spreading based on client observable network state. In this section, we instead focus on the RLC layer, where the user has limited control over behavior using configuration of parameters during session setup. The Radio Link control (RLC) layer [1] buffers upper layer service data units (SDU) on a per-session basis -- IP packets in this case, and segments each SDU into smaller protocol data units (PDU) of size S P DU and await transmission at lower layers. There are three transmission modes: transparent mode (TM), unacknowledged mode (UM) and acknowledged mode (AM). Only AM performs link-layer retransmissions if transmission in the lower layer fails. For error resiliency, we focus only on AM. In particular, we look at how SDUs are discarded in the RLC layer: using a method of retransmission-based discard (RBD), an SDU can be discarded before successful transmission. In a nutshell, an SDU is discarded if a predefined maximum number of retransmissions B has been reached before successful transmission of a PDU belonging to the SDU. We will investigate how the value B can be selected to trade off error resiliency with delay in Section 4.3. WING FOR WIRELESS INTERACTIVE NETWORK GAMING Before we discuss the three optimizations of our proposed gaming proxy WING in details in Section 4.2, 4.3 and 4.4, we first define a new type of transport data called variable deadline data in Section 4.1 -- a consequence of a prediction procedure used at a network game client to predict locations of other game players in the virtual game world. 208 0 100 200 300 400 500 1 2 3 4 5 6 delay in ms distortion distortion vs. delay for dead-reckoning random walk weighted random walk 0 100 200 300 400 500 0.2 0.4 0.6 0.8 1 delay in ms utility utility vs. delay for dead-reckoning random walk weighted random walk a) distortion vs. delay b) utility vs. delay Figure 2: Examples of Dead-Reckoning 4.1 Variable Deadline Data Delivery Unlike media streaming applications where a data unit containing media data is fully consumed if it is correctly delivered by a playback deadline and useless otherwise [9], the usefulness (utility) of a game datum is inversely proportional to the time it requires to deliver it. This relationship between utility and transmission delay is the behavioral result of a commonly used game view reconstruction procedure at a game client called dead-reckoning [2]. It works as follows. To maintain time-synchronized virtual world views among game players at time t 0 , a player P A predicts the location t 0 of another player P B and draws it in P A 's virtual world at time t 0 , extrapolating from previously received location updates of P B in the past, , &lt; t 0 . When location update t 0 arrives at P A from P B at a later time t 1 , P A updates its record of P B 's locations with (t 0 , t 0 ), in order to make an accurate prediction of (t 1 , t 1 ) for display in P A 's virtual world at time t 1 . Regardless of what prediction method is used at the client, it is clear that a smaller transmission delay will in general induce a smaller prediction error. We term this type of data with inversely proportional relationship between quantifiable utility and delay variable deadline data. We next show examples of how such utility-delay curve u(d) can be derived in practice given a player movement model and a prediction method. 4.1.1 Examples of Dead-Reckoning We first consider two simple movement models that model a game player in two-dimensional space (x, y). The first is random walk, where for each time increment t, probability mass function (pmf) of random variable of x-coordinate x t , p(x t ), is defined as follows: p(x t+1 = x t + 1) = 1/3 p(x t+1 = x t ) = 1/3 p(x t+1 = x t - 1) = 1/3 (1) Random variable of y-coordinate y t is calculated similarly and is independent of x t . The second movement model is weighted random walk, whose pmf is defined as follows: p (x t+1 = x t + ((x t - x t-1 + 1) mod 2)) = 1/6 p (x t+1 = x t + (x t - x t-1 )) = 2/3 p (x t+1 = x t + ((x t - x t-1 - 1) mod 2)) = 1/6 (2) In words, the player continues the same movement as done in the previous instant with probability 2/3, and changes to one of two other movements each with probability 1/6. Random variable y-coordinate y t is calculated similarly. We defined a simple prediction method called 0th-order prediction as follows: each unknown x t is simply set to the most recently updated x . Using each of the two movement models in combination with the prediction method, we constructed distortion-delay curves experimentally as shown in Figure 2a. As seen, 0th-order prediction is a better match to random walk than weighted random walk, inducing a smaller distortion for all delay values. Utility u(d) -- shown in Figure 2b is simply the reciprocal of distortion. Having derived u(d) gives us a quantifiable metric on which we can objectively evaluate game data transport systems. 4.2 Proxy-based Congestion Control We argue that by locating WING between the open wired Internet and the provisioned wireless networks to conduct split-connection data transfer, stable TCP-friendly congestion control can be maintained on top of UDP in the wired server-WING connection. Traditional congestion control algorithms like TCP-friendly Rate Control (TFRC) [11] space outgoing packets with interval T cc as a function of estimated packet loss rate (PLR) cc , RTT mean m cc and RTT variance 2 cc due to wired network congestion: T cc = m cc p2 cc /3 + 3(m cc + 4 2 cc ) cc `1 + 32 2 cc p3 cc /8 (3) Past end-to-end efforts [17, 6] have focused on methodologies to distinguish wired network congestion losses from wireless link losses, in order to avoid unnecessary rate reduction due to erroneous perception of wireless losses as network congestion. Split connection offers the same effect regarding PLR by completely shielding sender from packet losses due to wireless link failures. Moreover, by performing TFRC (3) in the server-WING connection using only stable wired network statistics, split connection shields the server-WING connection from large rate fluctuations due to large RTT variance in the last-hop 3G link as shown in [15]. For this reason, [15] showed experimentally that indeed proxy-based split-connection congestion control performs better than end-to-end counterparts, even in negligible wireless loss environments. Lastly, we note that split connection can benefit from a rate-mismatch environment [8, 9], where the available bandwidth R 1 in the server-WING connection is larger than the available bandwidth R 2 in the WING-client connection. In such case, the surplus bandwidth R 1 - R 2 can b e used for redundancy packets like forward-error correction (FEC) or retransmission to lower PLR in the server-WING connection . We refer interested readers to [8, 9] for further details. 4.3 Optimizing RLC Configuration Given utility-delay function u(d) in Section 4.1, we optimize configuration of RLC to maximize utility. More pre-cisely , we pick the value of maximum retransmission limit B -- inducing expected SDU loss rate l and delay d , so that the expected utility (1 - l )u(d ) is maximized. We assume a known average SDU size S SDU , PDU loss rate P DU , and probability density function (pdf) of PDU transmission delay () with mean m and variance 2 . First, the expected number of PDUs fragmented from an SDU is N = l S SDU S P DU m . For a given B, the expected SDU 209 loss rate l SDU can b e written simply: P P DU = B X i=1 i-1 P DU (1 P DU ) (4) l SDU = 1 - P N P DU (5) where P P DU is the probability that a PDU is successfully delivered given B. The delay d SDU experienced by a successfully delivered SDU is the sum of queuing delay d q SDU and transmission delay d t SDU . Queuing delay d q SDU is the delay experienced by an SDU while waiting for head-of-queue SDUs to clear due to early termination or delivery success. d t SDU is the expected wireless medium transmission delay given the SDU is successfully delivered. d t SDU is easier and can be calculated as: X P DU = 1 P P DU B X i=1 i i-1 P DU (1 P DU ) (6) d t SDU = N m X P DU (7) where X P DU is the expected number of PDU (re)transmissions given PDU delivery success. To calculate d q SDU , we assume a M/G/1 queue 2 with arrival rate q , mean service time m q , and variance of service time 2 q . Using Pollaczek-Khinchin mean value formula [14], d q SDU can b e written as: d q SDU = q m q `1 + 2 q /m 2 q 2 (1 q m q ) m q (8) In our application, q is the rate at which game data arrive at WING from the server, which we assume to be known. m q is the mean service rate for both cases of SDU delivery success and failure and can be derived as follows: Y SDU = 1 l SDU N X i=1 (B + (i - 1)X P DU ) B P DU P i-1 P DU (9) m q = (1 - l SDU ) d t SDU + l SDU m Y SDU (10) where Y SDU is the expected total number of PDU (re)transmissions in an SDU given SDU delivery failure. Similar analysis will show that the variance of service rate 2 q for our application is: 2 q = (1 - l SDU ) N 2 X 2 P DU 2 + l SDU Y 2 SDU 2 (11) We can now evaluate expected queuing delay d q SDU , from which we evaluate expected delay d SDU . Optimal B is one that maximizes (1 - l SDU ) u(d SDU ). 4.4 Loss-optimized Differential Coding If the location data -- player position updates sent to improve dead-reckoning discussed in Section 4.1 -- are in absolute values, then the size of the packet containing the data can be large, resulting in large delay due to many PDU fragmentation and spreading. The alternative is to describe the location in relative terms -- the difference in the location from a previous time slot. Differential values are smaller, resulting in fewer encoded bits and smaller packets, and hence 2 Our system is actually more similar to a D/G/1 queue, since the arrivals of game data are more likely to be deter-ministic than Markovian. Instead, we use M/G/1 queue as a first-order approximation. 1 2 3 4 ACK boundary 2 3 4 1 Figure 3: Example of Differential Coding mode mode marker ref. size coord. size total 0 00 0 32 2 + 64n 1 01 0 16 2 + 32n 2 10 2 8 4 + 16n 3 11 4 4 6 + 8n Table 1: Differential Coding Modes smaller transmission delay. This differential coding of location data is used today in networked games. The obvious disadvantage of differential coding is that the created dependency chain is vulnerable to network loss; a single loss can result in error propagation until the next absolute location data (refresh). To lessen the error propagation effect while maintaining the coding benefit of differential coding, one can reference a position in an earlier time slot. An example is shown in Figure 3, where we see position 3 ( 3 ) references 1 instead of 2 . This way, loss of packet containing 2 will not affect 3 , which depends only on 1 . The problem is then: for a new position t , how to select reference position t-r for differential coding such that the right tradeoff of error resilience and packet size can be selected? This selection must be done in an on-line manner as new position becomes available from the application to avoid additional delay. 4.4.1 Specifying Coding Modes To implement loss-optimized differential coding, we first define a coding specification that dictates how the receiver should decode location packets. For simplicity, we propose only four coding modes, where each mode is specified by a designated bit sequence (mode marker) in the packet. Assuming the original absolute position is specified by two 32-bit fixed point numbers, mode 0 encodes the unaltered absolute position in x-y order, resulting in data payload size of 2 + 64n bits for n game entities. Mode 1 uses the previous position as reference for differential encoding with 16 bits per coordinate, resulting in 2 + 32n bits for n entities. Mode 2 uses the first 2 bits to specify r in reference position t - r for differential encoding. Each coordinate takes 8 bits, resulting in 4 + 16n total bits for n entities. Mode 3 is similar to mode 2 with the exception that each of the reference marker and the two coordinate takes only 4 bits to encode, resulting in 6 + 8n bits for n entities. For given position t = (x t , y t ) and reference t-r = (x t-r , y t-r ), some modes may be infeasible due to the fixed coding bit budgets for reference and coordinate sizes. So limited to the set of feasible modes, we seek a reference position / mode pair that maximizes an objective function. 210 PLR RTT mean RTT variance Tokyo-Singapore (50) 0 94.125ms 178.46 Tokyo-Singapore (100) 0 95.131ms 385.30 Tokyo-Singapore (200) 0 96.921ms 445.63 HSDPA (50) 0 62.232ms 7956.8 HSDPA (100) 0 72.547ms 25084 HSDPA (200) 0 152.448ms 143390 Table 2: Comparison of Network Statistics 4.4.2 Finding Optimal Coding Modes For an IP packet of size s t containing position t that is sent at time t, we first define the probability that it is correctly delivered b y time as t ( ). t ( ) depends on expected PLR l(s t ) and delay d(s t ), resulting from retransmission limit B chosen in Section 4.3: N(s t ) = s t S P DU i (12) l(s t ) = 1 - (P P DU ) N(s t ) (13) d(s t ) = d q SDU + N(s t ) m X P DU (14) where N(s t ) is the number of PDUs fragmented from an SDU of size s t . l(s t ) is PLR in (5) generalized to SDU size s t . d(s t ) is the expected queuing delay in (8) plus the transmission delay in (7) generalized to SDU size s t . We can now approximate t ( ) as: t ( ) 1 if ACKed b y (1 - l(s t ))1 ( - t - d(s t )) o.w. (15) where 1(x) = 1 if x 0, and = 0 otherwise. If no acknowledgment packets (ACK) are sent from client to WING, then t ( ) is simply the second case in (15). We next define the probability that position t is correctly decoded b y time as P t ( ). Due to dependencies resulting from differential coding, P t ( ) is written as follows: P t ( ) = t ( ) Y jt j ( ) (16) where j i is the set of positions j's that precedes t in the dependency graph due to differential coding. Given utility function u(d) in Section 4.1 and decode probability (16), the optimal reference position / mode pair is one that maximizes the following objective function: max P t (t + d(s t )) u(d(s t )) (17) EXPERIMENTATION We first present network statistics for HSDPA and discuss the implications. We collected network statistics of 10,000 ping packets, of packet size 50, 100 and 200 bytes, spaced 200ms apart, between hosts in Tokyo and Singapore inside HP intranet. The results are shown in Table 2. We then conducted the same experiment over a network emu-lator called WiNe2 [16] emulating the HSDPA link with 10 competing ftp users each with mobility model Pedestrian A. We make two observations in Table 2. One, though results from both experiments had similar RTT means, HSDPA's RTT variances were very large, substantiating our claim that using split-connection to shield the server-WING connection from HSDPA's RTT variance would drastically improve TFRC bandwidth (3) of server-WING connection. number of entities n 4 frame rate 10 fps IP + UDP header 20 + 8 bytes RLC PDU size 40 bytes RLC PDU loss rate 0.1 to 0.3 average packet size 61 bytes shifted Gamma parameter g 2 shifted Gamma parameter g 0.1 shifted Gamma parameter g 10.0 Table 3: Simulation Parameters 1 2 3 4 5 6 7 8 9 10 50 100 150 200 250 300 350 400 retransmission limit expected delay in ms expected delay vs. retrans. limit epsilon=0.1 epsilon=0.2 epsilon=0.3 1 2 3 4 5 6 7 8 9 10 0.2 0.22 0.24 0.26 0.28 0.3 0.32 0.34 0.36 0.38 retransmission limit expected utility expected utility vs. retrans. limit epsilon=0.1 epsilon=0.2 epsilon=0.3 a) delay vs. retrans. limit B b) utility vs. retrans. limit B Figure 4: Delay andUtility vs. Retrans. Limit Two, larger packets entailed larger RTT means for HSDPA. This means that the differential coding discussed in Section 4.4 indeed has substantial performance improvement potential . We next used an internally developed network simulator called (mu)lti-path (n)etwork (s)imulator (muns) that was used in other simulations [7] to test RLC configurations and differential coding. For PDU transmission delay (), we used a shifted Gamma distribution: () = g g ( g ) g -1 e g (g ) ( g ) , g &lt; &lt; (18) where () is the Gamma function [14]. The parameters used are shown in Table 3. Figure 4 shows the expected delay and utility as a function of retransmission limit B for different PDU loss rates. As expected, when B increases, the expected delay increases. The expected utility, on the other hand, reaches a peak and decreases. For given PDU loss rate, we simply select B with the largest expected utility. Next, we compare the results of our loss-optimized differential coding optimization opt in Section 4.4 with two schemes: abs, which always encodes in absolute values; and, rel, which uses only previous frame for differential coding and refreshes with absolute values every 10 updates. abs P DU B abs(1) abs( B ) rel(1) rel( B ) opt 0.10 2 1.181 1.134 1.806 1.154 1.070 0.15 3 1.222 1.166 2.288 1.108 1.073 0.20 2 1.290 1.192 2.619 1.380 1.086 0.25 2 1.356 1.232 3.035 1.568 1.090 0.30 2 1.449 1.268 3.506 1.750 1.110 0.35 2 1.509 1.300 3.556 2.054 1.121 Table 4: Distortion Comparison 211 represents the most error resilient coding method in differential coding, while rel represents a reasonably coding-efficient method with periodical resynchronization. Note, however, that neither abs nor rel adapts differential coding in real time using client feedbacks. abs and rel were each tested twice. In the first trial, limit B was set to 1, and in the second, B was set to the optimal configured value as discussed in Section 4.3. 20000 data points were generated and averaged for each distortion value in Table 4. As we see in Table 4 for various PDU loss rate P DU , the resulting distortions for opt were always lower than abs's and rel's, particularly for high PDU loss rates. opt performed better than rel because of opt's error resiliency of loss-optimized differential coding, while opt performed better than abs because opt's smaller packets in-duced a smaller queuing delay and a smaller transmission delay due to smaller number of RLC fragmentations. This demonstrates that it is important not only to find an optimal RLC configuration, but a suitable differential coding scheme to match the resulting loss rate and delay of the configuration. CONCLUSION We propose a performance enhancing proxy called WING to improve the delivery of game data from a game server to 3G game players using three techniques: i) split-connection TCP-friendly congestion control, ii) network game optimized RLC configuration, and, iii) packet compression using differential coding. For future, we will investigate how similar techniques can be applied for the 3G uplink from game player to game server. ACKNOWLEDGMENTS The authors thank other members of the multimedia systems architecture team, Yasuhiro Araki and Takeaki Ota, for their valuable comments and discussions. REFERENCES [1] Universal Mobile Telecommunications System (UMTS); Radio Link Control (RLC) protocol specification (3GPP TS.25.322 version 5.12.0 Release 5). http://www.3gpp.org/ftp/Specs/archive/25 series/ 25.322/25322-5c0.zip, September 2005. [2] S. Aggarwal, H. Banavar, and A. Khandelwal. Accuracy in dead-reckoning based distributed multi-player games. In ACM SIGCOMM NetGames, Portland, OR, August 2004. [3] A. Akkawi, S. Schaller, O. Wellnitz, and L. Wolf. A mobile gaming platform for the IMS. In ACM SIGCOMM NetGames, Portland, OR, August 2004. [4] H. Balakrishnan, V. Padmanabhan, S. Seshan, and R. Katz. A comparison of mechanisms for improving TCP performance over wireless links. In IEEE/ACM Trans. Networking, volume 5, no.6, December 1997. [5] Q. Bi and S. Vitebsky. Performance analysis of 3G-1x EvDO high data rate system. In IEEE Wireless Communications and Networking Conference, Orlando, FL, March 2002. [6] M. Chen and A. Zakhor. AIO-TRFC: A light-weight rate control scheme for streaming over wireless. In IEEE WirelessCom, Maui, HI, June 2005. [7] G. Cheung, P. Sharma, and S. J. Lee. Striping delay-sensitive packets over multiple bursty wireless channels. In IEEE International Conference on Multimedia and Expo, Amsterdam, the Netherlands, July 2005. [8] G. Cheung and W. t. Tan. Streaming agent for wired network / wireless link rate-mismatch environment. In International Workshop on Multimedia Signal Processing, St. Thomas, Virgin Islands, December 2002. [9] G. Cheung, W. t. Tan, and T. Yoshimura. Double feedback streaming agent for real-time delivery of media over 3G wireless networks. In IEEE Transactions on Multimedia, volume 6, no.2, pages 304314, April 2004. [10] S. P. et al. Game transport protocol: A reliable lightweight transport protocol for massively multiplayer on-line games (MMPOGs). In SPIE-ITCOM, Boston, MA, July 2002. [11] S. Floyd, M. Handley, J. Padhye, and J. Widmer. Equation-based congestion control for unicast applications. In ACM SIGCOMM, Stockholm, Sweden, August 2000. [12] P. Ghosh, K. Basu, and S. Das. A cross-layer design to improve quality of service in online multiplayer wireless gaming networks. In IEEE Broadnets, Boston, MA, October 2005. [13] L. Huang, U. Horn, F. Hartung, and M. Kampmann. Proxy-based TCP-friendly streaming over mobile networks. In IEEE International Symposium on a World of Wireless, Mobile and Multimedia Networks, Atlanta, GA, September 2002. [14] A. Leon-Garcia. Probability and Random Processes for Electrical Engineering. Addison Wesley, 1994. [15] M. Meyer, J. Sachs, and M. Holzke. Performance evaluation of a TCP proxy in WCDMA networks. In IEEE Wireless Communications, October 2003. [16] Nomor Research GmbH. WiSe2. http://www.nomor.de. [17] F. Yang, Q. Zhang, W. Zhu, and Y.-Q. Zhang. Bit allocation for scalable video streaming over mobile wireless internet. In IEEE Infocom, Hong Kong, March 2004. [18] T. Yoshimura, T. Ohya, T. Kawahara, and M. Etoh. Rate and robustness control with RTP monitoring agent for mobile multimedia streaming. In IEEE International Conference on Communication, New York, NY, April 2002. 212

Wireless Networks;3G wireless network;time critical data;Network Gaming;congestion control;loss-optimized;RLC configuration;proxy architecture

148

Physically-Based Visual Simulation on Graphics Hardware

In this paper, we present a method for real-time visual simulation of diverse dynamic phenomena using programmable graphics hardware. The simulations we implement use an extension of cellular automata known as the coupled map lattice (CML). CML represents the state of a dynamic system as continuous values on a discrete lattice. In our implementation we store the lattice values in a texture, and use pixel-level programming to implement simple next-state computations on lattice nodes and their neighbors. We apply these computations successively to produce interactive visual simulations of convection, reaction-diffusion, and boiling. We have built an interactive framework for building and experimenting with CML simulations running on graphics hardware, and have integrated them into interactive 3D graphics applications.

Introduction Interactive 3D graphics environments, such as games, virtual environments, and training and flight simulators are becoming increasingly visually realistic, in part due to the power of graphics hardware. However, these scenes often lack rich dynamic phenomena, such as fluids, clouds, and smoke, which are common to the real world. A recent approach to the simulation of dynamic phenomena, the coupled map lattice [Kaneko 1993] , uses a set of simple local operations to model complex global behavior. When implemented using computer graphics hardware, coupled map lattices (CML) provide a simple, fast and flexible method for the visual simulation of a wide variety of dynamic systems and phenomena. In this paper we will describe the implementation of CML systems with current graphics hardware, and demonstrate the flexibility and performance of these systems by presenting several fast interactive 2D and 3D visual simulations. Our CML boiling simulation runs at speeds ranging from 8 iterations per second for a 128x128x128 lattice to over 1700 iterations per second for a 64x64 lattice. Section 2 describes CML and other methods for simulating natural phenomena. Section 3 details our implementation of CML simulations on programmable graphics hardware, and Section 4 describes the specific simulations we have implemented. In Section 5 we discuss limitations of current hardware and investigate some solutions. Section 6 concludes. CML and Related Work The standard approach to simulating natural phenomena is to solve equations that describe their global behavior. For example, multiple techniques have been applied to solving the Navier-Stokes fluid equations [Fedkiw, et al. 2001;Foster and Metaxas 1997;Stam 1999] . While their results are typically numerically and visually accurate, many of these simulations require too much computation (or small lattice sizes) to be integrated into interactive graphics applications such as games. CML models, instead of solving for the global behavior of a phenomenon, model the behavior by a number of very simple local operations. When aggregated, these local operations produce a visually accurate approximation to the desired global behavior. Figure 1: 3D coupled map lattice simulations running on graphics hardware. Left: Boiling. Right: Reaction-Diffusion . A coupled map lattice is a mapping of continuous dynamic state values to nodes on a lattice that interact (are `coupled') with a set of other nodes in the lattice according to specified rules. Coupled map lattices were developed by Kaneko for the purpose of studying spatio-temporal dynamics and chaos [Kaneko 1993] . Since their introduction, CML techniques have been used extensively in the fields of physics and mathematics for the simulation of a variety of phenomena, including boiling [Yanagita 1992] , convection [Yanagita and Kaneko 1993] , cloud formation [Yanagita and Kaneko 1997] , chemical reaction-diffusion [Kapral 1993] , and the formation of sand ripples and dunes [Nishimori and Ouchi 1993] . CML techniques were recently introduced to the field of computer graphics for the purpose of cloud modeling and animation [Miyazaki, et al. 2001] . Lattice Boltzmann computation is a similar technique that has been used for simulating fluids, particles, and other classes of phenomena [Qian, et al. 1996] . A CML is an extension of a cellular automaton (CA) [Toffoli and Margolus 1987;von Neumann 1966;Wolfram 1984] in which the discrete state values of CA cells are replaced with continuous real values. Like CA, CML are discrete in space and time and are a versatile technique for modeling a wide variety of phenomena. Methods for animating cloud formation using cellular automata were presented in [Dobashi, et al. 2000;Nagel and Raschke 1992] . Discrete-state automata typically require very large lattices in order to simulate real phenomena, because the discrete states must be filtered in order to compute real values. By using continuous-valued state, a CML is able to represent real physical quantities at each of its nodes. While a CML model can certainly be made both numerically and visually accurate [Kaneko 1993] , our implementation on graphics hardware introduces precision constraints that make numerically accurate simulation difficult. Therefore, our goal is instead to implement visually accurate simulation models on graphics hardware, in the hope that continuing improvement in the speed and precision of graphics hardware will allow numerically accurate simulation in the near future. The systems that have been found to be most amenable to CML implementation are multidimensional initial-value partial differential equations. These are the governing equations for a wide range of phenomena from fluid dynamics to reaction-diffusion. Based on a set of initial conditions, the simulation evolves forward in time. The only requirement is that the equation must first be explicitly discretized in space and time, which is a standard requirement for conventional numerical simulation. This flexibility means that the CML can serve as a model for a wide class of dynamic systems. 2.1 A CML Simulation Example To illustrate CML, we describe the boiling simulation of [Yanagita 1992] . The state of this simulation is the temperature of a liquid. A heat plate warms the lower layer of liquid, and temperature is diffused through the liquid. As the temperature reaches a threshold, the phase changes and "bubbles" of high temperature form. When phase changes occur, newly formed bubbles absorb latent heat from the liquid around them, and temperature differences cause them to float upward under buoyant force. Yanagita implements this global behavior using four local CML operations; Diffusion, Phase change, Buoyancy, and Latent heat. Each of these operations can be written as a simple equation. Figures 1, 2 and 7 (see color pate) show this simulation running on graphics hardware, and Section 4.1 gives details of our implementation. We will use this simulation as an example throughout this paper. Hardware Implementation Graphics hardware is an efficient processor of images it can use texture images as input, and outputs images via rendering. Images arrays of values map well to state values on a lattice. Two-dimensional lattices can be represented by 2D textures, and 3D lattices by 3D textures or collections of 2D textures. This natural correspondence, as well as the programmability and performance of graphics hardware, motivated our research. 3.1 Why Graphics Hardware? Our primary reason to use graphics hardware is its speed at imaging operations compared to a conventional CPU. The CML models we have implemented are very fast, making them well suited to interactive applications (See Section 4.1). GPUs were designed as efficient coprocessors for rendering and shading. The programmability now available in GPUs such as the NVIDIA GeForce 3 and 4 and the ATI Radeon 8500 makes them useful coprocessors for more diverse applications. Since the time between new generations of GPUs is currently much less than for CPUs, faster coprocessors are available more often than faster central processors. GPU performance tracks rapid improvements in semiconductor technology more closely than CPU performance. This is because CPUs are designed for high performance on sequential operations, while GPUs are optimized for the high parallelism of vertex and fragment processing [Lindholm, et al. 2001] . Additional transistors can Figure 2: A sequence of stills (10 iterations apart) from a 2D boiling simulation running on graphics hardware. 110 Harris, Coombe, Scheuermann, and Lastra / Simulation on Graphics Hardware The Eurographics Association 2002. therefore be used to greater effect in GPU architectures. In addition, programmable GPUs are inexpensive, readily available, easily upgradeable, and compatible with multiple operating systems and hardware architectures. More importantly, interactive computer graphics applications have many components vying for processing time. Often it is difficult to efficiently perform simulation, rendering, and other computational tasks simultaneously without a drop in performance. Since our intent is visual simulation, rendering is an essential part of any solution. By moving simulation onto the GPU that renders the results of a simulation, we not only reduce computational load on the main CPU, but also avoid the substantial bus traffic required to transmit the results of a CPU simulation to the GPU for rendering. In this way, methods of dynamic simulation on the GPU provide an additional tool for load balancing in complex interactive applications. Graphics hardware also has disadvantages. The main problems we have encountered are the difficulty of programming the GPU and the lack of high precision fragment operations and storage. These problems are related programming difficulty is increased by the effort required to ensure that precision is conserved wherever possible. These issues should disappear with time. Higher-level shading languages have been introduced that make hardware graphics programming easier [Peercy, et al. 2000;Proudfoot, et al. 2001] . The same or similar languages will be usable for programming simulations on graphics hardware. We believe that the precision of graphics hardware will continue to increase, and with it the full power of programmability will be realised. 3.2 General-Purpose Computation The use of computer graphics hardware for general-purpose computation has been an area of active research for many years, beginning on machines like the Ikonas [England 1978] , the Pixel Machine [Potmesil and Hoffert 1989] and Pixel-Planes 5 [Rhoades, et al. 1992] . The wide deployment of GPUs in the last several years has resulted in an increase in experimental research with graphics hardware. [Trendall and Steward 2000] gives a detailed summary of the types of computation available on modern GPUs. Within the realm of graphics applications, programmable graphics hardware has been used for procedural texturing and shading [Olano and Lastra 1998; Peercy, et al. 2000; Proudfoot, et al. 2001; Rhoades, et al. 1992] . Graphics hardware has also been used for volume visualization [Cabral, et al. 1994] . Recently, methods for using current and near-future GPUs for ray tracing computations have been described in [Carr, et al. 2002] and [Purcell, et al. 2002] , respectively. Other researchers have found ways to use graphics hardware for non-graphics applications. The use of rasterization hardware for robot motion planning is described in [Lengyel, et al. 1990] . [Hoff, et al. 1999] describes the use of z-buffer techniques for the computation of Voronoi diagrams. The PixelFlow SIMD graphics computer [Eyles, et al. 1997] was used to crack UNIX password encryption [Kedem and Ishihara 1999] , and graphics hardware has been used in the computation of artificial neural networks [Bohn 1998] . Our work uses CML to simulate dynamic phenomena that can be described by PDEs. Related to this is the visualization of flows described by PDEs, which has been implemented using graphics hardware to accelerate line integral convolution and Lagrangian-Eulerian advection [Heidrich, et al. 1999; Jobard, et al. 2001; Weiskopf, et al. 2001] . NVIDIA has demonstrated the Game of Life cellular automata running on their GPUs, as well as a 2D physically-based water simulation that operates much like our CML simulations [NVIDIA 2001a;NVIDIA 2001b] . 3.3 Common Operations A detailed description of the implementation of the specific simulations that we have modeled using CML would require more space than we have in this paper, so we will instead describe a few common CML operations, followed by details of their implementation. Our goal in these descriptions is to impart a feel for the kinds of operations that can be performed using a graphics hardware implementation of a CML model. 3.3.1 Diffusion and the Laplacian The divergence of the gradient of a scalar function is called the Laplacian [Weisstein 1999] : 2 2 2 2 2 ( , ) . T T T x y x y = + The Laplacian is one of the most useful tools for working with partial differential equations. It is an isotropic measure of the second spatial derivative of a scalar function. Intuitively, it can be used to detect regions of rapid change, and for this reason it is commonly used for edge detection in image processing. The discretized form of this equation is: 2 , 1, 1, , 1 , 1 , 4 i j i j i j i j i j i j T T T T T T + + = + + + . The Laplacian is used in all of the CML simulations that we have implemented. If the results of the application of a Laplacian operator at a node T i,j are scaled and then added to the value of T i,j itself, the result is diffusion [Weisstein 1999] : ' 2 , , , 4 d i j i j i j c T T T = + . (1) Here, c d is the coefficient of diffusion. Application of this diffusion operation to a lattice state will cause the state to diffuse through the lattice 1 . 3.3.2 Directional Forces Most dynamic simulations involve the application of force. Like all operations in a CML model, forces are applied via 1 See Appendix A for details of our diffusion implementation. 111 Harris, Coombe, Scheuermann, and Lastra / Simulation on Graphics Hardware The Eurographics Association 2002. computations on the state of a node and its neighbors. As an example, we describe a buoyancy operator used in convection and cloud formation simulations [Miyazaki, et al. 2001;Yanagita and Kaneko 1993;Yanagita and Kaneko 1997] . This buoyancy operator uses temperature state T to compute a buoyant velocity at a node and add it to the node's vertical velocity state, v: , , , 1, 1, 2 [2 ] b c i j i j i j i j i j v v T T T + = + . (2) Equation (2) expresses that a node is buoyed upward if its horizontal neighbors are cooler than it is, and pushed downward if they are warmer. The strength of the buoyancy is controlled via the parameter c b . 3.3.3 Computation on Neighbors Sometimes an operation requires more complex computation than the arithmetic of the simple buoyancy operation described above. The buoyancy operation of the boiling simulation described in Section 2.1 must also account for phase change, and is therefore more complicated: , , , , 1 , 1 2 [ ( ) ( )], ( ) tanh[ ( )]. i j i j i j i j i j c T T T T T T T T + = = (3) In Equation (3), s is the buoyancy strength coefficient, and (T) is an approximation of density relative to temperature, T. The hyperbolic tangent is used to simulate the rapid change of density of a substance around the phase change temperature, T c . A change in density of a lattice node relative to its vertical neighbors causes the temperature of the node to be buoyed upward or downward. The thing to notice in this equation is that simple arithmetic will not suffice the hyperbolic tangent function must be applied to the temperature at the neighbors above and below node (i,j). We will discuss how we can compute arbitrary functions using dependent texturing in Section 3.4. 3.4 State Representation and Storage Our goal is to maintain all state and operation of our simulations in the GPU and its associated memory. To this end, we use the frame buffer like a register array to hold transient state, and we use textures like main memory arrays for state storage. Since the frame buffer and textures are typically limited to storage of 8-bit unsigned integers, state values must be converted to this format before being written to texture. Texture storage can be used for both scalar and vector data. Because of the four color channels used in image generation, two-, three-, or four-dimensional vectors can be stored in each texel of an RGBA texture. If scalar data are needed, it is often advantageous to store more than one scalar state in a single texture by using different color channels. In our CML implementation of the Gray-Scott reaction-diffusion system, for example, we store the concentrations of both reactants in the same texture. This is not only efficient in storage but also in computation since operations that act equivalently on both concentrations can be performed in parallel. Physical simulation also requires the use of signed values. Most texture storage, however, uses unsigned fixed-point values. Although fragment-level programmability available in current GPUs uses signed arithmetic internally, the unsigned data stored in the textures must be biased and scaled before and after processing [NVIDIA 2002] . 3.5 Implementing CML Operations An iteration of a CML simulation consists of successive application of simple operations on the lattice. These operations consist of three steps: setup the graphics hardware rendering state, render a single quadrilateral fit to the view port, and store the rendered results into a texture. We refer to each of these setup-render-copy operations as a single pass. In practice, due to limited GPU resources (number of texture units, number of register combiners, etc.), a CML operation may span multiple passes. The setup portion of a pass simply sets the state of the hardware to correctly perform the rest of the pass. To be sure that the correct lattice nodes are sampled during the pass, texels in the input textures must map directly to pixels in the output of the graphics pipeline. To ensure that this is true, we set the view port to the resolution of the lattice, and the view frustum to an orthographic view fit to the lattice so that there is a one-to-one mapping between pixels in the rendering buffer and texels in the texture to be updated. The render-copy portion of each pass performs 4 suboperations: Neighbor Sampling, Computation on Neighbors, New State Computation, and State Update. Figure 3 illustrates the mapping of the suboperations to graphics hardware. Neighbor sampling and Computation on Neighbors are performed by the programmable texture mapping hardware. New State Computation performs arithmetic on the results of the previous suboperations using programmable texture blending. Finally, State Update feeds the results of one pass to the next by rendering or copying the texture blending results to a texture. Neighbor Sampling: Since state is stored in textures, neighbor sampling is performed by offsetting texture coordinates toward the neighbors of the texel being updated. For example, to sample the four nearest neighbor nodes of Figure 3: Components of a CML operation map to graphics hardware pipeline components. 112 Harris, Coombe, Scheuermann, and Lastra / Simulation on Graphics Hardware The Eurographics Association 2002. node (x,y), the texture coordinates at the corners of the quadrilateral mentioned above are offset in the direction of each neighbor by the width of a single texel. Texture coordinate interpolation ensures that as rasterization proceeds, every texel's neighbors will be correctly sampled. Note that beyond sampling just the nearest neighbors of a node, weighted averages of nearby nodes can be computed by exploiting the linear texture interpolation hardware available in GPUs. An example of this is our single-pass implementation of 2D diffusion, described in Appendix A. Care must be taken, though, since the precision used for the interpolation coefficients is sometimes lower than the rest of the texture pipeline. Computation on Neighbors: As described in Section 3.3.3, many simulations compute complex functions of the neighbors they sample. In many cases, these functions can be computed ahead of time and stored in a texture for use as a lookup table. The programmable texture shader functionality of recent GPUs provides several dependent texture addressing operations. We have implemented table lookups using the "DEPENDENT_GB_TEXTURE_ 2D_NV" texture shader of the GeForce 3. This shader provides memory indirect texture addressing the green and blue colors read from one texture unit are used as texture coordinates for a lookup into a second texture unit. By binding the precomputed lookup table texture to the second texture unit, we can implement arbitrary function operations on the values of the nodes (Figure 4). New State Computation: Once we have sampled the values of neighboring texels and optionally used them for function table lookups, we need to compute the new state of the lattice. We use programmable hardware texture blending to perform arithmetic operations including addition, multiplication, and dot products. On the GeForce 3 and 4, we implement this using register combiners [NVIDIA 2002] Register combiners take the output of texture shaders and rasterization as input, and provide arithmetic operations, user-defined constants, and temporary registers. The result of these computations is written to the frame buffer. State Update: Once the new state is computed, we must store it in a state texture. In our current implementation, we copy the newly-rendered frame buffer to a texture using the glCopyTexSubImage2D() instruction in OpenGL. Since all simulation state is stored in textures, our technique avoids large data transfers between the CPU and GPU during simulation and rendering. 3.6 Numerical Range of CML Simulations The physically based nature of CML simulations means that the ranges of state values for different simulations can vary widely. The graphics hardware we use to implement them, on the other hand, operates only on fixed-point fragment values in the range [0,1]. This means that we must normalize the range of a simulation into [0,1] before it can be implemented in graphics hardware. Because the hardware uses limited-precision fixed-point numbers, some simulations will be more robust to this normalization than others. The robustness of a simulation depends on several factors. Dynamic range is the ratio between a simulation's largest absolute value and its smallest non-zero absolute value. If a simulation has a high dynamic range, it may not be robust to normalization unless the precision of computation is high enough to represent the dynamic range. We refer to a simulation's resolution as the smallest absolute numerical difference that it must be able to discern. A simulation with a resolution finer than the resolution of the numbers used in its computation will not be robust. Finally, as the arithmetic complexity of a simulation increases, it will incur more roundoff error, which may reduce its robustness when using low-precision arithmetic. For example, the boiling simulation (Section 4.1) has a range of approximately [0,10], but its values do not get very close to zero, so its dynamic range is less than ten. Also, its resolution is fairly coarse, since the event to which it is most sensitive phase change is near the top of its range. For these reasons, boiling is fairly robust under normalization. Reaction-diffusion has a range of [0,1] so it does not require normalization. Its dynamic range, however, is on the order of 10 5 , which is much higher than that of the 8-bit numbers stored in textures. Fortunately, by scaling the coefficients of reaction-diffusion, we can reduce this dynamic range somewhat to get interesting results. However, as we describe in Section 4.3, it suffers from precision errors (See Section 5.1 for more discussion of precision issues). As more precision becomes available in graphics hardware, normalization will become less of an issue. When floating point computation is made available, simulations can be run within their natural ranges. Results We have designed and built an interactive framework, "CMLlab", for constructing and experimenting with CML simulations (Figure 5). The user constructs a simulation from a set of general purpose operations, such as diffusion and advection, or special purpose operations designed for specific simulations, such as the buoyancy operations described in Section 3.3. Each operation processes a set of input textures and produces a single output texture. The user connects the outputs and inputs of the selected operations into a directed acyclic graph. An iteration of the simulation consists of traversing the graph in depth-first fashion so that each operation is performed in order. The state textures resulting from an iteration are used as input state for the next iteration, and for displaying the simulated system. The Figure 4: Arbitrary function lookups are implemented using dependent texturing in graphics hardware. 113 Harris, Coombe, Scheuermann, and Lastra / Simulation on Graphics Hardware The Eurographics Association 2002. results of intermediate passes in a simulation iteration can be displayed to the user in place of the result textures. This is useful for visually debugging the operation of a new simulation. While 2D simulations in our framework use only 2D textures for storage of lattice state, 3D simulations can be implemented in two ways. The obvious way is to use 3D textures. However, the poor performance of copying to 3D textures in current driver implementations would make our simulations run much slower. Instead, we implement 3D simulations using a collection of 2D slices to represent the 3D volume. This has disadvantages over using true 3D textures. For example, we must implement linear filtering and texture boundary conditions (clamp or repeat) in software, wheras 3D texture functionality provides these in hardware. It is worth noting that we trade optimal performance for flexibility in the CMLLab framework. Because we want to allow a variety of models to be built from a set of operations, we often incur the expense of some extra texture copies in order to keep operations separate. Thus, our implementation is not optimal even faster rates are achievable on the same hardware by sacrificing operator reuse. To demonstrate the utility of hardware CML simulation in interactive 3D graphics applications, we have integrated the simulation system into a virtual environment built on a 3D game engine, "Wild Magic" [Eberly 2001] . Figure 7 (see color plate) is an image of a boiling witch's brew captured from a real-time demo we built with the engine. The demo uses our 3D boiling simulation (Section 4.1) and runs at 45 frames per second. We will now describe three of the CML simulations that we have implemented. The test computer we used is a PC with a single 2.0 GHz Pentium 4 processor and 512 MB of RAM. Tests were performed on this machine with both an NVIDIA GeForce 3 Ti 500 GPU with 64 MB of RAM, and an NVIDIA GeForce 4 Ti 4600 GPU with 128 MB of RAM. 4.1 Boiling We have implemented 2D and 3D boiling simulations as described in [Yanagita 1992] . Rather than simulate all components of the boiling phenomenon (temperature, pressure, velocity, phase of matter, etc.), their model simulates only the temperature of the liquid as it boils. The simulation is composed of successive application of thermal diffusion, bubble formation and buoyancy, latent heat transfer. Sections 3.3.1 and 3.3.3 described the first two of these, and Section 2.1 gave an overview of the model. For details of the latent heat transfer computation, we refer the reader to [Yanagita 1992] . Our implementation requires seven passes per iteration for the 2D simulation, and 9 passes per slice for the 3D simulation. Table 1 shows the simulation speed for a range of resolutions. For details of our boiling simulation implementation, see [Harris 2002b] . 4.2 Convection The Rayleigh-Bnard convection CML model of [Yanagita and Kaneko 1993] simulates convection using four CML operations: buoyancy (described in 3.3.2), thermal diffusion, temperature and velocity advection, and viscosity and pressure effect. The viscosity and pressure effect is implemented as 2 grad(div ) 4 v p k v v v k v = + + , where v is the velocity, k v is the viscosity ratio and k p is the coefficient of the pressure effect. The first two terms of this equation account for diffusion of the velocity, and the last term is the flow caused by the gradient of the mass flow around the lattice [Miyazaki, et al. 2001] . See [Miyazaki, et al. 2001;Yanagita and Kaneko 1993] for details of the discrete implementation of this operation. The remaining operation is advection of temperature and velocity by the velocity field. [Yanagita and Kaneko 1993] implements this by distributing state from a node to its neighbors according to the velocity at the node. In our implementation, this was made difficult by the precision Iterations Per Second Resolution Software GeForce 3 GeForce 4 Speedup 64x64 266.5 1252.9 1752.5 4.7 / 6.6 128x128 61.8 679.0 926.6 11.0 / 15.0 256x256 13.9 221.3 286.6 15.9 / 20.6 512x512 3.3 61.2 82.3 18.5 / 24.9 1024x1024 .9 15.5 21.6 17.2 / 24 32x32x32 25.5 104.3 145.8 4.1 / 5.7 64x64x64 3.2 37.2 61.8 11.6 / 19.3 128x128x128 .4 NA 8.3 NA / 20.8 Table 1: A speed comparison of our hardware CML boiling simulation to a software version. The speedup column gives the speedup for both GeForce 3 and 4. Figure 5: CMLlab, our interactive framework for building and experimenting with CML simulations. 114 Harris, Coombe, Scheuermann, and Lastra / Simulation on Graphics Hardware The Eurographics Association 2002. limitations of the hardware, so we used a texture shader-based advection operation instead. This operation advects state stored in a texture using the GL_OFFSET_TEXTURE_ 2D_NV dependent texture addressing mode of the GeForce 3 and 4. A description of this method can be found in [Weiskopf, et al. 2001] . Our 2D convection implementation (Figure 8 in the color plate section) requires 10 passes per iteration. We have not implemented a 3D convection simulation because GeForce 3 and 4 do not have a 3D equivalent of the offset texture operation. Due to the precision limitations of the graphics hardware, our implementation of convection did not behave exactly as described by [Yanagita and Kaneko 1993] . We do observe the formation of convective rolls, but the motion of both the temperature and velocity fields is quite turbulent. We believe that this is a result of low-precision arithmetic. 4.3 Reaction-Diffusion Reaction-Diffusion processes were proposed by [Turing 1952] and introduced to computer graphics by [Turk 1991;Witkin and Kass 1991] . They are a well-studied model for the interaction of chemical reactants, and are interesting due to their complex and often chaotic behavior. The patterns that emerge are reminiscent of patterns occurring in nature [Lee, et al. 1993] . We implemented the Gray-Scott model, as described in [Pearson 1993] . This is a two-chemical system defined by the initial value partial differential equations: 2 2 2 2 (1 ) ( ) , u v U D U UV F U t V D V UV F k V t = + = + + where F, k, D u , and D v . are parameters given in [Pearson 1993] . We have implemented 2D and 3D versions of this process, as shown in Figure 5 (2D), and Figures 1 and 9 (3D, on color plate). We found reaction-diffusion relatively simple to implement in our framework because we were able to reuse our existing diffusion operator. In 2D this simulation requires two passes per iteration, and in 3D it requires three passes per slice. A 256x256 lattice runs at 400 iterations per second in our interactive framework, and a 128x128x32 lattice runs at 60 iterations per second. The low precision of the GeForce 3 and 4 reduces the variety of patterns that our implementation of the Gray-Scott model produces. We have seen a variety of results, but much less diversity than produced by a floating point implementation. As with convection, this appears to be caused by the effects of low-precision arithmetic. Hardware Limitations While current GPUs make a good platform for CML simulation, they are not without problems. Some of these problems are performance problems of the current implementation, and may not be issues in the near future. NVIDIA has shown in the past that slow performance can often be alleviated via optimization of the software drivers that accompany the GPU. Other limitations are more fundamental. Most of the implementation limitations that we encountered were limitations that affected performance. We have found glCopyTexSubImage3D(), which copies the frame buffer to a slice of a 3D texture, to be much slower (up to three orders of magnitude) than glCopyTexSubImage2D() for the same amount of data. This prevented us from using 3D textures in our implementation. Once this problem is alleviated, we expect a 3D texture implementation to be faster and easier to implement, since it will remove the need to bind multiple textures to sample neighbors in the third dimension. Also, 3D textures provide hardware linear interpolation and boundary conditions (periodic or fixed) in all three dimensions. With our slice-based implementation, we must interpolate and handle boundary conditions in the third dimension in software. The ability to render to texture will also provide a speed improvement, as we estimate that in a complex 3D simulation, much of the processing time is spent copying rendered data from the frame buffer to textures (typically one copy per pass). When using 3D textures, we will need the ability to render to a slice of a 3D texture. 5.1 Precision The hardware limitation that causes the most problems to our implementation is precision. The register combiners in the GeForce 3 and 4 perform arithmetic using nine-bit signed fixed-point values. Without floating point, the programmer must scale and bias values to maintain them in ranges that maximize precision. This is not only difficult, it is subject to arithmetic error. Some simulations (such as boiling) handle this error well, and behave as predicted by a floating point implementation. Others, such as our reaction-diffusion implementation, are more sensitive to precision errors. We have done some analysis of the error introduced by low precision and experiments to determine how much precision is needed (For full details, see [Harris 2002a] ). We hypothesize that the diffusion operation is very susceptible to Figure 6: High-precision fragment computations in near future graphics hardware will enable accurate simulation of reaction-diffusion at hundreds of iterations per second. 115 Harris, Coombe, Scheuermann, and Lastra / Simulation on Graphics Hardware The Eurographics Association 2002. roundoff error, because in our experiments in CMLlab, iterated application of a diffusion operator never fully diffuses its input. We derive the error induced by each application of diffusion (in 2D) to a node (i,j) as , 3 (3 ) 4 d i j d x + + , where d is the diffusion coefficient, x i,j is the value at node (i, j), and is the amount of roundoff error in each arithmetic operation. Since d and x i,j are in the range [0,1], this error is bounded above by 4.75 d . With 8 bits of precision, is at most 2 -9 . This error is fairly large, meaning that a simulation that is sensitive to small numbers will quickly diverge. In an attempt to better understand the precision needs of our more sensitive simulations, we implemented a software version of our reaction-diffusion simulation with adjustable fixed-point precision. Through experimentation, we have found that with 14 or more bits of fixed-point precision, the behavior of this simulation is visually very similar to our single-precision floating-point implementation. Like the floating-point version, a diverse variety of patterns grow, evolve, and sometimes develop unstable formations that never cease to change. Figure 6 shows a variety of patterns generated with this 14-bit fixed-point simulation. Graphics hardware manufacturers are quickly moving toward higher-quality pixels. This goal, along with increasing programmability, makes high-precision computation essential. Higher precision, including floating-point fragment values, will become a standard feature of GPUs in the near future [Spitzer 2002] . With the increasing precision and programmability of GPUs, we believe that CML methods for simulating natural phenomena using graphics hardware will become very useful. Conclusions and Future Work In this paper, we have described a method for simulating a variety of dynamic phenomena using graphics hardware. We presented the coupled map lattice as a simple and flexible simulation technique, and showed how CML operations map to computer graphics hardware operations. We have described common CML operations and how they can be implemented on programmable GPUs. Our hardware CML implementation shows a substantial speed increase (up to 25 times on a GeForce 4) over the same simulations implemented to run on a Pentium 4 CPU. However, this comparison (and the speedup numbers in Table 1) should be taken with a grain of salt. While our CPU-based CML simulator is an efficient, straightforward implementation that obeys common cache coherence principles, it is not highly optimized, and could be accelerated by using vectorized CPU instructions. Our graphics hardware implementation is not highly optimized either. We sacrifice optimal speed for flexibility. The CPU version is also written to use single precision floats, while the GPU version uses fixed-point numbers with much less precision. Nevertheless, we feel that it would be difficult, if not impossible, to achieve a 25x speedup over our current CPU implementation by optimizing the code and using lower precision numbers. A more careful comparison and optimized simulations on both platforms would be useful in the future. "CMLlab", our flexible framework for building CML models, allows a user to experiment with simulations running on graphics processors. We have described various 2D and 3D simulations that we have implemented in this framework. We have also integrated our CML framework with a 3D game engine to demonstrate the use of 3D CML models in interactive scenes and virtual environments. In the future, we would like to add more flexibility to CMLlab. Users currently cannot define new, custom operations without writing C++ code. It would be possible, however, to provide generic, scriptable operators, since the user microcode that runs on the GPU can be dynamically loaded. We have described the problems we encountered in implementing CML in graphics hardware, such as limited precision and 3D texturing performance problems. We believe that these problems will be alleviated in near future generations of graphics hardware. With the continued addition of more texture units, memory, precision, and more flexible programmability, graphics hardware will become an even more powerful platform for visual simulation. Some relatively simple extensions to current graphics hardware and APIs would benefit CML and PDE simulation. For example, the ability to render to 3D textures could simplify and accelerate each pass of our simulations. One avenue for future research is to increase parallelization of simulations on graphics hardware. Currently, it is difficult to add multiple GPUs to a single computer because PCs have a single AGP port. If future PC hardware adds support for multiple GPUs, powerful multiprocessor machines could be built with these inexpensive processors. We plan to continue exploring the use of CML on current and future generations of graphics hardware. We are interested in porting our system to ATI Radeon hardware. The Radeon 8500 can sample more textures per pass and has more programmable texture addressing than GeForce 3, which could add power to CML simulations. Also, our current framework relies mostly on the power of the fragment processing pipeline, and uses none of the power available in the programmable vertex engine. We could greatly increase the complexity of simulations by taking advantage of this. Currently, this would incur additional cost for feedback of the output of the fragment pipeline (through the main memory) and back into the vertex pipeline, but depending on the application, it may be worth the expense. GPU manufacturers could improve the performance of this feedback by allowing textures in memory to be interpreted as vertex meshes for processing by the vertex engine, thus avoiding unneccessary transfers back to the host. We hope to implement the cloud simulation described by [Miyazaki, et al. 2001] in the near future, as well as other 116 Harris, Coombe, Scheuermann, and Lastra / Simulation on Graphics Hardware The Eurographics Association 2002. dynamic phenomena. Also, since the boiling simulation of [Yanagita 1992] models only temperature, and disregards surface tension, the bubbles are not round. We are interested in extending this simulation to improve its realism. We plan to continue exploring the use of computer graphics hardware for general computation. As an example, the anisotropic diffusion that can be performed on a GPU may be useful for image-processing and computer vision applications. Acknowledgements The authors would like to thank Steve Molnar, John Spitzer and the NVIDIA Developer Relations team for answering many questions. This work was supported in part by NVIDIA Corporation, US NIH National Center for Research Resources Grant Number P41 RR 02170, US Office of Naval Research N00014-01-1-0061, US Department of Energy ASCI program, and National Science Foundation grants ACR-9876914 and IIS-0121293. A Implementation of Diffusion On GeForce 3 hardware, the diffusion operation can be implemented more efficiently than the Laplacian operator itself. To do so, we rewrite Equation (1) as ' , , 1, 1, , 1 , 1 4 , ( , ) 1 (1 ) ( ) 4 1 [(1 ) ], 4 k d i j d i j i j i j i j i j d i j d n i j k c T c T T T T T c T c T + + = = + + + + = + where n k (x,y) represents the kth nearest neighbor of (x, y). In this form, we see that the diffusion operator is the average of four weighted sums of the center texel, T i,j and its four nearest neighbor texels. These weighted sums are actually linear interpolation computations, with c d as the parameter of interpolation. This means that we can implement the diffusion operation described by Equation 3 by enabling linear texture filtering, and using texture coordinate offsets of c d w , where w is the width of a texel as described in Section 3.5. References [Bohn 1998] Bohn, C.-A. Kohonen Feature Mapping Through Graphics Hardware. In Proceedings of 3rd Int. Conference on Computational Intelligence and Neurosciences 1998. 1998. [Cabral, et al. 1994] Cabral, B., Cam, N. and Foran, J. Accelerated Volume Rendering and Tomographic Reconstruction Using Texture Mapping Hardware. In Proceedings of Symposium on Volume Visualization 1994, 91-98. 1994. [Carr, et al. 2002] Carr, N.A., Hall, J.D. and Hart, J.C. The Ray Engine. In Proceedings of SIGGRAPH / Eurographics Workshop on Graphics Hardware 2002. 2002. [Dobashi, et al. 2000] Dobashi, Y., Kaneda, K., Yamashita, H., Okita, T. and Nishita, T. A Simple, Efficient Method for Realistic Animation of Clouds. In Proceedings of SIGGRAPH 2000, ACM Press / ACM SIGGRAPH, 19-28. 2000. [Eberly 2001] Eberly, D.H. 3D Game Engine Design. Morgan Kaufmann Publishers. 2001. [England 1978] England, J.N. A system for interactive modeling of physical curved surface objects. In Proceedings of SIGGRAPH 78 1978, 336-340. 1978. [Eyles, et al. 1997] Eyles, J., Molnar, S., Poulton, J., Greer, T. and Lastra, A. PixelFlow: The Realization. In Proceedings of 1997 SIGGRAPH / Eurographics Workshop on Graphics Hardware 1997, ACM Press, 57-68. 1997. [Fedkiw, et al. 2001] Fedkiw, R., Stam, J. and Jensen, H.W. Visual Simulation of Smoke. In Proceedings of SIGGRAPH 2001, ACM Press / ACM SIGGRAPH. 2001. [Foster and Metaxas 1997] Foster, N. and Metaxas, D. Modeling the Motion of a Hot, Turbulent Gas. In Proceedings of SIGGRAPH 1997, ACM Press / ACM SIGGRAPH, 181-188. 1997. [Harris 2002a] Harris, M.J. Analysis of Error in a CML Diffusion Operation. University of North Carolina Technical Report TR02-015. http://www.cs.unc.edu/~harrism/cml/dl/HarrisTR02-015.pdf . 2002a. [Harris 2002b] Harris, M.J. Implementation of a CML Boiling Simulation using Graphics Hardware. University of North Carolina Technical Report TR02-016. http://www.cs.unc.edu/~harrism/cml/dl/HarrisTR02-016.pdf . 2002b. [Heidrich, et al. 1999] Heidrich, W., Westermann, R., Seidel, H.-P. and Ertl, T. Applications of Pixel Textures in Visualization and Realistic Image Synthesis. In Proceedings of ACM Symposium on Interactive 3D Graphics 1999. 1999. [Hoff, et al. 1999] Hoff, K.E.I., Culver, T., Keyser, J., Lin, M. and Manocha, D. Fast Computation of Generalized Voronoi Diagrams Using Graphics Hardware. In Proceedings of SIGGRAPH 1999, ACM / ACM Press, 277-286 . 1999. [Jobard, et al. 2001] Jobard, B., Erlebacher, G. and Hussaini, M.Y. Lagrangian-Eulerian Advection for Unsteady Flow Visualization. In Proceedings of IEEE Visualization 2001. 2001. [Kaneko 1993] Kaneko, K. (ed.), Theory and applications of coupled map lattices. Wiley, 1993. [Kapral 1993] Kapral, R. Chemical Waves and Coupled Map Lattices. in Kaneko, K. ed. Theory and Applications of Coupled Map Lattices, Wiley, 135-168. 1993. [Kedem and Ishihara 1999] Kedem, G. and Ishihara, Y. Brute Force Attack on UNIX Passwords with SIMD Computer. In Proceedings of The 8th USENIX Security Symposium 1999. 1999. [Lee, et al. 1993] Lee, K.J., McCormick, W.D., Ouyang, Q. and Swinn, H.L. Pattern Formation by Interacting Chemical Fronts. Science, 261. 192-194. 1993. [Lengyel, et al. 1990] Lengyel, J., Reichert, M., Donald, B.R. and Greenberg, D.P. Real-Time Robot Motion Planning Using Rasterizing Computer Graphics Hardware. In Proceedings of SIGGRAPH 1990, 327-335. 1990. 117 Harris, Coombe, Scheuermann, and Lastra / Simulation on Graphics Hardware The Eurographics Association 2002. [Lindholm, et al. 2001] Lindholm, E., Kilgard, M. and Moreton, H. A User Programmable Vertex Engine. In Proceedings of SIGGRAPH 2001, ACM Press / ACM SIGGRAPH, 149-158. 2001. [Miyazaki, et al. 2001] Miyazaki, R., Yoshida, S., Dobashi, Y. and Nishita, T. A Method for Modeling Clouds Based on Atmospheric Fluid Dynamics. In Proceedings of The Ninth Pacific Conference on Computer Graphics and Applications 2001, IEEE Computer Society Press, 363-372. 2001. [Nagel and Raschke 1992] Nagel, K. and Raschke, E. Self-organizing criticality in cloud formation? Physica A, 182. 519-531. 1992. [Nishimori and Ouchi 1993] Nishimori, H. and Ouchi, N. Formation of Ripple Patterns and Dunes by Wind-Blown Sand. Physical Review Letters, 71 1. 197-200. 1993. [NVIDIA 2002] NVIDIA. NVIDIA OpenGL Extension Specifications. http://developer.nvidia.com/view.asp?IO=nvidia_opengl_specs . 2002. [NVIDIA 2001a] NVIDIA. NVIDIA OpenGL Game Of Life Demo. http://developer.nvidia.com/view.asp?IO=ogl_gameoflife . 2001a. [NVIDIA 2001b] NVIDIA. NVIDIA Procedural Texture Physics Demo. http://developer.nvidia.com/view.asp?IO=ogl_dynamic_bumpreflection . 2001b. [Olano and Lastra 1998] Olano, M. and Lastra, A. A Shading Language on Graphics Hardware: The PixelFlow Shading System. In Proceedings of SIGGRAPH 1998, ACM / ACM Press, 159-168. 1998. [Pearson 1993] Pearson, J.E. Complex Patterns in a Simple System. Science, 261. 189-192. 1993. [Peercy, et al. 2000] Peercy, M.S., Olano, M., Airey, J. and Ungar, P.J. Interactive Multi-Pass Programmable Shading. In Proceedings of SIGGRAPH 2000, ACM Press / ACM SIGGRAPH, 425-432. 2000. [Potmesil and Hoffert 1989] Potmesil, M. and Hoffert, E.M. The Pixel Machine: A Parallel Image Computer. In Proceedings of SIGGRAPH 89 1989, ACM, 69-78. 1989. [Proudfoot, et al. 2001] Proudfoot, K., Mark, W.R., Tzvetkov, S. and Hanrahan, P. A Real-Time Procedural Shading System for Programmable Graphics Hardware. In Proceedings of SIGGRAPH 2001, ACM Press / ACM SIGGRAPH, 159-170. 2001. [Purcell, et al. 2002] Purcell, T.J., Buck, I., Mark, W.R. and Hanrahan, P. Ray Tracing on Programmable Graphics Hardware. In Proceedings of SIGGRAPH 2002, ACM / ACM Press. 2002. [Qian, et al. 1996] Qian, Y.H., Succi, S. and Orszag, S.A. Recent Advances in Lattice Boltzmann Computing. in Stauffer, D. ed. Annual Reviews of Computational Physics III, World Scientific, 195-242. 1996. [Rhoades, et al. 1992] Rhoades, J., Turk, G., Bell, A., State, A., Neumann, U. and Varshney, A. Real-Time Procedural Textures. In Proceedings of Symposium on Interactive 3D Graphics 1992, ACM / ACM Press, 95-100. 1992. [Spitzer 2002] Spitzer, J. Shading and Game Development (Presentation on NVIDIA Technology). IBM EDGE Workshop. 2002. [Stam 1999] Stam, J. Stable Fluids. In Proceedings of SIGGRAPH 1999, ACM Press / ACM SIGGRAPH, 121-128 . 1999. [Toffoli and Margolus 1987] Toffoli, T. and Margolus, N. Cellular Automata Machines. The MIT Press. 1987. [Trendall and Steward 2000] Trendall, C. and Steward, A.J. General Calculations using Graphics Hardware, with Applications to Interactive Caustics. In Proceedings of Eurogaphics Workshop on Rendering 2000, Springer, 287-298 . 2000. [Turing 1952] Turing, A.M. The chemical basis of morphogenesis. Transactions of the Royal Society of London, B237. 37-72. 1952. [Turk 1991] Turk, G. Generating Textures on Arbitrary Surfaces Using Reaction-Diffusion. In Proceedings of SIGGRAPH 1991, ACM Press / ACM SIGGRAPH, 289-298 . 1991. [von Neumann 1966] von Neumann, J. Theory of Self-Reproducing Automata. University of Illinois Press. 1966. [Weiskopf, et al. 2001] Weiskopf, D., Hopf, M. and Ertl, T. Hardware-Accelerated Visualization of Time-Varying 2D and 3D Vector Fields by Texture Advection via Programmable Per-Pixel Operations. In Proceedings of Vision, Modeling, and Visualization 2001, 439-446. 2001. [Weisstein 1999] Weisstein, E.W. CRC Concise Encyclopedia of Mathematics. CRC Press. 1999. [Witkin and Kass 1991] Witkin, A. and Kass, M. Reaction-Diffusion Textures. In Proceedings of SIGGRAPH 1991, ACM Press / ACM SIGGRAPH, 299-308. 1991. [Wolfram 1984] Wolfram, S. Cellular automata as models of complexity. Nature, 311. 419-424. 1984. [Yanagita 1992] Yanagita, T. Phenomenology of boiling: A coupled map lattice model. Chaos, 2 3. 343-350. 1992. [Yanagita and Kaneko 1993] Yanagita, T. and Kaneko, K. Coupled map lattice model for convection. Physics Letters A, 175. 415-420. 1993. [Yanagita and Kaneko 1997] Yanagita, T. and Kaneko, K. Modeling and Characterization of Cloud Dynamics. Physical Review Letters, 78 22. 4297-4300. 1997. 118 Harris, Coombe, Scheuermann, and Lastra / Simulation on Graphics Hardware The Eurographics Association 2002. Figure 7: A CML boiling simulation running in an interactive 3D environment (the steam is a particle system). Figure 9: A sequence from our 3D version of the Gray-Scott reaction-diffusion model. Figure 8: A CML convection simulation. The left panel shows temperature; the right panel shows 2D velocity encoded in the blue and green color channels . 160

Coupled Map Lattice;Visual Simulation;Reaction-Diffusion;dynamic phenomena;Multipass Rendering;simulation;CML;graphic hardware;Graphics Hardware

149

Physiological Measures of Presence in Stressful Virtual Environments

A common measure of the quality or effectiveness of a virtual environment (VE) is the amount of presence it evokes in users. Presence is often defined as the sense of being there in a VE. There has been much debate about the best way to measure presence, and presence researchers need, and have sought, a measure that is reliable, valid, sensitive, and objective. We hypothesized that to the degree that a VE seems real, it would evoke physiological responses similar to those evoked by the corresponding real environment, and that greater presence would evoke a greater response. To examine this, we conducted three experiments, the results of which support the use of physiological reaction as a reliable, valid, sensitive, and objective presence measure. The experiments compared participants' physiological reactions to a non-threatening virtual room and their reactions to a stressful virtual height situation. We found that change in heart rate satisfied our requirements for a measure of presence, change in skin conductance did to a lesser extent, and that change in skin temperature did not. Moreover, the results showed that inclusion of a passive haptic element in the VE significantly increased presence and that for presence evoked: 30FPS &gt; 20FPS &gt; 15FPS.

Introduction Virtual environments (VEs) are the most sophisticated human-computer interfaces yet developed. The effectiveness of a VE might be defined in terms of enhancement of task performance, effectiveness for training, improvement of data comprehension, etc. A common metric of VE quality is the degree to which the VE creates in the user the subjective illusion of presence a sense of being in the virtual, as opposed to the real, environment. Since presence is a subjective condition, it has most commonly been measured by self-reporting, either during the VE experience or immediately afterwards by questionnaires. There has been vigorous debate as to how to best measure presence [Barfield et al. 1995; Ellis 1996; Freeman et al. 1998; IJsselsteijn and de Ridder 1998; Lombard and Ditton 1997; Regenbrecht and Schubert 1997; Schubert et al. 1999; Sheridan 1996; Slater 1999; Witmer and Singer 1998]. In order to study a VE's effectiveness in evoking presence, researchers need a well-designed and verified measure of the phenomena. This paper reports our evaluation of three physiological measures heart rate, skin conductance, and skin temperature as alternate operational measures of presence in stressful VEs. Since the concept and idea of measuring presence are heavily debated, finding a measure that could find wide acceptance would be ideal. In that hope, we investigated the reliability, validity, sensitivity, and objectivity of each physiological measure. Figure 1. Side view of the virtual environment. Subjects start in the Training Room and later enter the Pit Room. 1.2. Physiological Reaction as a Surrogate Measure of Presence As VE system and technology designers, we have sought for a presence measure that is Reliable produces repeatable results, both from trial to trial on the same subject and across subjects, Valid measures subjective presence, or at least correlates with well-established subjective presence measures, Sensitive discriminates among multiple levels of presence, and Objective is well shielded from both subject and experimenter bias. We hypothesize that to the degree that a VE seems real, it will evoke physiological responses similar to those evoked by the corresponding real environment, and that greater presence will evoke a greater response. If so, these responses can serve as objective surrogate measures of subjective presence. Of the three physiological measures in our studies, Change in Heart Rate performs best. It consistently differentiates among conditions with more sensitivity and more statistical power than the other physiological measures, and more than most of the self-reported measures. It also best correlates with the reported measures. Figure 2. View of the 20' pit from the wooden ledge. Change in Skin Temperature is less sensitive, less powerful, and slower responding than Change in Heart Rate, although its response curves are similar. It also correlates with reported measures. Our results and the literature on skin temperature reactions suggest that Change in Skin Temperature would differentiate among conditions better if the exposures to the stimulus were at least 2 minutes [McMurray 1999; Slonim 1974]. Ours averaged 1.5 minutes in each experiment. Change in Skin Conductance Level yielded significant differentiation in some experiments but was not so consistent as Change in Heart Rate. More investigation is needed to establish whether it can reliably differentiate among multiple levels of presence. Since Change in Heart Rate best followed the hypotheses, the remainder of this paper will treat chiefly the results for it. For a full account of all measures, please see [Meehan 2001]. 1.3. Our Environment and Measures We use a derivative of the compelling VE reported by Usoh et al. [1999]. Figure 1 shows the environment: a Training Room, quite ordinary, and an adjacent Pit Room, with an unguarded hole in the floor leading to a room 20 ft. below. On the upper level the Pit Room is bordered with a 2-foot wide walkway. The 18x32 foot, 2-room virtual space fits entirely within the working space of our lab's wide-area ceiling tracker. Users, equipped with a head-tracked stereoscopic head-mounted display, practice walking about and picking up and placing objects in the Training Room. Then they are told to carry an object into the next room and place it at a designated spot. The door opens, and they walk through it to an unexpected hazard, a virtual drop of 20 ft. if they move off the walkway. Below is a furnished Living Room (Figure 2). Users report feeling frightened. Some report vertigo. Some will not walk out on the ledge and ask to stop the experiment or demo at the doorway. A few boldly walk out over the hole, as if there were a solid glass floor. For most of us, doing that, if we can, requires conscious mustering of will. This environment, with its ability to elicit a fear reaction in users, enables investigation of physiological reaction as a measure of presence. If so strong a stress-inducing VE does not produce significant physiological reactions, a less stressful VE won't. This investigation is a first step. Follow-on research should investigate whether less stressful environments also elicit statistically significant physiological reactions. This remainder section will discuss the physiological measures we tested and the reported measures we used to evaluate validity. 1.3.1. The Physiological Measures As stated above, we investigated three physiological metrics that measure stress in real environments [Andreassi 1995; Guyton 1986; Weiderhold et al. 1998]: Change in heart rate ( Heart Rate). The heart beats faster in stress. Change in skin conductance ( Skin Conductance Level). The skin of the palm sweats more in stress, independently of temperature, so its conductance rises. Change in skin temperature ( Skin Temperature). Circulation slows in the extremities in stress, causing skin temperature to drop. Each of these measures was constructed to increase when the physiological reaction to the Pit Room was greater. Heart Rate = mean HR Pit Room mean HR Training Room. Skin Conductance = mean SC Pit Room mean SC Training Room Skin Temperature = mean ST Training Room mean ST Pit Room We first measured heart rate with a convenient finger-mounted blood-pulse plethysmograph, but the noise generated by the sensor moving on the finger made the signal unstable and unusable. We then went to more cumbersome chest-attached three-electrode electrocardiography (ECG). This gave a good signal. Skin conductivity and skin temperature were successfully measured on the fingers. Once connected, users reported forgetting about the physiological sensors they did not cause breaks in presence during the experiments. Figure 3 shows a subject wearing the physiological monitoring equipment. 1.3.2. The Reported Measures Reported Presence. We used the University College London (UCL) questionnaire [Slater et al. 1995; Usoh et al. 1999]. The UCL questionnaire contains seven questions that measure presence (Reported Presence), three questions that measure behavioral presence (Reported Behavioral Presence) does the user act as if in a similar real environment and three that measure ease of locomotion (Ease of Locomotion). Responses for each question are on a scale of 1 to 7. Reported Ease of Locomotion was administered for consistency with earlier experiments, but we do not report on it in this paper. Figure 3. Subject wearing HMD and physiological monitoring equipment in the "Pit Room". 646 Even though each question is rated on a scale of 1-7, Slater et al. use it only to yield a High-Presence/ Low-Presence result. A judgment must be made as to the high-low threshold. Slater et al. have investigated the use of 6 and 7 as "high" responses [ 6] and the use of 5, 6, and 7 as "high" responses [ 5] as well as other constructions: addition of raw scores, and a combination based on principal-components analysis. They have found that [ 6] better followed conditions [Slater et al. 1994], and, therefore they chose that construction. We found that the [ 5] construction better follows presence conditions but has lower correlations with our physiological measures. Therefore, in order to best follow the original intention of the measures, irrespective of the lower correlations with our measures, we choose the [ 5] construction. On the study for which data is published, Slater's subjects rarely (&lt;10%) reported "5" values; over 25% of our subjects did. One explanation for this difference in subjects' reporting may be that university students today expect more technically of a VE than they did several years ago and, therefore, are more likely to report lower values (5s) even for the most presence-inducing VEs. Reported Behavioral Presence. Three questions asked subjects if they behaved as if present when in the VE. The count of high scores [ 5] on these questions made up the Reported Behavioral Presence measure. Multiple Exposures Passive Haptics Frame Rate Presence in VEs Does presence decrease with exposures? Passive Haptics increase presence? Higher Frame Rate increases presence? Reliability of Measures Are repeated measures highly correlated? Regardless of condition, will the Pit Room evoke similar physiological reactions on every exposure? Validity Do results correlate with reported measures? Sensitivity of Measures Do measures detect any effect? Do measures distinguish between 2 conditions? Do measures distinguish among 4 conditions? Table 1. Questions investigated in each study. 1.4. Methods and Procedures 1.4.1. Experimental procedures. We conducted three experiments: Effects of Multiple Exposures on Presence (Multiple Exposures), Effects of Passive Haptics on Presence (Passive Haptics), and Effects of Frame Rate on Presence (Frame Rate). Each of the three studies investigated some interesting aspect of VEs and the properties of the physiological measures themselves. Table 1 summarizes all the questions studied. For all studies we excluded subjects who had previously experienced VEs more than three times. The experiments were also limited to subjects who were ambulatory, could use stereopsis for depth perception, had no history of epilepsy or seizure, were not overly prone to motion sickness, were in their usual state of good physical fitness at the time of the experiment, and were comfortable with the equipment. Multiple Exposures: 10 subjects (average age 24.4; = 8.2; 7 female, 3 male) were trained to pick up books and move about in the Training Room at which time a physiological baseline was taken. Subjects then carried a virtual book from the Training Room and placed it on a virtual chair on the far side of the Pit Room. After that, they returned to the Training Room. The subjects performed this task three times per day on four separate days. We investigated whether the presence-evoking power of a VE declines with multiple exposures. Heart Rate was not successfully measured in this study due to problems with the sensor. Figure 4. Subject in slippers with toes over 1.5-inch ledge. Passive Haptics: 52 subjects (average age 21.4; = 4.3; 16 female, 36 male) reported on two days. Subjects experienced the VE with the 1.5-inch wooden ledge on one of their two days. The 1.5-inch height was selected so that the edge-probing foot did not normally contact the real laboratory floor where the virtual pit was seen. On their other day, subjects experienced the VE without the ledge. Subjects were counterbalanced as to the order of presentation of the physical ledge. Subjects performed all exposures to the VE wearing only thin sock-like slippers (Figure 4). The task was the same as in the Multiple Exposures study except subjects were instructed to walk to the edge of the wooden platform, place their toes over the edge, and count to ten before they proceeded to the chair on the far side of the room to drop the book. We investigated whether the 1.5-inch wooden ledge increased the presence-evoking power of the VE. Frame Rate: 33 participants (average age 22.3; = 3.6; 8 female, 25 male) entered the VE four times on one day and were presented the same VE with a different frame rate each time. The four frame rates were 10, 15, 20, and 30 frames-per-second (FPS). Subjects were counterbalanced as to the order of presentation of the four All exposures First Exposure Only (Between Subjects) Study Variable Mean P % &gt; 0 N Mean P % &gt; 0 N Skin Conductance 2.3 mSiemens &lt; .001 99% 112 2.9 mSiemens .002 100% 9 Multiple Exposures Skin Temperature 0.6 o F &lt; .001 77% 94 1.2 o F .015 100% 7 Heart Rate 6.3 BPM &lt; .001 89% 92 6.2 BPM &lt; .001 85% 46 Skin Conductance 4.8 mSiemens &lt; .001 100% 100 4.7 mSiemens &lt; .001 100% 50 Passive Haptics Skin Temperature 1.1 o F &lt; .001 90% 98 1.1 o F &lt; .001 94% 49 Heart Rate 6. 3 BPM &lt; .001 91% 132 8.1 BPM &lt; .001 91% 33 Skin Conductance 2.0 mSiemens &lt; .001 87% 132 2.6 mSiemens &lt; .001 97% 33 Frame Rate Skin Temperature 0.8 o F &lt; .001 100% 132 1.0 o F &lt; .001 100% 33 Table 2. Summary of means and significance of differences ( ) between Training Room and Pit Room. The mean, P-value for the one-sample t-test, percentage of times the measure was &gt; 0, and number of samples are shown. The left side shows the means and significances of all exposures. The right side shows these for only subjects' first exposures. The greater mean is shown in bold. 647 frame rates. Subjects were trained to pick up and drop blocks in the Training Room and then carried a red block to the Pit Room and dropped it on a red X-target on the floor of the Living Room, a procedural improvement that forced subjects to look down into the pit. They then plucked from the air two other colored blocks floating in the Pit Room and dropped each on the correspondingly-colored Xs on the floor of the Living Room. The X-targets and the green and blue blocks are visible in Figures 1 and 2. In this study, we investigated the effect of several different frame rates on presence and hypothesized that the higher the frame rate, the greater the presence evoked. In all three studies, the amount of physical activity (walking, manipulating objects) was approximately balanced between the Pit and Training Rooms. This lessened any difference between the two rooms in physiological reaction due to physical activity. 1.4.2. Statistics In this paper, we define statistical significance at the 5% level, i.e. P &lt; 0.050. Findings significant at the 5% level are discussed as "demonstrated" or "shown". To find the best statistical model for each measure, we used Stepwise Selection and Elimination as described by Kleinbaum et al. [1998]. As they suggest, to account better (statistically) for variation in the dependent variable (e.g., Heart Rate), we included all variables in the statistical models that were significant at the P &lt; 0.100 level. The analysis of differences in physiological reaction between the Pit Room and the Training Room for all studies (Table 2) was performed with a One-Sample T-Test. The correlations among measures were performed using the Bivariate Pearson Correlation. We analyzed order effects and the effects on presence of passive haptics and frame rate with the Univariate General Linear Model, using the repeated measure technique described in the SAS 6.0 Manual [SAS 1990]. This technique allows one to investigate the effect of the condition while taking into account inter-subject variation, order effects, and the effects of factors that change from exposure to exposure such as loss of balance on the 1.5-inch ledge. Section 2 details our evaluation of physiological measures as a surrogate for presence. In Section 3, we analyze physiological reactions as between-subject measures. In Section 4, we summarize the results as they pertain to interesting aspects of VEs. Physiological measures of presence In this section, we discuss the reliability, validity, sensitivity, and objectivity of the physiological measures. 2.1. Reliability Reliability is "the extent to which the same test applied on different occasions ... yields the same result" [Sutherland 1996]. Specifically, we wanted to know whether the virtual environment would consistently evoke similar physiological reactions as the subject entered and remained in the Pit Room on several occasions. Inconsistency could manifest itself as either a systematic increase or decrease in reactions or in uncorrelated measures for repeated exposure to the same VE. In the Multiple Exposures study the condition was the same each time, so this was our purest measure of reliability. We also hypothesized that in the Passive Haptics and Frame Rate studies, regardless of condition, that the Pit Room would also evoke similar physiological reactions on every exposure. We hypothesized that simply being exposed to the Pit Room would cause a greater physiological reaction than the difference between "high" and "low" presence conditions. Therefore, all three studies provide information on reliability. As we hypothesized, the environment consistently evoked physiological reactions over multiple exposures to the Pit Room. When analyzing the data from all exposures, we found there were significant physiological reactions to the Pit Room: heart rate and skin conductance were significantly higher and skin temperature was significantly lower in the Pit Room in all three studies. Heart rate was higher in the Pit Room for 90% of the exposures to the VE, skin conductance was higher for nearly 95%, and skin temperature was lower for 90%. Table 2 shows the mean difference, t-test, percentage of occurrences where the measure was above zero, and the total count for each physiological measure for each study. It shows results both for all exposures taken together, which is the approach discussed for most of the paper, and for analysis of the first exposure only, which we discuss in Section 4. We also wanted to know whether the physiological reactions to the environment would diminish over multiple exposures. Since our hypotheses relied on presence in the VE evoking a stress reaction over multiple exposures (2-12 exposures), we wanted to know whether physiological reactions to the VE would drop to zero or become unusably small due to habituation. In fact, Skin Temperature, Reported Presence, Reported Behavioral Presence, and Heart Rate each decreased with multiple exposures in every study (although this effect was not always statistically significant), and Skin Conductance decreased in all but one study. None decreased to zero, though, even after twelve exposures to the VE. Table 3 shows the significant order effects. A decrease in physiological reaction over multiple exposures would not necessarily weaken validity, since the literature shows that habituation diminishes the stress reactions to real heights and other stressors [Abelson and Curtis, 1989; Andreassi 1995]. Since, however, the reported presence measures, not just the physiological stress measures, decrease over multiple exposures, the decreases may not be due to habituation to the stressor; there may also be, as Heeter hypothesized, a decrease in a VE's ability to evoke presence as novelty wears off [Heeter 1992]. Orienting Effect. In general, each measure decreased after the first exposure. Moreover, for each measure except Heart Rate, there was a significant decrease after the first exposure in at least one of the studies (see Table 3). For physiological responses, this is called an orienting effect a higher physiological reaction when one sees something novel [Andreassi 1995]. Though this term traditionally refers to physiological reactions, we will also use the term for the initial spike in the reported measures. We attempted, with only partial success, to overcome the orienting effect by exposing subjects to the environment once as part of their orientation to the experimental setup and prior to the data-gathering portion of the experiment. In the Passive Haptics and Frame Rate studies, subjects entered the VE for approximately two Order Effects Heart Rate ( BPM) Skin Conductance ( mSiemens) Skin Temperature ( o F) Reported Presence (Count "high") Reported Behavioral Presence (Count "high") Multiple Exposures NA -0.7 (1 st ) -0.9 (1 st ) -0.7 (1 st ) Passive Haptics - - -0.8 (1 st ) -0.4 (1 st ) Frame Rate -1.0 (Task) -0.8 (1 st ) -0.3 (1 st ) -0.2 (Task) Table 3. Significant order effects for each measure in each study. "(1 st )" indicates a decrease after the first exposure only. "(Task)" indicates a decrease over tasks on the same day. There was an order effect for each measure in at least one study. NA is "Not available". Significant results are listed at the P &lt; 0.050 level (bold) and P &lt; 0.100 (normal text). Full details given in [Meehan 2001]. 648 minutes and were shown both virtual rooms before the experiment started. These pre-exposures reduced but did not eliminate the orienting effects. 2.2. Validity Validity is "the extent to which a test or experiment genuinely measures what it purports to measure" [Sutherland 1996]. The concept of presence has been operationalized in questionnaires so the validity of the physiological measures can be established by investigating how well the physiological reactions correlate with one or more of the questionnaire-based measures of presence. We investigated their correlations with two such measures: Reported Presence and Reported Behavioral Presence. Reported Presence. Of the physiological measures, Heart Rate correlated best with the Reported Presence. There was a significant correlation in the Frame Rate study (corr. = 0.265, P &lt; 0.005) and no correlation (corr. = 0.034, P = 0.743) in the Passive Haptics study. In the Multiple Exposures study, where Heart Rate was not available, Skin Conductance had the highest correlation with Reported Presence (corr. = 0.245, P &lt; 0.010). Reported Behavioral Presence. Heart Rate had the highest correlation, and a significant one, with Reported Behavioral Presence in the Frame Rate study (corr. = 0.192, P &lt; 0.050), and there was no correlation between the two (corr. = 0.004, P = 0.972) in the Passive Haptics study. In the Multiple Exposures study, where Heart Rate was not measured, Skin Conductance had the highest correlation with reported behavioral presence (corr. = 0.290, P &lt; 0.005). The correlations of the physiological measures with the reported measures give some support to their validities. The validity of Heart Rate appears to be better established by its correlation with the well-established reported measures. There was also some support for the validity of Skin Conductance from its correlation with reported measures. Following hypothesized relationships. According to Singleton, the validation process includes "examining the theory underlying the concept being measured," and "the more evidence that supports the hypothesized relationships [between the measure and the underlying concept], the greater one's confidence that a particular operational definition is a valid measure of the concept" [Singleton et al. 1993]. We hypothesized that presence should increase with frame rate and with the inclusion of the 1.5-inch wooden ledge, since each of these conditions provides increased sensory stimulation fidelity. As presented in the next section, our physiological measures did increase with frame rate and with inclusion of the 1.5-inch wooden ledge. This helps validate the physiological reactions as measures of presence. 2.3. Sensitivity and multi-level sensitivity Sensitivity is "the likelihood that an effect, if present, will be detected" [Lipsey 1998]. The fact that the physiological measures reliably distinguished between subjects reaction in the Pit Room versus the Training Room in every study assured us of at least a minimal sensitivity. For example, heart rate increased an average across all conditions of 6.3 beats / minute (BPM) in the Pit Room (P &lt; 0.001) compared to the Training Room in both the Passive Haptics and Frame Rate studies. See Table 2 for a full account of sensitivity of physiological measures to the difference between the two rooms. Acrophobic patients', when climbing to the second story of a fire escape (with a handrail), waiting one minute, and looking down, averaged an increase in heart rate of 13.4 BPM [Emmelkamp and Felten 1985]. Our subjects were non-phobic, and our height was virtual; so, we would expect, and did find, our subjects' heart rate reactions to be lower but in the same direction. Multi-level sensitivity. For guiding VE technological development and for better understanding of the psychological phenomena of VEs, we need a measure that reliably yields a higher value as a VE is improved along some "goodness" dimension, i.e., is sensitive to multiple condition values. We distinguish this from sensitivity as described above and call this multi-level sensitivity. The Passive Haptics study provided us some evidence of the measures' ability to discriminate between two "high presence" situations. We have informally observed that walking into the Pit Room causes a strong reaction in users, and this reaction seems greater in magnitude than the differences in reaction to the Pit Room between any two experimental conditions (e.g., with and without the 1.5-inch wooden ledge). Therefore, we expected the differences in reaction among the conditions to be less than the differences between the two rooms. For example, in Passive Haptics, we expected there to be a significant difference in the physiological measures between the two conditions (with and without the 1.5-inch wooden ledge), but expected it to be less than the difference between the Training Room and Pit Room in the "lower" presence condition (without the 1.5-inch wooden ledge). For Heart Rate, we did find a significant difference between the two conditions of 2.7 BPM (P &lt; 0.050), and it was less than the inter-room difference for the without-ledge condition: 4.9 BPM. See Figure 5. Figure 6 shows that the differences among the conditions in the FR study are smaller in magnitude as compared to the differences between the two rooms. 0 2 4 6 8 10 No Physical Ledge With Physical Ledge Change in beats/minute 2.7, P&lt;0.050 4.9, P &lt;0.001 7.6, P &lt;0.001 Figure 5. Heart Rate in Passive Haptics study. In the Passive Haptics study, we investigated the multi-level sensitivity of the measures by testing whether presence was significantly higher with the 1.5-inch wooden ledge. Presence as measured by each of Heart Rate (2.7 BPM; P &lt; 0.050), Skin Conductance (0.8 mSiemens; P &lt; 0.050), and Reported Behavioral Presence (0.5 more "high" responses; P &lt; 0.005) was significantly higher with the wooden ledge. Reported Presence had a strong trend in the same direction (0.5 more "high" responses; P = 0.060). In the Frame Rate study, we investigated the multi-level sensitivity of the measures by testing whether presence increased significantly as graphic frame update rates increased. We hypothesized that physiological reactions would increase monotonically with frame rates of 10, 15, 20, and 30 FPS. They did not do exactly that (see Figure 6). During the 10 FPS condition, there was an anomalous reaction for all of the physiological measures and for Reported Behavioral Presence. That is, at 10 FPS, subjects had higher physiological reaction and reported more behavioral presence. We believe that this reaction at 10 FPS was due to discomfort, added lag, and reduced temporal fidelity while in the ostensibly dangerous situation of walking next to a 20-foot pit [Meehan 2001]. 649 We also observed that subjects often lost their balance while trying to inch to the edge of the wooden platform at this low frame rate; their heart rate jumped an average of 3.5 BPM each time they lost their balance (P &lt; 0.050). Statistically controlling for these Loss of Balance incidents improved the significance of the statistical model for Heart Rate and brought the patterns of responses closer to the hypothesized monotonic increase in presence with frame rate but did not completely account for the increased physiological reaction at 10 FPS. Loss of Balance was not significant in any other model. 0 1 2 3 4 5 6 7 8 9 10 10 15 20 30 Frame Rate Change in beats/minute Figure 6. heart rate, after correcting for Loss of Balance, at 10, 15, 20, and 30 frames per second. Beyond 10 FPS, Heart Rate followed the hypothesis. After we statistically controlled for Loss of Balance, Heart Rate significantly increased between 15 FPS and 30 FPS (3.2 BPM; P &lt; 0.005) and between 15 FPS and 20 FPS (2.4 BPM; P &lt; 0.050). There was also a non-significant increase between 20 FPS and 30 FPS (0.7 BPM; P = 0.483) and a non-significant decrease between 10 FPS and 15 FPS (1.6 BPM; P = 0.134). Reported Presence, and Reported Behavioral Presence also increased with frame rate from 15-20-30 FPS, but with less distinguishing power. These findings support the multi-level sensitivity of Heart Rate. 2.4. Objectivity The measure properties of reliability, validity, and multi-level sensitivity are established quantitatively. Objectivity can only be argued logically. We argue that physiological measures are inherently better shielded from both subject bias and experimenter bias than are either reported measures or measures based on behavior observations. Reported measures are liable to subject bias the subject reporting what he believes the experimenter wants. Post-experiment questionnaires are also vulnerable to inaccurate recollection and to modification of impressions garnered early in a run by impressions from later. Having subjects report during the session, whether by voice report or by hand-held instrument, intrudes on the very presence illusion one is trying to measure. Behavioral measures, while not intrusive, are subject to bias on the part of the experimenters who score the behaviors. Physiological measures, on the other hand, are much harder for subjects to affect, especially with no biofeedback. These measures are not liable to experimenter bias, if instructions given to the participants are properly limited and uniform. We read instructions from a script in the Multiple Exposures study. We improved our procedure in the later Passive Haptics and Frame Rate studies by playing instructions from a compact disk player located in the real laboratory and represented by a virtual radio in the VE. 2.5. Summary and discussion The data presented here show that physiological reactions can be used as reliable, valid, multi-level sensitive, and objective measures of presence in stressful VEs. Of the physiological measures, Heart Rate performed the best. There was also some support for Skin Conductance. Heart Rate significantly differentiated between the Training Room and the Pit Room, and although this reaction faded over multiple exposures, it never decreased to zero. It correlated with the well-established reported measure, the UCL questionnaire. It distinguished between the presence and absence of passive haptics and among frame rates at and above 15 FPS. As we argued above, it is objective. In total, it satisfies all of the requirements for a reliable, valid, multi-level sensitive, and objective measure of presence in a stressful VE. Skin Conductance has some, but not all, of the properties we desire in a measure of presence. In particular, it did not differentiate among frame rates. We do not have a theory as to why. Although, Heart Rate satisfied the requirements for a presence measure for our VE, which evokes a strong reaction, it may not for less stressful VEs. To determine whether physiological reaction can more generally measure presence, a wider range of VEs must be tested, including less stressful, non-stressful, and relaxing environments. Investigation is currently under way to look at physiological reaction in relaxing 3D Television environments [Dillon et al. 2001]. The height reaction elicited by our VE could be due to vertigo, fear, or other innate or learned response. The reactions are well known in the literature and manifest as increased heart rate and skin conductance and decreased skin temperature [Andreassi 1995; Guyton 1986]. We hypothesized that the more present a user feels in our stressful environment, the more physiological reaction the user will exhibit. What causes this higher presence and higher physiological reaction? Is it due to a more realistic flow of visual information? Is it due to more coherence between the visual and haptic information? Is it due to the improved visual realism? All of these are likely to improve presence. We cannot, however, answer these questions definitively. We can say, though, that we have empirically shown that physiological reaction and reported presence are both higher when we present a "higher presence" VE. Whatever it is that causes the higher reported presence and physiological reaction, it causes more as we improve the VE. An additional desirable aspect of a measure is ease of use in the experimental setting. We did not record the time needed for each measure, but after running many subjects we can say with some confidence that use of the physiological monitoring and of the presence questionnaire each added approximately the same amount of time to the experiment. It took about five minutes per exposure to put on and take off the physiological sensors. It took about an extra minute at the beginning and end of each set of exposures to put on and take off the ECG sensor it was left on between exposures on the same day. It took subjects about five minutes to fill out the UCL Presence Questionnaire. It took some training for experimenters to learn the proper placement of the physiological equipment on the hands and chest of the subject thirty minutes would probably be sufficient. Another aspect of ease of use is the amount of difficulty participants have with the measure and to what extent the measure, if concurrent with an experimental task, interferes with the task. No subjects reported difficulties with the questionnaires. Only 650 about one in ten subjects reported noticing the physiological monitoring equipment on the hands during the VE exposures. Our experiment, though, was designed to use only the right hand, keeping the sensor-laden left hand free from necessary activity. No subjects reported noticing the ECG sensor once it was attached to the chest. In fact, many subjects reported forgetting about the ECG electrodes when prompted to take them off at the end of the day. There are groups investigating less cumbersome equipment, which would probably improve ease of use, including a physiological monitoring system that subjects wear like a shirt [Cowings et al. 2001]. Overall, questionnaires and physiological monitoring were both easy to use and non-intrusive. Physiological reactions as between-subjects measures We conducted all of the studies as within-subjects to avoid the variance due to natural human differences. That is, each subject experienced all of the conditions for the study in which she participated. This allowed us to look at relative differences in subject reaction among conditions and to overcome the differences among subjects in reporting and physiological reaction. The UCL questionnaire has been used successfully between-subjects [Usoh et al. 1999]. We suspected, however, that physiological reaction would not perform as well if taken between-subjects . We expected the variance among subjects would mask, at least in part, the differences in physiological reaction evoked by the different conditions. We investigated this hypotheses by analyzing the data using only the first task for each subject eliminating order effects and treating the reduced data sets as between-subjects experiments. That is, we treat each experiment as if only the first task for each subject was run. This means that the analysis uses only 10 data points (10 subjects first exposure only) for the Multiple Exposures study, 52 data points for the Passive Haptics study, and 33 data points for the Frame Rate study. Reliability between-subjects: Physiological reaction in the Pit Room. Even between subjects, we expected that there would be a consistent physiological reaction to the Pit Room, since we expected such a reaction for every exposure to the VE. We expected the significance to be lower, however, because of the reduced size of the data set. We found exactly that. The right half of Table 2 shows the values of the physiological measures averaged across conditions for the between-subjects analysis. As compared to the full data set, the between-subjects data have lower significance values, but subjects still have strong physiological reactions to the Pit Room. Table 2 demonstrates that the physiological orienting effects caused the averages for the first exposures to be higher than for the full data set. Validity between-subjects: Correlation with established measures. We expected correlations with the reported measures to be lower when taken between subjects since there were fewer data points and individual differences in physiological reaction and reporting would confound the correlations. This was the case. No physiological measure correlated significantly with any reported measure when analyzing between-subjects. Multi-level sensitivity between-subjects: Differentiating among presence conditions. We expected inter-subject variation in physiological reaction to mask the differences in physiological reactions evoked by the presence conditions (e.g., various frame rates). Contrary to this expectation, however, we found strong trends in the physiological measures among conditions in both the Passive Haptics and Frame Rate studies. (The condition was not varied in the Multiple Exposures study.) In the Passive Haptics study, both Heart Rate and Skin Conductance both varied in the expected direction non-significantly (3.3 BPM, P = 0.097; 1.0 mSiemens, P = 0.137, respectively). In the Frame Rate study, Heart Rate followed hypothesized patterns, but Skin Conductance did not. After the anomalous reaction at 10 FPS (as in full data set compare Figures 6 and 7), Heart Rate differentiated among presence conditions: at 30 FPS it was higher than at 15 FPS, and this difference was nearly significant (7.2 BPM; P = 0.054). Overall, Heart Rate shows promise as a between-subjects measure of presence. Though it did not correlate well with the reported measures (between-subjects), it did differentiate among the conditions with some statistical power in Passive Haptics and Frame Rate. Skin Conductance did not show as much promise as a between-subjects measure. For more discussion of physiological reactions as between-subjects measures of presence, see [Meehan 2001]. 0 2 4 6 8 10 12 14 10 15 20 30 Frame Rate Change in beat s/minut e Figure 7. Between-subjects analysis: Heart Rate. VE Effectiveness results Above we described the experiments as they related to the testing of the physiological presence measures, below we discuss each experiment with respect to the aspect of VEs it investigated. Effect of Multiple Exposures on Presence. As described in Section 1.4.1, ten users go through the same VE twelve times (over four days) in order to study whether the presence inducing power of a VE declines, or becomes unusably small, over multiple exposures. We did find significant decreases in each presence measure (reported and physiological) in either this experiment or one of the subsequent two experiments (see Table 3). However, none of the measures decreased to zero nor did any become unusably small. The findings support our hypothesis that all presence measures decrease over multiple exposures to the same VE, but not to zero. Effect of Passive Haptics on Presence. Our hypothesis was that supplementing a visual-aural VE with even rudimentary, low-fidelity passive haptics cues significantly increases presence. This experiment was only one of a set of studies investigating the passive haptics hypothesis. The detailed design, results, and discussion for the set are reported elsewhere [Insko 2001]. We found significant support for the hypothesis in that, with the inclusion of the 1.5-inch ledge, presence as measured by Heart Rate, Reported Behavioral Presence, and Skin Conductance was significantly higher at the P &lt; 0.05 level. Reported Presence also had a strong trend (P &lt; 0.10) in the same direction. 651 Effect of Frame Rate on Presence. Our hypothesis was that as frame rate increases from 10, 15, 20, 30 frames/second, presence increases. For frame rates of 15 frames/second and above, the hypothesis was largely confirmed. It was confirmed with statistical significance for 15 to 20 FPS and 15 to 30 FPS. 20 to 30 FPS though not statistically significant was in the same direction. 10 FPS gave anomalous results on all measures except Reported Presence, which increased monotonically with frame rate with no statistical significance. Future Work Given a compelling VE and a sensitive, quantitative presence measure, the obvious strategy is to degrade quantitative VE quality parameters in order to answer the questions: What makes a VE compelling? What are the combinations of minimum system characteristics to achieve this? For example, we would like to study the effect of Latency Self-avatar fidelity Aural localization Visual Detail Lighting Realism Realistic physics in interactions with objects Interactions with other people or agents Then we hope to begin to establish trade-offs for presence evoked: Is it more important to have latency below 50 ms or frame rate above 20 FPS? Additionally, we must eliminate the cables that tether subjects to the monitoring, tracking, and rendering equipment. Our subjects reported this encumbrance as the greatest cause of breaks in presence. Acknowledgements We would like to thank the University of North Carolina (UNC) Graduate School, the Link Foundation, and the National Institutes of Health National Center for Research Resources (Grant Number P41 RR 02170) for funding this project. We would like to thank the members of the Effective Virtual Environments group, the UNC Computer Science Department, and Dr. McMurray of the UNC Applied Physiology Department. Without their hard work, none of this research would have been possible. We would like to thank Drs. Slater, Usoh, and Steed of the University College of London who built much of the foundation for this work. We would also like to thank the reviewers for their thoughtful comments and suggestions. References Abelson, J. L. and G. C. Curtis (1989). Cardiac and neuroendocrine responses to exposure therapy in height phobics. Behavior Research and Therapy, 27(5): 561-567. Andreassi, J. L. (1995). Psychophysiology: Human behavior and physiological response. Hillsdale, N.J., Lawrence Erlbaum Associates. Barfield, W., T. Sheridan, D. Zeltzer and M. Slater (1995). Presence and performance within virtual environments. In W. Barfield and T. Furness, Eds., Virtual environments and advanced interface design. London, Oxford University Press. Cowings, P., S. Jensen, D. Bergner and W. Toscano (2001). A lightweight ambulatory physiological monitoring system. NASA Ames, California. Dillon, C., E. Keogh, J. Freeman and J. Davidoff (2001). Presence: Is your heart in it? 4th Int. Wkshp. on Presence, Philadelphia. Ellis, S. R. (1996). Presence of mind: A reaction to Thomas Sheridan's &quot;Further musings on the psychophysics of presence&quot;. Presence: Teleoperators and Virtual Environments, 5(2): 247-259. Emmelkamp, P. and M. Felten (1985). The process of exposure in vivo: cognitive and physiological changes during treatment of acrophobia. Behavior Research and Therapy, 23(2): 219. Freeman, J., S. E. Avons, D. Pearson, D. Harrison and N. Lodge (1998). Behavioral realism as a metric of presence. 1st Int. Wkshp. on Presence. Guyton, A. C. (1986). Basic characteristics of the sympathetic and parasympathetic function. In Textbook of Medical Physiology, 688-697. Philadelphia, W.B. Saunders Company. Heeter, C. (1992). Being there: The subjective experience of presence. Presence: Teleoperators and Virtual Environments, 1: 262-271. IJsselsteijn, W. A. and H. d. Ridder (1998). Measuring temporal variations in presence. 1st Int. Wkshp. on Presence. B. Insko (2001). Passive haptics significantly enhance virtual environments, Doctoral Dissertation. Computer Science. University of North Carolina, Chapel Hill, NC, USA. Kleinbaum, D., L. Kupper, K. Muller and A. Nizam (1998). Applied regression analysis and other multivariate methods. Lipsey, M. W. (1998). Design sensitivity: Statistical power for applied experimental research. In L. Brickman and D. J. Rog, Eds., Handbook of applied social research methods, 39-68. Thousand Oaks, California, Sage Publications, Inc. Lombard, M. and T. Ditton (1997). At the heart of it all: The concept of presence. Journal of Computer Mediated Communication, 3(2). McMurray, D. R. (1999). Director of Applied Physiology lab, University of North Carolina. Personal Communication. M. Meehan (2001). Physiological reaction as an objective measure of presence in virtual environments. Doctoral Dissertation. Computer Science. University of North Carolina, Chapel Hill, NC, USA. Regenbrecht, H. T. and T. W. Schubert (1997). Measuring presence in virtual environments. In Proc. of Human Computer Interface International, San Francisco. SAS (1990). SAS/ STAT User's Guide, Version 6, Fourth Edition. Cary, NC, USA, SAS Institute Inc. Schubert, T., F. Friedmann and H. Regenbrecht (1999). Embodied presence in virtual environments. In R. Paton and I. Neilson, Eds., Visual Representations and Interpretations. London, Springer-Verlag. Sheridan, T. B. (1996). Further musings on the psychophysics of presence. Presence: Teleoperators and Virtual Environments, 5(2): 241-246. Singleton, R. A., B. C. Straits and M. M. Straits (1993). Approaches to Social Research. New York, Oxford University Press. Slater, M., M. Usoh and A. Steed (1994). Depth of presence in virtual environments. Presence: Teleoperators and Virtual Environments, 3(2): 130-144. Slater, M., M. Usoh and A. Steed (1995). Taking steps: The influence of a walking technique on presence in virtual reality. ACM Transactions on Computer Human Interaction (TOCHI), 2(3): 201-219. Slater, M. (1999). Measuring Presence: A Response to the Witmer and Singer Presence Questionnaire. Presence: Teleoperators and Virtual Environments, 8(5): 560-565. Slonim, N. B., Ed. (1974). Environmental Physiology. Saint Louis. The C. V. Mosby Company. Sutherland, S. (1996). The international dictionary of psychology. New York, The Crossroads Publishing Company. Usoh, M., K. Arthur, M. Whitton, R. Bastos, A. Steed, M. Slater and F. Brooks (1999). Walking &gt; walking-in-place &gt; flying in virtual environments. In Proc. of ACM SIGGRAPH 99. ACM Press/ ACM SIGGRAPH. Weiderhold, B. K., R. Gervirtz and M. D. Wiederhold (1998). Fear of flying: A case report using virtual reality therapy with physiological monitoring. CyberPsychology and Behavior, 1(2): 97-104. Witmer, B. G. and M. J. Singer (1998). Measuring presence in virtual environments: A presence questionnaire. Presence: Teleoperators and Virtual Environments, 7(3): 225-240. 652

presence;Haptics;measurement;Frame Rate;virtual environment;Presence;Physiology

15

A New Statistical Formula for Chinese Text Segmentation Incorporating Contextual Information

A new statistical formula for identifying 2-character words in Chinese text, called the contextual information formula, was developed empirically by performing stepwise logistic regression using a sample of sentences that had been manually segmented. Contextual information in the form of the frequency of characters that are adjacent to the bigram being processed as well as the weighted document frequency of the overlapping bigrams were found to be significant factors for predicting the probablity that the bigram constitutes a word. Local information (the number of times the bigram occurs in the document being segmented) and the position of the bigram in the sentence were not found to be useful in determining words. The contextual information formula was found to be significantly and substantially better than the mutual information formula in identifying 2-character words. The method can also be used for identifying multi-word terms in English text.

INTRODUCTION Chinese text is different from English text in that there is no explicit word boundary. In English text, words are separated by spaces. Chinese text (as well as text of other Oriental languages) is made up of ideographic characters, and a word can comprise one, two or more such characters, without explicit indication where one word ends and another begins. This has implications for natural language processing and information retrieval with Chinese text. Text processing techniques that have been developed for Western languages deal with words as meaningful text units and assume that words are easy to identify. These techniques may not work well for Chinese text without some adjustments. To apply these techniques to Chinese text, automatic methods for identifying word boundaries accurately have to be developed. The process of identifying word boundaries has been referred to as text segmentation or, more accurately, word segmentation. Several techniques have been developed for Chinese text segmentation. They can be divided into: 1. statistical methods, based on statistical properties and frequencies of characters and character strings in a corpus (e.g. [13] and [16]). 2. dictionary-based methods, often complemented with grammar rules. This approach uses a dictionary of words to identify word boundaries. Grammar rules are often used to resolve conflicts (choose between alternative segmentations) and to improve the segmentation (e.g. [4], [8], [19] and [20]). 3. syntax-based methods, which integrate the word segmentation process with syntactic parsing or part-of-speech tagging (e.g. [1]). 4. conceptual methods, that make use of some kind of semantic processing to extract information and store it in a knowledge representation scheme. Domain knowledge is used for disambiguation (e.g. [9]). Many researchers use a combination of methods (e.g. [14]). The objective of this study was to empirically develop a statistical formula for Chinese text segmentation. Researchers have used different statistical methods in segmentation, most of which were based on theoretical considerations or adopted from other fields. In this study, we developed a statistical formula empirically by performing stepwise logistic regression using a sample of sentences that had been manually segmented. This paper reports the new formula developed for identifying 2-character words, and the effectiveness of this formula compared with the mutual information formula. This study has the following novel aspects: The statistical formula was derived empirically using regression analysis. The manual segmentation was performed to identify meaningful words rather than simple words. Meaningful words include phrasal words and multi-word terms. In addition to the relative frequencies of bigrams and characters often used in other studies, our study also investigated the use of document frequencies and weighted Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SIGIR '99 8/99 Berkley, CA USA Copyright 1999 ACM 1-58113-096-1/99/0007 . . . $5.00 82 ) ( * ) ( ) ( log 2 C freq B freq BC freq document frequencies. Weighted document frequencies are similar to document frequencies but each document is weighted by the square of the number of times the character or bigram occurs in the document. Contextual information was included in the study. To predict whether the bigram BC in the character string A B C D constitutes a word, we investigated whether the frequencies for AB, CD, A and D should be included in the formula. Local frequencies were included in the study. We investigated character and bigram frequencies within the document in which the sentence occurs (i.e. the number of times the character or bigram appears in the document being segmented). We investigated whether the position of the bigram (at the beginning of the sentence, before a punctuation mark, or after a punctuation mark) had a significant effect. We developed a segmentation algorithm to apply the statistical formula to segment sentences and resolve conflicts. In this study, our objective was to segment text into meaningful words rather than simple words . A simple word is the smallest independent unit of a sentence that has meaning on its own. A meaningful word can be a simple word or a compound word comprising 2 or more simple words depending on the context. In many cases, the meaning of a compound word is more than just a combination of the meanings of the constituent simple words, i.e. some meaning is lost when the compound word is segmented into simple words. Furthermore, some phrases are used so often that native speakers perceive them and use them as a unit. Admittedly, there is some subjectivity in the manual segmentation of text. But the fact that statistical models can be developed to predict the manually segmented words substantially better than chance indicates some level of consistency in the manual segmentation. The problem of identifying meaningful words is not limited to Chinese and oriental languages. Identifying multi-word terms is also a problem in text processing with English and other Western languages, and researchers have used the mutual information formula and other statistical approaches for identifying such terms (e.g. [3], [6] and [7]). PREVIOUS STUDIES There are few studies using a purely statistical approach to Chinese text segmentation. One statistical formula that has been used by other researchers (e.g. [11] and [16]) is the mutual information formula. Given a character string A B C D , the mutual information for the bigram BC is given by the formula: MI(BC) = = log 2 freq(BC) log 2 freq(B) log 2 freq(C) where freq refers to the relative frequency of the character or bigram in the corpus (i.e. the number of times the character or bigram occurs in the corpus divided by the number of characters in the corpus). Mutual information is a measure of how strongly the two characters are associated, and can be used as a measure of how likely the pair of characters constitutes a word. Sproat & Shih [16] obtained recall and precision values of 94% using mutual information to identify words. This study probably segmented text into simple words rather than meaningful words. In our study, text was segmented into meaningful words and we obtained much poorer results for the mutual information formula. Lua [12] and Lua & Gan [13] applied information theory to the problem of Chinese text segmentation. They calculated the information content of characters and words using the information entropy formula I = - log 2 P, where P is the probability of occurrence of the character or word. If the information content of a character string is less than the sum of the information content of the constituent characters, then the character string is likely to constitute a word. The formula for calculating this loss of information content when a word is formed is identical to the mutual information formula. Lua & Gan [13] obtained an accuracy of 99% (measured in terms of the number of errors per 100 characters). Tung & Lee [18] also used information entropy to identify unknown words in a corpus. However, instead of calculating the entropy value for the character string that is hypothesized to be a word (i.e. the candidate word), they identified all the characters that occurred to the left of the candidate word in the corpus. For each left character, they calculated the probability and entropy value for that character given that it occurs to the left of the candidate word. The same is done for the characters to the right of the candidate word. If the sum of the entropy values for the left characters and the sum of the entropy values for the right characters are both high, than the candidate word is considered likely to be a word. In other words, a character string is likely to be a word if it has several different characters to the left and to the right of it in the corpus, and none of the left and right characters predominate (i.e. not strongly associated with the character string). Ogawa & Matsuda [15] developed a statistical method to segment Japanese text. Instead of attempting to identify words directly, they developed a formula to estimate the probability that a bigram straddles a word boundary. They referred to this as the segmentation probability. This was complemented with some syntactic information about which class of characters could be combined with which other class. All the above mathematical formulas used for identifying words and word boundaries were developed based on theoretical considerations and not derived empirically. Other researchers have developed statistical methods to find the best segmentation for the whole sentence rather than focusing on identifying individual words. Sproat et al. [17] developed a stochastic finite state model for segmenting text. In their model, a word dictionary is represented as a weighted finite state transducer. Each weight represents the estimated cost of the word (calculated using the negative log probability). Basically, the system selects the sentence segmentation that has the smallest total cost. Chang & Chen [1] developed a method for word segmentation and part-of-speech tagging based on a first-order hidden Markov model. 83 RESEARCH METHOD The purpose of this study was to empirically develop a statistical formula for identifying 2-character words as well as to investigate the usefulness of various factors for identifying the words. A sample of 400 sentences was randomly selected from 2 months (August and September 1995) of news articles from the Xin Hua News Agency, comprising around 2.3 million characters. The sample sentences were manually segmented. The segmentation rules described in [10] were followed fairly closely. More details of the manual segmentation process, especially with regard to identifying meaningful words will be given in [5]. 300 sentences were used for model building, i.e. using regression analysis to develop a statistical formula. 100 sentences were set aside for model validation to evaluate the formula developed in the regression analysis. The sample sentences were broken up into overlapping bigrams. In the regression analysis, the dependent variable was whether a bigram was a two-character word according to the manual segmentation. The independent variables were various corpus statistics derived from the corpus (2 months of news articles). The types of frequency information investigated were: 1. Relative frequency of individual characters and bigrams (character pairs) in the corpus, i.e. the number of times the character or bigram occurs in the corpus divided by the total number of characters in the corpus. 2. Document frequency of characters and bigrams, i.e. the number of documents in the corpus containing the character or bigram divided by the total number of documents in the corpus. 3. Weighted document frequency of characters and bigrams. To calculate the weighted document frequency of a character string, each document containing the character string is assigned a score equal to the square of the number of times the character string occurs in the document. The scores for all the documents containing the character string are then summed and divided by the total number of documents in the corpus to obtain the weighted document frequency for the character string. The rationale is that if a character string occurs several times within the same document, this is stronger evidence that the character string constitutes a word, than if the character string occurs once in several documents. Two or more characters can occur together by chance in several different documents. It is less likely for two characters to occur together several times within the same document by chance. 4. Local frequency in the form of within-document frequency of characters and bigrams, i.e. the number of times the character or bigram occurs in the document being segmented. 5. Contextual information. Frequency information of characters adjacent to a bigram is used to help determine whether the bigram is a word. For the character string A B C D , to determine whether the bigram BC is a word, frequency information for the adjacent characters A and D, as well as the overlapping bigrams AB and BC were considered. 6. Positional information. We studied whether the position of a character string (at the beginning, middle or end of a sentence) gave some indication of whether the character string was a word. The statistical model was developed using forward stepwise logistic regression, using the Proc Logistic function in the SAS v.6.12 statistical package for Windows. Logistic regression is an appropriate regression technique when the dependent variable is binary valued (takes the value 0 or 1). The formula developed using logistic regression predicts the probability (more accurately, the log of the odds) that a bigram is a meaningful word. In the stepwise regression, the threshold for a variable to enter the model was set at the 0.001 significance level and the threshold for retaining a variable in the model was set at 0.01. In addition, preference was given to relative frequencies and local frequencies because they are easier to calculate than document frequencies and weighted document frequencies. Also, relative frequencies are commonly used in previous studies. Furthermore, a variable was entered in a model only if it gave a noticeable improvement to the effectiveness of the model. During regression analysis, the effectiveness of the model was estimated using the measure of concordance that was automatically output by the SAS statistical program. A variable was accepted into the model only if the measure of concordance improved by at least 2% when the variable was entered into the model. We evaluated the accuracy of the segmentation using measures of recall and precision. Recall and precision in this context are defined as follows: Recall = No. of 2-character words identified in the automatic segmentation that are correct No. of 2-character words identified in the manual segmentation Precision = No. of 2-character words identified in the automatic segmentation that are correct No. of 2-character words identified in the automatic segmentation STATISTICAL FORMULAS DEVELOPED The formula that was developed for 2-character words is as follows. Given a character string A B C D , the association strength for bigram BC is: Assoc(BC) = 0.35 * log 2 freq(BC) + 0.37 * log 2 freq(A) + 0.32 log 2 freq(D) 0.36 * log 2 docfreq wt (AB) 0.29 * log 2 docfreq wt (CD) + 5.91 where freq refers to the relative frequency in the corpus and docfreq wt refers to the weighted document frequency. We refer to this formula as the contextual information formula. More details of the regression model are given in Table 1. The formula indicates that contextual information is helpful in identifying word boundaries. A in the formula refers to the character preceding the bigram that is being processed, whereas D is the character following the bigram. The formula indicates that if the character preceding and the character following the bigram have high relative frequencies, then the bigram is more likely to be a word. 84 Contextual information involving the weighted document frequency was also found to be significant. The formula indicates that if the overlapping bigrams AB and CD have high weighted document frequencies, then the bigram BC is less likely to be a word. We tried replacing the weighted document frequencies with the unweighted document frequencies as well as the relative frequencies. These were found to give a lower concordance score. Even with docfreq (AB) and docfreq (CD) in the model, docfreq wt (AB) and docfreq wt (CD) were found to improve the model significantly. However, local frequencies were surprisingly not found to be useful in predicting 2-character words. We investigated whether the position of the bigram in the sentence was a significant factor. We included a variable to indicate whether the bigram occurred just after a punctuation mark or at the beginning of the sentence, and another variable to indicate whether the bigram occurred just before a punctuation mark or at the end of a sentence. The interaction between each of the position variables and the various relative frequencies were not significant. However, it was found that whether or not the bigram was at the end of a sentence or just before a punctuation mark was a significant factor. Bigrams at the end of a sentence or just before a punctuation mark tend to be words. However, since this factor did not improve the concordance score by 2%, the effect was deemed too small to be included in the model. It should be noted that the contextual information used in the study already incorporates some positional information. The frequency of character A (the character preceding the bigram) was given the value 0 if the bigram was preceded by a punctuation mark or was at the beginning of a sentence. Similarly, the frequency of character D (the character following the bigram) was given the value 0 if the bigram preceded a punctuation mark. We also investigated whether the model would be different for high and low frequency words. We included in the regression analysis the interaction between the relative frequency of the bigram and the other relative frequencies. The interaction terms were not found to be significant. Finally, it is noted that the coefficients for the various factors are nearly the same, hovering around 0.34. 4.2 Improved Mutual Information Formula In this study, the contextual information formula (CIF) was evaluated by comparing it with the mutual information formula (MIF). We wanted to find out whether the segmentation results using the CIF was better than the segmentation results using the MIF. In the CIF model, the coefficients of the variables were determined using regression analysis. If CIF was found to give better results than MIF, it could be because the coefficients for the variables in CIF had been determined empirically and not because of the types of variables in the formula. To reject this explanation, regression analysis was used to determine the coefficients for the factors in the mutual information formula. We refer to this new version of the formula as the improved mutual information formula. Given a character string A B C D , the improved mutual information formula is: Improved MI(BC) = 0.39 * log 2 freq(BC) - 0.28 * log 2 freq(B) 0 .23 log 2 freq(C) - 0.32 The coefficients are all close to 0.3. The formula is thus quite similar to the mutual information formula, except for a multiplier of 0.3. SEGMENTATION ALGORITHMS The automatic segmentation process has the following steps: 1. The statistical formula is used to calculate a score for each bigram to indicate its association strength (or how likely the bigram is a word). 2. A threshold value is then set and used to decide which bigram is a word. If a bigram obtains a score above the threshold value, then it is selected as a word. Different threshold values can be used, depending on whether the user prefers high recall or high precision. 3. A segmentation algorithm is used to resolve conflict. If two overlapping bigrams both have association scores above the Parameter Standard Wald Pr &gt; Standardized Variable DF Estimate Error Chi-Square Chi-Square Estimate INTERCPT 1 5.9144 0.1719 1184.0532 0.0001 . Log freq(BC) 1 0.3502 0.0106 1088.7291 0.0001 0.638740 Log freq(A) 1 0.3730 0.0113 1092.1382 0.0001 0.709621 Log freq(D) 1 0.3171 0.0107 886.4446 0.0001 0.607326 Log docfreq wt (AB) 1 -0.3580 0.0111 1034.0948 0.0001 -0.800520 Log docfreq wt (CD) 1 -0.2867 0.0104 754.2276 0.0001 -0.635704 Note: freq refers to the relative frequency, and docfreq wt refers to the weighted document frequency. Association of Predicted Probabilities and Observed Responses Concordant = 90.1% Somers' D = 0.803 Discordant = 9.8% Gamma = 0.803 Tied = 0.1% Tau-a = 0.295 (23875432 pairs) c = 0.901 Table 1. Final regression model for 2-character words 85 threshold value, then there is conflict or ambiguity. The frequency of such conflicts will rise as the threshold value is lowered. The segmentation algorithm resolves the conflict and selects one of the bigrams as a word. One simple segmentation algorithm is the forward match algorithm. Consider the sentence A B C D E . The segmentation process proceeds from the beginning of the sentence to the end. First the bigram AB is considered. If the association score is above the threshold, then AB is taken as a word, and the bigram CD is next considered. If the association score of AB is below the threshold, the character A is taken as a 1-character word. And the bigram BC is next considered. In effect, if the association score of both AB and BC are above threshold, the forward match algorithm selects AB as a word and not BC. The forward match method for resolving ambiguity is somewhat arbitrary and not satisfactory. When overlapping bigrams exceed the threshold value, it simply decides in favour of the earlier bigram. Another segmentation algorithm was developed in this study which we refer to as the comparative forward match algorithm. This has an additional step: If 2 overlapping bigrams AB and BC both have scores above the threshold value then their scores are compared. If AB has a higher value, then it is selected as a word, and the program next considers the bigrams CD and DE. On the other hand, if AB has a lower value, then character A is selected as a 1-character word, and the program next considers bigrams BC and CD. The comparative forward match method (CFM) was compared with the forward match method (FM) by applying them to the 3 statistical formulas (the contextual information formula, the mutual information formula and the improved mutual information formula). One way to compare the effectiveness of the 2 segmentation algorithms is by comparing their precision figures at the same recall levels. The precision figures for selected recall levels are given in Table 2. The results are based on the sample of 300 sentences. The comparative forward match algorithm gave better results for the mutual information and improved mutual information formulas especially at low threshold values when a large number of conflicts are likely. Furthermore, for the forward match method, the recall didn t go substantially higher than 80% even at low threshold values. For the contextual information formula, the comparative forward match method did not perform better than forward match, except at very low threshold values when the recall was above 90%. This was expected because the contextual information formula already incorporates information about neighboring characters within the formula. The formula gave very few conflicting segmentations. There were very few cases of overlapping bigrams both having association scores above the threshold except when threshold values were below 1.5. EVALUATION In this section we compare the effectiveness of the contextual information formula with the mutual information formula and the improved mutual information formula using the 100 sentences that had been set aside for evaluation purposes. For the contextual information formula, the forward match segmentation algorithm was used. The comparative forward match algorithm was used for the mutual information and the improved mutual information formulas. The three statistical formulas were compared by comparing their precision figures at 4 recall levels at 60%, 70%, 80% and 90%. For each of the three statistical formulas, we identified the threshold values that would give a recall of 60%, 70%, 80% and 90%. We then determined the precision values at these threshold values to find out whether the contextual information formula gave better precision than the other two formulas at 60%, 70%, 80% and 90% recall. These recall levels were selected because a recall of 50% or less is probably unacceptable for most applications. The precision figures for the 4 recall levels are given in Table 3. The recall-precision graphs for the 3 formulas are given in Fig. 1. The contextual information formula substantially outperforms the mutual information and the improved mutual information formulas. At the 90% recall level, the contextual information Precision Recall Comparative Forward Match Forward Match Improvement Mutual Information 90% 51% 80% 52% 47% 5% 70% 53% 51% 2% 60% 54% 52% 2% Improved Mutual Information 90% 51% 80% 53% 46% 7% 70% 54% 52% 2% 60% 55% 54% 1% Contextual Information Formula 90% 55% 54% 1% 80% 62% 62% 0% 70% 65% 65% 0% 60% 68% 68% 0% Table 2. Recall and precision values for the comparative forward match segmentation algorithm vs. forward match Precision Recall Mutual Information Improved Mutual Information Contextual Information 90% 57% (0.0) 57% (-2.5) 61% (-1.5) 80% 59% (3.7) 59% (-1.5) 66% (-0.8) 70% 59% (4.7) 60% (-1.0) 70% (-0.3) 60% 60% (5.6) 62% (-0.7) 74% (0.0) * Threshold values are given in parenthesis. Table 3. Recall and precision for three statistical formulas 86 formula was better by about 4%. At the 60% recall level, it outperformed the mutual information formula by 14% (giving a relative improvement of 23%). The results also indicate that the improved mutual information formula does not perform better than the mutual information formula. 6.2 Statistical Test of Significance In order to perform a statistical test, recall and precision figures were calculated for each of the 100 sentences used in the evaluation. The average recall and the average precision across the 100 sentences were then calculated for the three statistical formulas. In the previous section, recall and precision were calculated for all the 100 sentences combined. Here, recall and precision were obtained for individual sentences and then the average across the 100 sentences was calculated. The average precision for 60%, 70%, 80% and 90% average recall are given in Table 4. For each recall level, an analysis of variance with repeated measures was carried out to find out whether the differences in precision were significant. Pairwise comparisons using Tukey s HSD test was also carried out. The contextual information formula was significantly better ( =0.001) than the mutual information and the improved mutual information formulas at all 4 recall levels. The improved mutual information formula was not found to be significantly better than mutual information. ANALYSIS OF ERRORS The errors that arose from using the contextual information formula were analyzed to gain insights into the weaknesses of the model and how the model can be improved. There are two types of errors: errors of commission and errors of omission. Errors of commission are bigrams that are identified by the automatic segmentation to be words when in fact they are not (according to the manual segmentation). Errors of omission are bigrams that are not identified by the automatic segmentation to be words but in fact they are. The errors depend of course on the threshold values used. A high threshold (e.g. 1.0) emphasizes precision and a low threshold (e.g. 1.0) emphasizes recall. 50 sentences were selected from the 100 sample sentences to find the distribution of errors at different regions of threshold values. Association Score &gt;1.0 (definite errors) will through telegraph [on the] day [31 July] Association Score Between 1.0 and 1.0 (borderline errors) still to will be people etc. I want Person's name ( ) Wan Wen Ju Place name ( ) a village name in China ( ) Canada Name of an organization/institution ( ) Xin Hua Agency ( ) The State Department Table 6. Bigrams incorrectly identified as words 55 60 65 70 75 60 65 70 75 80 85 90 95 Recall(%) Precision(%) Contextual information Mutual information Improved mutual information Fig. 1. Recall-precision graph for the three statistical models. Association Score&gt;1.0 (definite errors) ( ) university (agricultural university) ( ) geology (geologic age) ( ) plant (upland plant) ( ) sovereignty (sovereign state) Association Score Between 1.0 and 1.0 (borderline errors) ( ) statistics (statistical data) ( ) calamity (natural calamity) ( ) resources (manpower resources) ( ) professor (associate professor) ( ) poor (pauperization) ( ) fourteen (the 14 th day) ( ) twenty (twenty pieces) Table 5. Simple words that are part of a longer meaningful word Avg Precision Avg Recall Mutual Information Improved Mutual Information Contextual Information 90% 57% (1.0) 58% (-2.3) 61% (-1.5) 80% 60% (3.8) 60% (-1.4) 67% (-0.7) 70% 59% (4.8) 60% (-1.0) 70% (-0.3) 60% 60% (5.6) 63% (-0.6) 73% (0.0) * Threshold values are given in parenthesis. Table 4. Average recall and average precision for the three statistical formulas 87 We divide the errors of commission (bigrams that are incorrectly identified as words by the automatic segmentation) into 2 groups: 1. Definite errors: bigrams with association scores above 1.0 but are not words 2. Borderline errors: bigrams with association scores between 1.0 and 1.0 and are not words We also divide the errors of omission (bigrams that are words but are not identified by the automatic segmentation) into 2 groups: 1. Definite errors: bigrams with association scores below 1.0 but are words 2. Borderline errors: bigrams with association scores between 1.0 and 1.0 and are words. 7.1 Errors of Commission Errors of commission can be divided into 2 types: 1. The bigram is a simple word that is part of a longer meaningful word. 2. The bigram is not a word (neither simple word nor meaningful word). Errors of the first type are illustrated in Table 5. The words within parenthesis are actually meaningful words but segmented as simple words (words on the left). The words lose part of the meaning when segmented as simple words. These errors occurred mostly with 3 or 4-character meaningful words. Errors of the second type are illustrated in Table 6. Many of the errors are caused by incorrectly linking a character with a function word or pronoun. Some of the errors can easily be removed by using a list of function words and pronouns to identify these characters. 7.2 Errors of Omission Examples of definite errors of omission (bigrams with association scores below 1.0 but are words) are given in Table 7. Most of the errors are rare words and time words. Some are ancient names, rare and unknown place names, as well as technical terms. Since our corpus comprises general news articles, these types of words are not frequent in the corpus. Time words like dates usually have low association values because they change everyday! These errors can be reduced by incorporating a separate algorithm for recognizing them. The proportion of errors of the various types are given in Table 8. CONCLUSION A new statistical formula for identifying 2-character words in Chinese text, called the contextual information formula, was developed empirically using regression analysis. The focus was on identifying meaningful words (including multi-word terms and idioms) rather than simple words. The formula was found to give significantly and substantially better results than the mutual information formula. Contextual information in the form of the frequency of characters that are adjacent to the bigram being processed as well as the weighted document frequency of the overlapping bigrams were found to be significant factors for predicting the probablity that the bigram constitutes a word. Local information (e.g. the number of times the bigram occurs in the document being segmented) and the position of the bigram in the sentence were not found to be useful in determining words. Of the bigrams that the formula erroneously identified as words, about 80% of them were actually simple words. Of the rest, many involved incorrect linking with a function words. Of the words that the formula failed to identify as words, more than a third of them were rare words or time words. The proportion of rare words increased as the threshold value used was lowered. These rare words cannot be identified using statistical techniques. This study investigated a purely statistical approach to text Association Score between -1.0 and -2.0 the northern section of a construction project fragments of ancient books Association Score &lt; -2.0 September 3rd day (name of a district in China ) (name of an institution) the Book of Changes Table 7. 2-character words with association score below -1.0 Errors of Commission Association score &gt; 1.0 (No. of errors=34) Borderline Cases Association score: 1.0 to1.0 (No. of cases: 210) Errors of Omission Association score &lt; 1.0 Association score: 1.0 to 2.0 (No. of errors=43) Association score &lt; 2.0 (No. of errors=22) Simple words 82.3% Not words 17.7% Simple words 55.2% Not words 20.5% Meaningful words 24.3% Rare words & time words 23.2% Others 76.8% Rare words & time words 63.6% Others 36.4% Table 8. Proportion of errors of different types 88 segmentation. The advantage of the statistical approach is that it can be applied to any domain, provided that the document collection is sufficiently large to provide frequency information. A domain-specific dictionary of words is not required. In fact, the statistical formula can be used to generate a shortlist of candidate words for such a dictionary. On the other hand, the statistical method cannot identify rare words and proper names. It is also fooled by combinations of function words that occur frequently and by function words that co-occur with other words. It is well-known that a combination of methods is needed to give the best segmentation results. The segmentation quality in this study can be improved by using a list of function words and segmenting the function words as single character words. A dictionary of common and well-known names (including names of persons, places, institutions, government bodies and classic books) could be used by the system to identify proper names that occur infrequently in the corpus. Chang et al. [2] developed a method for recognizing proper nouns using a dictionary of family names in combination with a statistical method for identifying the end of the name. An algorithm for identifying time and dates would also be helpful. It is not clear whether syntactic processing can be used to improve the segmentation results substantially. Our current work includes developing statistical formulas for identifying 3 and 4-character words, as well as investigating whether the statistical formula developed here can be used with other corpora. The approach adopted in this study can also be used to develop statistical models for identifying multi-word terms in English text. It would be interesting to see whether the regression model developed for English text is similar to the one developed in this study for Chinese text. Frantzi, Ananiadou & Tsujii [7], using a different statistical approach, found that contextual information could be used to improve the identification of multi-word terms in English text. REFERENCES [1] Chang, C.-H., and Chen, C.-D. A study of integrating Chinese word segmentation and part-of-speech tagging. Communications of COLIPS, 3, 1 (1993), 69-77. [2] Chang, J.-S., Chen, S.-D., Ker, S.-J., Chen, Y., and Liu, J.S. A multiple-corpus approach to recognition of proper names in Chinese texts. Computer Processing of Chinese and Oriental Languages, 8, 1 (June 1994), 75-85. [3] Church, K.W., and Hanks, P. Word association norms, mutual information and lexicography. In Proceedings of the 27 th Annual Meeting of the Association for Computational Linguistics (Vancouver, June 1989), 76-83. [4] Dai, J.C., and Lee, H.J. A generalized unification-based LR parser for Chinese. Computer Processing of Chinese and Oriental Languages, 8, 1 (1994), 1-18. [5] Dai, Y. Developing a new statistical method for Chinese text segmentation. (Master s thesis in preparation) [6] Damerau, F.J. Generating and evaluating domain-oriented multi-word terms from texts. Information Processing & Management, 29, 4 (1993), 433-447. [7] Frantzi, K.T., Ananiadou, S., and Tsujii, J. The C-value/NC -value method of automatic recognition for multi-word terms. In C. Nikolaou and C. Stephanidis (eds.), Research and Advanced Technology for Digital Libraries, 2 nd European Conference, ECDL 98 (Heraklion, Crete, September 1998), Springer-Verlag, 585-604. [8] Liang, N.Y. The knowledge of Chinese words segmentation [in Chinese]. Journal of Chinese Information Processing, 4, 2 (1990), 42-49. [9] Liu, I.M. Descriptive-unit analysis of sentences: Toward a model natural language processing. Computer Processing of Chinese & Oriental Languages, 4, 4 (1990), 314-355. [10] Liu, Y., Tan, Q., and Shen, X.K. Xin xi chu li yong xian dai han yu fen ci gui fan ji zi dong fen ci fang fa [ Modern Chinese Word Segmentation Rules and Automatic Word Segmentation Methods for Information Processing ]. Qing Hua University Press, Beijing, 1994. [11] Lua, K.T. Experiments on the use of bigram mutual information in Chinese natural language processing. Presented at the 1995 International Conference on Computer Processing of Oriental Languages (ICCPOL) (Hawaii, November 1995). Available: http://137.132.89.143/luakt/ publication.html [12] Lua, K.T. From character to word - An application of information theory. Computer Processing of Chinese & Oriental Languages, 4, 4 (1990), 304-312. [13] Lua, K.T., and Gan, G.W. An application of information theory in Chinese word segmentation. Computer Processing of Chinese & Oriental Languages, 8, 1 (1994), 115-124. [14] Nie, J.Y., Hannan, M.L., and Jin, W.Y. Unknown word detection and segmentation of Chinese using statistical and heuristic knowledge. Communications of COLIPS, 5, 1&2 (1995), 47-57. [15] Ogawa, Y., and Matsuda, T. Overlapping statistical word indexing: A new indexing method for Japanese text. In Proceedings of the 20th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (Philadelphia, July 1997), ACM, 226-234. [16] Sproat, R., and Shih, C.L. A statistical method for finding word boundaries in Chinese text. Computer Processing of Chinese & Oriental Languages, 4, 4 (1990), 336-351. [17] Sproat, R., Shih, C., Gale, W., and Chang, N. A stochastic finite-state word-segmentation algorithm for Chinese. Computational Lingustics, 22, 3 (1996), 377-404. [18] Tung, C.-H., and Lee, H.-J. Identification of unknown words from a corpus. Computer Processing of Chinese and Oriental Languages, 8 (Supplement, Dec. 1994), 131-145. [19] Wu, Z., and Tseng, G. ACTS: An automatic Chinese text segmentation system for full text retrieval. Journal of the American Society for Information Science, 46, 2 (1995), 83-96 . [20] Yeh, C.L., and Lee, H.J. Rule-based word identification for mandarin Chinese sentences: A unification approach. Computer Processing of Chinese and Oriental Languages, 5, 2 (1991), 97-118. 89

logistic regression;statistical formula;word boundary identification;Chinese text segmentation;word boundary;natural language processing;mutual information;regression model;contextual information;multi-word terms

150

Preventing Attribute Information Leakage in Automated Trust Negotiation

Automated trust negotiation is an approach which establishes trust between strangers through the bilateral, iterative disclosure of digital credentials. Sensitive credentials are protected by access control policies which may also be communicated to the other party. Ideally, sensitive information should not be known by others unless its access control policy has been satisfied. However, due to bilateral information exchange, information may flow to others in a variety of forms, many of which cannot be protected by access control policies alone. In particular, sensitive information may be inferred by observing negotiation participants' behavior even when access control policies are strictly enforced. In this paper, we propose a general framework for the safety of trust negotiation systems. Compared to the existing safety model, our framework focuses on the actual information gain during trust negotiation instead of the exchanged messages. Thus, it directly reflects the essence of safety in sensitive information protection. Based on the proposed framework, we develop policy databases as a mechanism to help prevent unauthorized information inferences during trust negotiation. We show that policy databases achieve the same protection of sensitive information as existing solutions without imposing additional complications to the interaction between negotiation participants or restricting users' autonomy in defining their own policies.

INTRODUCTION Automated trust negotiation (ATN) is an approach to access control and authentication in open, flexible systems such as the Internet. ATN enables open computing by as-Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. CCS'05, November 711, 2005, Alexandria, Virginia, USA. Copyright 2005 ACM 1-59593-226-7/05/0011 ... $ 5.00. signing an access control policy to each resource that is to be made available to entities from different domains. An access control policy describes the attributes of the entities allowed to access that resource, in contrast to the traditional approach of listing their identities. To satisfy an access control policy, a user has to demonstrate that they have the attributes named in the policy through the use of digital credentials. Since one's attributes may also be sensitive, the disclosure of digital credentials is also protected by access control policies. A trust negotiation is triggered when one party requests access to a resource owned by another party. Since each party may have policies that the other needs to satisfy, trust is established incrementally through bilateral disclosures of credentials and requests for credentials, a characteristic that distinguishes trust negotiation from other trust establishment approaches [2, 11]. Access control policies play a central role in protecting privacy during trust negotiation. Ideally, an entity's sensitive information should not be known by others unless they have satisfied the corresponding access control policy. However , depending on the way two parties interact with each other, one's private information may flow to others in various forms, which are not always controlled by access control policies. In particular, the different behaviors of a negotiation participant may be exploited to infer sensitive information , even if credentials containing that information are never directly disclosed. For example, suppose a resource's policy requires Alice to prove a sensitive attribute such as employment by the CIA. If Alice has this attribute, then she likely protects it with an access control policy. Thus, as a response, Alice will ask the resource provider to satisfy her policy. On the other hand, if Alice does not have the attribute, then a natural response would be for her to terminate the negotiation since there is no way that she can access the resource. Thus, merely from Alice's response, the resource provider may infer with high confidence whether or not Alice is working for the CIA, even though her access control policy is strictly enforced. The problem of unauthorized information flow in ATN has been noted by several groups of researchers [20, 22, 27]. A variety of approaches have been proposed, which mainly fall into two categories. Approaches in the first category try to "break" the correlation between different information. Intuitively , if the disclosed policy for an attribute is independent from the possession of the attribute, then the above inference is impossible. A representative approach in this category is by Seamons et al. [20], where an entity possessing a sensi-36 tive credential always responds with a cover policy of f alse to pretend the opposite. Only when the actual policy is satisfied by the credentials disclosed by the opponent will the entity disclose the credential. Clearly, since the disclosed policy is always f alse, it is not correlated to the possession of the credential. One obvious problem with this approach, however, is that a potentially successful negotiation may fail because an entity pretends to not have the credential. Approaches in the second category aim to make the correlation between different information "safe", i.e., when an opponent is able to infer some sensitive information through the correlation, it is already entitled to know that information . For example, Winsborough and Li [23] proposed the use of acknowledgement policies ("Ack policies" for short) as a solution. Their approach is based on the principle "discuss sensitive topics only with appropriate parties". Therefore, besides an access control policy P , Alice also associates an Ack policy P Ack with a sensitive attribute A. Intuitively, P Ack determines when Alice can tell others whether or not she has attribute A. During a negotiation, when the attribute is requested, the Ack policy P Ack is first sent back as a reply. Only when P Ack is satisfied by the other party, will Alice disclose whether or not she has A and may then ask the other party to satisfy the access control policy P . In order to prevent additional correlation introduced by Ack policies, it is required that all entities use the same Ack policy to protect a given attribute regardless of whether or not they have A. In [23], Winsborough and Li also formally defined the safety requirements in trust negotiation based on Ack policies. Though the approach of Ack policies can provide protection against unauthorized inferences, it has a significant disadvantage . One benefit of automated trust negotiation is that it gives each entity the autonomy to determine the appropriate protection for its own resources and credentials. The perceived sensitivity of possessing an attribute may be very different for different entities. For example, some may consider the possession of a certificate showing eligibility for food stamps highly sensitive, and thus would like to have a very strict Ack policy for it. Some others may not care as much, and have a less strict Ack policy, because they are more concerned with their ability to get services than their privacy. The Ack Policy system, however, requires that all entities use the same Ack policy for a given attribute, which. deprives entities of the autonomy to make their own decisions . This will inevitably be over-protective for some and under-protective for others. And either situation will result in users preferring not to participate in the system. In this paper, we first propose a general framework for safe information flow in automated trust negotiation. Compared with that proposed by Winsborough and Li, our framework focuses on modeling the actual information gain caused by information flow instead of the messages exchanged. Therefore it directly reflects the essence of safety in sensitive information protection. Based on this framework, we propose policy databases as a solution to the above problem. Policy databases not only prevent unauthorized inferences as described above but also preserve users' autonomy in deciding their own policies. In order to do this, we focus on severing the correlation between attributes and policies by introducing randomness, rather than adding additional layers or fixed policies as in the Ack Policy system. In our approach, there is a central database of policies for each possession sensitive attribute. Users who possess the attribute submit their policies to the database anonymously. Users who do not possess the attribute can then draw a policy at random from the database. The result of this process is that the distributions of policies for a given possession sensitive attribute are identical for users who have the attribute and users who do not. Thus, an opponent cannot infer whether or not users possess the attribute by looking at their policies. The rest of the paper is organized as follows. In section 2, we propose a formal definition of safety for automated trust negotiation. In section 3, we discuss the specifics of our approach, including what assumptions underlie it, how well it satisfies our safety principle, both theoretically and in practical situations, and what practical concerns to implementing it exist. Closely related work to this paper is reported in section 4. We conclude this paper in section 5 SAFETY IN TRUST NEGOTIATION In [23], Winsborough and Li put forth several definitions of safety in trust negotiation based on an underlying notion of indistinguishability. The essence of indistinguishability is that if an opponent is given the opportunity to interact with a user in two states corresponding to two different potential sets of attributes, the opponent cannot detect a difference in those sets of attributes based on the messages sent. In the definition of deterministic indistinguishability, the messages sent in the two states must be precisely the same. In the definition of probabilistic indistinguishability, they must have precisely the same distribution. These definitions, however, are overly strict. To determine whether or not a given user has a credential, it is not sufficient for an opponent to know that the user acts differently depending on whether or not that user has the credential: the opponent also has to be able to figure out which behavior corresponds to having the credential and which corresponds to lacking the credential. Otherwise, the opponent has not actually gained any information about the user. Example 1. Suppose we have a system in which there is only one attribute and two policies, p 1 and p 2 . Half of the users use p 1 when they have the attribute and p 2 when they do not. The other half of the users use p 2 when they have the attribute and p 1 when they do not. Every user's messages would be distinguishable under the definition of indistinguishability presented in [23] because for each user the distribution of messages is different. However, if a fraction r of the users have the attribute and a fraction 1 - r do not, then 1 2 r + 1 2 (1 - r) = 1 2 of the users display policy p 1 and the other half of the users display policy p 2 . As such the number of users displaying p 1 or p 2 does not change as r changes. Hence, they are independent. Since the policy displayed is independent of the attribute when users are viewed as a whole, seeing either policy does not reveal any information about whether or not the user in question has the attribute. As such, Winsborough and Li's definitions of indistinguishability restrict a number of valid systems where a given user will act differently in the two cases, but an opponent cannot actually distinguish which case is which. In fact, their definitions allow only systems greatly similar to the Ack Policy system that they proposed in [22]. Instead we propose a definition of safety based directly on information 37 gain instead of the message exchange sequences between the two parties. Before we formally define safety, we first discuss what safety means informally. In any trust negotiation system, there is some set of objects which are protected by policies . Usually this includes credentials, information about attribute possession, and sometimes even some of the policies in the system. All of these can be structured as digital information, and the aim of the system is to disclose that information only to appropriate parties. The straight-forward part of the idea of safety is that an object's value should not be revealed unless its policy has been satisfied. However, we do not want to simply avoid an object's value being known with complete certainty, but also the value being guessed with significant likelihood. As such, we can define the change in safety as the change in the probability of guessing the value of an object. If there are two secrets, s 1 and s 2 , we can define the conditional safety of s 1 upon the disclosure of s 2 as the conditional probability of guessing s 1 given s 2 . Thus, we define absolute safety in a system as being the property that no disclosure of objects whose policies have been satisfied results in any change in the probability of guessing the value of any object whose policy has not been satisfied regardless of what inferences might be possible. There exists a simple system which can satisfy this level of safety, which is the all-or-nothing system, a system in which all of every user's objects are required to be protected by a single policy which is the same for all users. Clearly in such a system there are only two states, all objects revealed or no objects revealed. As such, there can be no inferences between objects which are revealed and objects which are not. This system, however, has undesirable properties which outweigh its safety guarantees, namely the lack of autonomy, flexibility, and fine-grained access control. Because of the necessity of protecting against every possible inference which could occur, it is like that any system which achieves ideal safety would be similarly inflexible. Since there have been no practical systems proposed which meet the ideal safety condition, describing ideal safety is not sufficient unto itself. We wish to explore not just ideal safety, but also safety relative to certain types of attacks. This will help us develop a more complete view of safety in the likely event that no useful system which is ideally safe is found. If a system does not have ideal safety, then there must be some inferences which can cause a leakage of information between revealed objects and protected objects. But this does not mean that every single object revealed leaks information about every protected object. As such, we can potentially describe what sort of inferences a system does protect against. For example, Ack Policy systems are moti-vated by a desire to prevent inferences from a policy to the possession of the attribute that it protects. Inferences from one attribute to another are not prevented by such a system (for example, users who are AARP members are more likely to be retired than ones who are not). Hence, it is desirable to describe what it means for a system to be safe relative to certain types of inferences. Next we present a formal framework to model safety in trust negotiation. The formalism which we are using in this paper is based on that used by Winsborough and Li, but is substantially revised. 2.0.1 Trust Negotiation Systems A Trust Negotiation System is comprised of the following elements: A finite set, K, of principals, uniquely identified by a randomly chosen public key, P ub k . Each principal knows the associated private key, and can produce a proof of identity. A finite set, T , of attributes. An attribute is something which each user either possesses or lacks. An example would be status as a licensed driver or enrollment at a university. A set, G, of configurations, each of which is a subset of T . If a principal k is in a configuration g G, then k possesses the attributes in g and no other attributes. A set, P, of possible policies, each of which is a logical proposition comprised of a combination of and, or, and attributes in T . We define an attribute in a policy to be true with respect to a principal k if k has that attribute. We consider all logically equivalent policies to be the same policy. Objects. Every principal k has objects which may be protected which include the following: - A set, S, of services provided by a principal. Every principal offers some set of services to all other principals. These services are each protected by some policy, as we will describe later. A simple service which each principal offers is a proof of attribute possession. If another principal satisfies the appropriate policy, the principal will offer some proof that he holds the attribute. This service is denoted s t for any attribute t T . - A set, A, of attribute status objects. Since the set of all attributes is already known, we want to protect the information about whether or not a given user has an attribute. As such we formally define A as a set of boolean valued random variables, a t . The value of a t for a principal k, which we denote a t (k) is defined to be true if k possesses t T and false otherwise. Thus A = {a t |t T }. - A set, Q of policy mapping objects. A system may desire to protect an object's policy either because of correlations between policies and sensitive attributes or because in some systems the policies themselves may be considered sensitive. Similar to attribute status objects, we do not protect a policy , per se, but instead the pairing of a policy with what it is protecting. As such, each policy mapping object is a random variable q o with range P where o is an object. The value of q o for a given principal k, denoted q o (k) is the policy that k has chosen to protect object o. Every system should define which objects are protected. It is expected that all systems should protect the services, S, and the attribute status objects, A. In some systems, there will also be policies which protect policies. Thus protected objects may also include a subset of Q. We call the set of protected objects O, where O S A Q. If an object is not protected, this is equivalent to it having a policy equal to true. For convenience, we define Q X to be the members of Q which are policies protecting members of X , where X is a set of objects. Formally, Q X = {q o Q|o X }. Some subset of the information objects are considered to be sensitive objects. These are the objects about which we want an opponent to gain no information unless they have satisfied that object's policy. Full information about any object, sensitive or insensitive, is not released by the system 38 until its policy has been satisfied, but it is acceptable for inferences to cause the leakage of information which is not considered sensitive. A set, N , of negotiation strategies. A negotiation strategy is the means that a principal uses to interact with other principals. Established strategies include the eager strategy [24] and the trust-target graph strategy [22]. A negotiation strategy, n, is defined as an interactive, deterministic, Turing-equivalent computational machine augmented by a random tape. The random tape serves as a random oracle which allows us to discuss randomized strategies. A negotiation strategy takes as initial input the public knowledge needed to operate in a system, the principal's attributes, its services, and the policies which protect its objects. It then produces additional inputs and outputs by interacting with other strategies. It can output policies, credentials, and any additional information which is useful. We do not further define the specifics of the information communicated between strategies except to note that all the strategies in a system should have compatible input and output protocols. We refrain from further specifics of strategies since they are not required in our later discussion. An adversary, M , is defined as a set of principals coordinating to discover the value of sensitive information objects belonging to some k M . Preventing this discovery is the security goal of a trust negotiation system. We assume that adversaries may only interact with principals through trust negotiation and are limited to proving possession of attributes which they actually possess. In other words, the trust negotiation system provides a means of proof which is resistant to attempts at forgery. A set, I, of all inferences. Each inference is a minimal subset of information objects such that the joint distribution of the set differs from the product of the individual distributions of the items in the set. 1 These then allow a partitioning, C, of the information objects into inference components. We define a relation such that o 1 o 2 iff i I|o 1 , o 2 i. C is the transitive closure of . In general, we assume that all of the information objects in our framework are static. We do not model changes in a principal's attribute status or policies. If such is necessary, the model would need to be adapted. It should also be noted that there is an additional constraint on policies that protect policies which we have not described. This is because in most systems there is a way to gain information about what a policy is, which is to satisfy it. When a policy is satisfied, this generally results in some service being rendered or information being released. As such, this will let the other party know that they have satisfied the policy for that object. Therefore, the effective policy protecting a policy status object must be the logical or of the policy in the policy status object and the policy which protects it. 2.0.2 The Ack Policy System To help illustrate the model, let us describe how the Ack Policy system maps onto the model. The mapping of oppo-1 A system need not define the particulars of inferences, but should discuss what sort of inferences it can deal with, and hence what sort of inferences are assumed to exist. nents, the sets of principals, attributes, configurations, and policies in the Ack Policy system is straightforward. In an Ack Policy system, any mutually compatible set of negotiation strategies is acceptable. There are policies for protecting services, protecting attribute status objects, and protecting policies which protect attribute proving services. As such, the set of protected objects, O = S A Q S . According to the definition of the Ack Policy system, for a given attribute, the policy that protects the proof service for that attribute is protected by the same policy that protects the attribute status object. Formally, t T , k K, q a t (k) = q q st (k). Further, the Ack policy for an attribute is required to be the same for all principals. Thus we know t T p Pk K|q a t (k) = p. Two basic assumptions about the set of inferences, I, exist in Ack Policy systems, which also lead us to conclusions about the inference components, C. It is assumed that inferences between the policy which protects the attribute proving service, q s t (k), and the attribute status object, a t (k), exist. As such, those two objects should always be in the same inference component. Because Ack Policies are uniform for all principals, they are uncorrelated to any other information object and they cannot be part of any inference. Hence, each Ack Policy is in an inference component of its own. 2.0.3 Safety in Trust Negotiation Systems In order to formally define safety in trust negotiation, we need to define the specifics of the opponent. We need to model the potential capabilities of an opponent and the information initially available to the opponent. Obviously, no system is safe against an opponent with unlimited capabilities or unlimited knowledge. As such, we restrict the opponent to having some tactic, for forming trust negotiation messages, processing responses to those messages, and, finally, forming a guess about the values of unrevealed information objects. We model the tactic as an interactive, deterministic, Turing-equivalent computational machine. This model is a very powerful model, and we argue that it describes any reasonable opponent. This model, however, restricts the opponent to calculating things which are computable from its input and implies that the opponent behaves in a deterministic fashion. The input available to the machine at the start is the knowledge available to the opponent before any trust negotiation has taken place. What this knowledge is varies depending on the particulars of a trust negotiation system. However, in every system this should include the knowledge available to the principals who are a part of the opponent , such as their public and private keys and their credentials . And it should also include public information such as how the system works, the public keys of the attribute authorities, and other information that every user knows. In most systems, information about the distribution of attributes and credentials and knowledge of inference rules should also be considered as public information. All responses from principals in different configurations become available as input to the tactic as they are made. The tactic must output both a sequence of responses and, at the end, guesses about the unknown objects of all users. We observe that an opponent will have probabilistic knowledge about information objects in a system. Initially, the probabilities will be based only on publicly available knowl-39 edge, so we can use the publicly available knowledge to describe the a priori probabilities. For instance, in most systems, it would be reasonable to assume that the opponent will have knowledge of the odds that any particular member of the population has a given attribute . Thus, if a fraction h t of the population is expected to possess attribute t T , the opponent should begin with an assumption that some given principal has a h t chance of having attribute t. Hence, h t represents the a priori probability of any given principal possessing t. Note that we assume that the opponent only knows the odds of a given principal having an attribute, but does not know for certain that a fixed percentage of the users have a given attribute. As such, knowledge about the value of an object belonging to some set of users does not imply any knowledge about the value of objects belonging to some other user. Definition 1. A trust negotiation system is safe relative to a set of possible inferences if for all allowed mappings between principals and configurations there exists no opponent which can guess the value of sensitive information objects whose security policies have not been met with odds better than the a priori odds over all principals which are not in the opponent, over all values of all random tapes, and over all mappings between public key values and principals. Definition 1 differs from Winsborough and Li's definitions in several ways. The first is that it is concerned with multiple users. It both requires that the opponent form guesses over all users and allows the opponent to interact with all users. Instead of simply having a sequence of messages sent to a single principal, the tactic we have defined may interact with a variety of users, analyzing incoming messages, and then use them to form new messages. It is allowed to talk to the users in any order and to interleave communications with multiple users, thus it is more general than those in [23]. The second is that we are concerned only with the information which the opponent can glean from the communication, not the distribution of the communication itself. As such, our definition more clearly reflects the fundamental idea of safety. We next introduce a theorem which will be helpful in proving the safety of systems. Theorem 1. There exists no opponent which can beat the a priori odds of guessing the value of an object, o, given only information about objects which are not in the same inference component as o, over all principals not in M and whose policy for o M cannot satisfy, over all random tapes, and over all mappings between public keys and principals. The formal proof for this theorem can be found in Appendix A. Intuitively, since the opponent only gains information about objects not correlated to o, its guess of the value of o is not affected. With theorem 1, let us take a brief moment to prove the safety of the Ack Policy systems under our framework. Specifically, we examine Ack Policy systems in which the distribution of strategies is independent of the distributions of attributes, an assumption implicitly made in [23]. In Ack Policy systems the Ack Policy is a policy which protects two objects in our model: an attribute's status object and its policy for that attribute's proof service. Ack Policies are required to be uniform for all users, which ensures that they are independent of all objects. Ack Policy systems are designed to prevent inferences from an attribute's policy to an attribute's status for attributes which are sensitive. So, let us assume an appropriate set of inference components in order to prove that Ack Policy systems are safe relative to that form of inference. As we described earlier, each attribute status object should be in the same inference component with the policy which protects that attribute's proof service, and the Ack policy for each attribute should be in its own inference component. The Ack Policy system also assumes that different attributes are independent of each other. As such, each attribute status object should be in a different inference group. This set of inference components excludes all other possible types of inferences. The set of sensitive objects is the set of attribute status objects whose value is true. Due to Theorem 1, we know then that no opponent will be able to gain any information based on objects in different inference components. So the only potential source of inference for whether or not a given attribute's status object, a t , has a value of true or f alse is the policy protecting the attribute proof service, s t . However, we know that the same policy, P , protects both of these objects. As such unauthorized inference between them is impossible without satisfying P . 2 Thus, the odds for a t do not change. Therefore, the Ack Policy system is secure against inferences from an attribute's policy to its attribute status. POLICY DATABASE We propose a new trust negotiation system designed to be safe under the definition we proposed, but to also allow the users who have sensitive attributes complete freedom to determine their own policies. It also does not rely on any particular strategy being used. Potentially, a combination of strategies could even be used so long as the strategy chosen is not in any way correlated to the attributes possessed. This system is based on the observation that there is more than one way to deal with a correlation. A simple ideal system which prevents the inference from policies to attribute possession information is to have each user draw a random policy. This system obviously does not allow users the freedom to create their own policies. Instead we propose a system which looks like the policies are random even though they are not. This system is similar to the existing trust negotiation systems except for the addition of a new element: the policy database. The policy database is a database run by a trusted third party which collects anonymized information about the policies which are in use. In the policy database system, a 2 Except that one of these is a policy mapping object which is being protected by a policy. As such, we have to keep in mind that there exists a possibility that the opponent could gain information about the policy without satisfying it. Specifically, the opponent can figure out what attributes do not satisfy it by proving that he possesses those attributes. However, in an Ack Policy system, the policy protecting an attribute proof object of an attribute which a user does not hold is always f alse. No opponent can distinguish between two policies which they cannot satisfy since all they know is that they have failed to satisfy them. And we are unconcerned with policies which they have satisfied. Thus, we know that the opponent cannot gain any useful information about the policies which they have not satisfied, and hence cannot beat the a priori odds for those policies. 40 user who has a given sensitive attribute chooses their own policy and submits it anonymously to the policy database for that attribute. The policy database uses pseudonymous certificates to verify that users who submit policies actually have the attribute, in a manner that will be discussed later in section 3.2. Then users who do not have the attribute will pull policies at random from the database to use as their own. The contents of the policy database are public, so any user who wishes to can draw a random policy from the database. In our system, each user uses a single policy to protect all the information objects associated with an attribute. They neither acknowledge that they have the attribute nor prove that they do until the policy has been satisfied. This means that users are allowed to have policies which protect attributes which they do not hold. The policy in our system may be seen as the combination of the Ack policy and a traditional access control policy for attribute proofs. The goal of this system is to ensure that the policy is in a separate inference component from the attribute status object, thus guaranteeing that inferences between policies and attribute status objects cannot be made. This system is workable because of the following. We know that policies cannot require a lack of an attribute, thus users who do not have a given attribute will never suffer from their policy for that attribute being too strong. Changes in the policy which protects an attribute that they do not have may vary the length of the trust negotiation, but it will never cause them to be unable to complete transactions which they would otherwise be able to complete. Also, we deal only with possession sensitive attributes. We do not deal with attributes where it is at all sensitive to lack them. As such, users who do not have the attribute cannot have their policies be too weak. Since there is no penalty for those users for their policies being either too weak or too strong, they can have whatever policy is most helpful for helping disguise the users who do possess the attribute. This also means that users who do not have the attribute do not need to trust the policy database since no policy which the database gives them would be unacceptable to them. Users who have the attribute, however, do need to trust that the policy database will actually randomly distribute policies to help camouflage their policies. They do not, however, need to trust the policy database to act appropriately with their sensitive information because all information is anonymized. 3.1 Safety of the Approach of Policy Databases Let us describe the Policy Database system in terms of our model. Again the opponent and the sets of principals, attributes, configurations, and policies need no special comment . Because we only have policies protecting the services and attribute status objects, the set of protected objects , O = S A. Also, each attribute proving service and attribute status object are protected by the same policy. t T , k K, q a t (k) = q s t (k). This system is only designed to deal with inferences from policies to attribute possession, so we assume that every attribute status object is in a different inference component. If the policies do actually appear to be completely random, then policies and attribute status objects should be in separate inference components as well. The obvious question is whether Policy Database systems actually guarantee that this occurs. The answer is that they do not guarantee it with any finite number of users due to the distribution of policies being unlikely to be absolutely, precisely the same. This is largely due to a combination of rounding issues and the odds being against things coming out precisely evenly distributed. However, as the number of users in the system approaches infinity, the system approaches this condition. In an ideal system, the distribution of policies would be completely random. If an opponent observes that some number of principals had a given policy for some attribute, this would give them no information about whether or not any of those users had the attribute. However, in the Policy Database system, every policy which is held is known to be held by at least one user who has the attribute. As such, we need to worry about how even the distributions of different policies are. We can describe and quantify the difference which exists between a real implementation of our system and the ideal. There are two reasons for a difference to exist. The first is difference due to distributions being discrete. For example, let us say that there are five users in our system, two of which have some attribute and three who do not. Let us also say that the two users with the attribute each have different policies. For the distributions to be identical, each of those policies would need to be selected by one and a half of the remaining three users. This, obviously, cannot happen. We refer to this difference as rounding error. The second is difference due to the natural unevenness of random selection. The distributions tend towards evenness as the number of samples increases, but with any finite number of users, the distributions are quite likely to vary some from the ideal. These differences can both be quantified the same way: as a difference between the expected number of principals who have a policy and the actual number. If the opponent knows that one half of the principals have an attribute and one half do not, and they observe that among four users, there are two policies, one of which is held by three users and the other by one user, then they can know that the user with the unique policy holds the attribute. In general, any time the number of users who share a policy is less than the expectation, it is more likely that a user who has that policy also has the attribute. Information is leaked when there is a difference between the expected number of principals who have a policy and the actual number of principals who have that policy in proportion to the ratio between them. Theorem 2. The limit of the difference between the expected number of principals who have a policy and the actual number of principals who have the policy as the number of users goes to infinity is 0. The proof of Theorem 2 can be found in Appendix B. The intuition behind it is that as the number of samples grows very large, the actual distribution approaches the ideal distribution and the rounding errors shrink towards zero. 3.2 Attacks and Countermeasures Until now, we have only proven things about a system which is assumed to be in some instantaneous unchanging state. In the real world we have to deal with issues related to how policies change over time and multiple interactions. 41 Therefore, we also want the policy which a given user randomly selects from the database to be persistent. Otherwise an adversary would simply make multiple requests to the same user over time and see if the policy changed. If it did, especially if it changed erratically, it would indicate that the user was repeatedly drawing random policies. Instead, the user should hold some value which designates which policy the user has. An obvious answer would be to have the user hold onto the policy itself, but this would open the user up to a new attack . If users lacking a given attribute simply grabbed onto a policy and never changed it, this itself could be a tell. If there were some event which occurred which made having a given attribute suddenly more sensitive than it used to be, then rational users who have the attribute would increase the stringency of their policies. For example, if a country undertook an action which was politically unpopular on a global scale, holders of passports issued by that country would likely consider that more sensitive information now and would increase their policies appropriately. The result would then be that the average policy for people who had cached a previously fetched policy would then be less stringent than those who were making their own policies. Instead of a permanent policy, it would be more sensible for a principal to receive a cookie which could get it the policy from a particular principal so that when principals who posses the attribute changed their policies, principals who do not possess it would too. We also need to guard against stacking the deck. Obviously we can restrict the database to users who actually have the attribute by requiring the presentation of a pseudonymous certificate [6, 7, 8, 9, 10, 18] which proves that they have the attribute. However, we also need to assure that a legitimate attribute holder cannot submit multiple policies in order to skew the set of policies. To this end, we require that each policy be submitted initially with a one-time-show pseudonymous credential [8]. The attribute authorities can be restricted so that they will only issue each user a single one-time-show pseudonymous credential for each Policy Database use. Then we can accept the policy, knowing it to come from a unique user who holds the attribute, and issue them a secret key which they can later use to verify that they were the submitter of a given policy and to replace it with an updated policy. This does not prevent a user who has the attribute from submitting a single false policy, perhaps one which is distinctly different from normal policies. The result would be that users who draw that policy would be known to not have the attribute. However, under the assumptions of our system , not having the attribute is not sensitive, so this does not compromise safety. 3.3 Limitations We assume that for the attribute being protected, it is not significantly sensitive to lack the attribute. This assumption means that our system likely cannot be used in practice to protect all attributes. Most notably it fails when lacking an attribute implies having or being highly likely to have some other attribute. For example, not having a valid passport probably means that you are a permanent resident of the country you are currently in (although users could be an illegal immigrants or citizens of a defunct nation). It also fails when the lack of an attribute is more sensitive than having it. For instance, few people are going to wish to prevent people from knowing that they have graduated from high school, but many would consider their lack of a high school graduation attribute to be sensitive. However, we argue that no system can adequately handle such a case because those who do have the attribute would likely be unwilling to accept any system which would result in them having to not disclose the attribute when it was useful for them to do so. And if they still easily disclose their attribute, then it becomes impossible for those without to disguise their lack. Similarly to the Ack Policy system, policy databases also do not generally handle any form of probabilistic inference rule between attributes. The existence of such a rule would likely imply certain relationships between policies which most users would enforce. If the possession of a city library card suggested with strong probability that the user was a city resident, then perhaps all users who have both would have a policy protecting their library card which is stricter than the policy protecting their city residency. However, as there is variety in the policies of individuals, a user could pick a random pair of policies which did not have this property. That would then be a sure tell that he did not actually have both of those attributes. Another drawback of the system is that it requires a policy database service be available on-line. This decreases the decentralized nature of trust negotiation. However, our approach is still less centralized than Ack Policies, which require that users cooperate to determine a universally accepted Ack policy. And this centralization may be able to be decreased by decentralizing the database itself. Although we discuss the database as if it were a single monolithic entity, it could be made of a number of different entities acting together. The only requirement is that it accepts policies from unique users who have the attribute and distributes them randomly. RELATED WORK The framework of automated trust negotiation was first proposed by Winsborough et al. [24]. Since then, great efforts have been put forward to address challenges in a variety of aspects of trust negotiation. An introduction to trust negotiation and related trust management issues can be found in [25]. As described in detail there, a number of trust negotiation systems and supporting middleware have been proposed and/or implemented in a variety of contexts (e.g., [3, 4, 11, 12, 14, 17, 19]). Information leakage during trust negotiation is studied in [13, 5, 15, 20, 21, 22, 23]. The work by Winsborough and Li has been discussed in detail in previous sections. Next, we discuss several other approaches. In [20], non-response is proposed as a way to protect possession-sensitive attributes. The basic idea is to have Alice, the owner of a sensitive attribute, act as if she does not have the attribute. Only later when the other party accidentally satisfies her policy for that attribute will Alice disclose that attribute. This approach is easy to deploy in trust negotiation. But clearly it will often cause a potentially successful negotiation to fail because of Alice's conservative response. Yu and Winslett [26] introduce a technique called policy migration to mitigate the problem of unauthorized inference. In policy migration, Alice dynamically integrates her poli-42 cies for sensitive attributes with those of other attributes, so that she does not need to explicitly disclose policies for sensitive attributes. Meanwhile, policy migration makes sure that "migrated" policies are logically equivalent to original policies, and thus guarantees the success of the negotiation whenever possible. On the other hand, policy migration is not a universal solution, in the sense that it may not be applicable to all the possible configurations of a negotiation. Further, it is subject to a variety of attacks. In other words, it only seeks to make unauthorized inference harder instead of preventing it completely. Most existing trust negotiation frameworks [16, 17, 28] assume that the appropriate access control policies can be shown to Bob when he requests access to Alice's resource. However, realistic access control policies also tend to contain sensitive information, because the details of Alice's policy for the disclosure of a credential C tends to give hints about C's contents. More generally, a company's internal and external policies are part of its corporate assets, and it will not wish to indiscriminately broadcast its policies in their entirety. Several schemes have been proposed to protect the disclosure of sensitive policies. In [4], Bonatti and Samarati suggests dividing a policy into two parts prerequisite rules and requisite rules. The constraints in a requisite rule will not be disclosed until those in prerequisite rules are satisfied. In [19], Seamons et al. proposed organizing a policy into a directed graph so that constraints in a policy can be disclosed gradually. In [26], access control policies are treated as first-class resources, thus can be protected in the same manner as services and credentials. Recently, much work has been done on mutual authentication and authorization through the use of cryptographic techniques that offer improved privacy guarantees. For example , Balfanz et al. [1] designed a secret-handshake scheme where two parties reveal their memberships in a group to each other if and only if they belong to the same group. Li et al. [15] proposed a mutual signature verification scheme to solve the problem of cyclic policy interdependency in trust negotiation. Under their scheme, Alice can see the content of Bob's credential signed by a certification authority CA only if she herself has a valid certificate also signed by CA and containing the content she sent to Bob earlier. A similar idea was independently explored by researchers [5, 13] to handle more complex access control policies. Note that approaches based on cryptographic techniques usually impose more constraints on access control policies. Therefore, policy databases are complementary to the above work. CONCLUSION AND FUTURE WORK In this paper, we have proposed a general framework for safety in automated trust negotiation. The framework is based strictly on information gain, instead of on communication . It thus more directly reflects the essence of safe information flow in trust negotiation. We have also shown that Ack policy systems are safe under our framework. Based on the framework, we have presented policy databases, a new, safe trust negotiation system. Compared with existing systems, policy databases do not introduce extra layers of policies or other complications to the negotiation between users. Further, policy databases preserve user's autonomy in defining their own policies instead of imposing uniform policies across all users. Therefore they are more flexible and easier to deploy than other systems. Further, we have discussed a number of practical issues which would be involved in implementing our system. In the future, we plan to address how our system can be used in the presence of delegated credentials. And we plan to attempt to broaden the system to account for probabilistic inferences rules which are publicly known. Acknowledgments This research was sponsored by NSF through IIS CyberTrust grant number 0430166 (NCSU). We also thank anonymous reviewers for their helpful comments. REFERENCES [1] D. Balfanz, G. Durfee, N. Shankar, D. Smetters, J. Staddon, and H. Wong. Secret Handshakes from Pairing-Based Key Agreements. In IEEE Symposium on Security and Privacy, Berkeley, CA, May 2003. [2] M. Blaze, J. Feigenbaum, J. Ioannidis, and A. Keromytis. The KeyNote Trust Management System Version 2. In Internet Draft RFC 2704, September 1999. [3] M. Blaze, J. Feigenbaum, and A. D. Keromytis. KeyNote: Trust Management for Public-Key Infrastructures. In Security Protocols Workshop, Cambridge, UK, 1998. [4] P. Bonatti and P. Samarati. Regulating Service Access and Information Release on the Web. In Conference on Computer and Communications Security, Athens, November 2000. [5] R.W. Bradshaw, J.E. Holt, and K.E. Seamons. Concealing Complex Policies in Hidden Credentials. In ACM Conference on Computer and Communications Security, Washington, DC, October 2004. [6] S. Brands. Rethinking Public Key Infrastructures and Digital Certificates: Building in Privacy. The MIT Press, 2000. [7] J. Camenisch and E.V. Herreweghen. Design and Implementation of the Idemix Anonymous Credential System. In ACM Conference on Computer and Communications Security, Washington D.C., November 2002. [8] J. Camenisch and A. Lysyanskaya. Efficient Non-Transferable Anonymous Multi-Show Credential System with Optional Anonymity Revocation. In EUROCRYPT 2001, volume 2045 of Lecture Notes in Computer Science. Springer, 2001. [9] D. Chaum. Security without Identification: Transactions Systems to Make Big Brother Obsolete. Communications of the ACM, 24(2), 1985. [10] I.B. Damg ard. Payment Systems and Credential Mechanism with Provable Security Against Abuse by Individuals. In CRYPTO'88, volume 403 of Lecture Notes in Computer Science. Springer, 1990. [11] A. Herzberg, J. Mihaeli, Y. Mass, D. Naor, and Y. Ravid. Access Control Meets Public Key Infrastructure, Or: Assigning Roles to Strangers. In IEEE Symposium on Security and Privacy, Oakland, CA, May 2000. [12] A. Hess, J. Jacobson, H. Mills, R. Wamsley, K. Seamons, and B. Smith. Advanced Client/Server Authentication in TLS. In Network and Distributed System Security Symposium, San Diego, CA, February 2002. [13] J. Holt, R. bradshaw, K.E. Seamons, and H. Orman. Hidden Credentials. In ACM Workshop on Privacy in the Electronic Society, Washington, DC, October 2003. [14] W. Johnson, S. Mudumbai, and M. Thompson. Authorization and Attribute Certificates for Widely Distributed Access Control. In IEEE International Workshop on Enabling Technologies: Infrastructure for Collaborative Enterprises, 1998. [15] N. Li, W. Du, and D. Boneh. Oblivious Signature-Based Envelope. In ACM Symposium on Principles of Distributed Computing, New York City, NY, July 2003. [16] N. Li, J.C. Mitchell, and W. Winsborough. Design of a Role-based Trust-management Framework. In IEEE Symposium on Security and Privacy, Berkeley, California, May 2002. [17] N. Li, W. Winsborough, and J.C. Mitchell. Distributed Credential Chain Discovery in Trust Management. Journal of Computer Security, 11(1), February 2003. [18] A. Lysyanskaya, R. Rivest, A. Sahai, and S. Wolf. Pseudonym Systems. In Selected Areas in Cryptography, 1999, volume 1758 of Lecture Notes in Computer Science. Springer, 2000. [19] K. Seamons, M. Winslett, and T. Yu. Limiting the Disclosure of Access Control Policies during Automated Trust 43 Negotiation. In Network and Distributed System Security Symposium, San Diego, CA, February 2001. [20] K. Seamons, M. Winslett, T. Yu, L. Yu, and R. Jarvis. Protecting Privacy during On-line Trust Negotiation. In 2nd Workshop on Privacy Enhancing Technologies, San Francisco, CA, April 2002. [21] W. Winsborough and N. Li. Protecting Sensitive Attributes in Automated Trust Negotiation. In ACM Workshop on Privacy in the Electronic Society, Washington, DC, November 2002. [22] W. Winsborough and N. Li. Towards Practical Automated Trust Negotiation. In 3rd International Workshop on Policies for Distributed Systems and Networks, Monterey, California, June 2002. [23] W. Winsborough and N. Li. Safety in Automated Trust Negotiation. In IEEE Symposium on Security and Privacy, Oakland, CA, May 2004. [24] W. Winsborough, K. Seamons, and V. Jones. Automated Trust Negotiation. In DARPA Information Survivability Conference and Exposition, Hilton Head Island, SC, January 2000. [25] M. Winslett, T. Yu, K.E. Seamons, A. Hess, J. Jarvis, B. Smith, and L. Yu. Negotiating Trust on the Web. IEEE Internet Computing, special issue on trust management, 6(6), November 2002. [26] T. Yu and M. Winslett. A Unified Scheme for Resource Protection in Automated Trust Negotiation. In IEEE Symposium on Security and Privacy, Oakland, CA, May 2003. [27] T. Yu and M. Winslett. Policy Migration for Sensitive Credentials in Trust Negotiation. In ACM Workshop on Privacy in the Electronic Society, Washington, DC, October 2003. [28] T. Yu, M. Winslett, and K. Seamons. Supporting Structured Credentials and Sensitive Policies through Interoperable Strategies in Automated Trust Negotiation. ACM Transactions on Information and System Security, 6(1), February 2003. APPENDIX A. PROOF OF THEOREM 1 Our goal is to prove the following theorem: There exists no opponent which can beat the a priori odds of guessing the value of an object, o given only information about objects which are not in the same inference component as o, over all principals not in M and whose policy for o M cannot satisfy, over all random tapes, and over all mappings of public key values to principals. Now it follows that if the opponent can beat the a priori odds of guessing the value of an object, o, then the opponent can beat the a priori odds of guessing the parity of o. Hence, if no opponent can beat the a priori odds of guessing the parity of an object, then none can beat the odds of guessing the value of the object. Lemma 1. There exists no opponent which can beat the a priori odds of guessing the parity of an object, o given only information about objects which are not in the same inference component as o, over all principals not in M and whose policy for o M cannot satisfy, over all random tapes, and over all mappings of public key values to principals. To prove this, we begin with an assumption that there exists some tactic which can successfully guess the parity of o with odds better than the a priori odds for at least some public key mappings. We are going to prove that any such tactic cannot beat the a priori odds on average across all mappings because there must be more mappings where it fails to beat the a priori odds than where it beats them. Just to be clear, the tactic is allowed to interact with principals whose policy for o it can satisfy. It just does not get to guess about the value of o for those principals, as it is entitled to beat the a priori odds for them. Hence, doing so is not considered a leakage in the system. Because the tactic is a deterministic Turing-equivalent computational machine, when it outputs its final guesses, it must output them in some order. We will define n to be the number of users, |K|. We will number the series of principals k 1 , k 2 , ..., k n . Without loss of generality, we can assume that every principal's strategy's random tape has some fixed value, resulting in them behaving in a strictly deterministic manner. Therefore, as the tactic and strategies are deterministic, the only remaining variable is the mapping of public keys to principals. Next we will fix the sequence of public keys. Because public keys are randomly chosen to begin with, and we are varying over the set of all public-key to user mappings, we can do this without loss of generality. The order in which guesses are made must in some way depend only on the a priori knowledge, the public keys, and the communications which the tactic has with the strategies. So, if all of these things are kept constant, the guesses will not change. Let us suppose that a fraction h of the population whose policy for o has not been satisfied has one parity value, and a fraction 1 - h of the population has the other. Without loss of generality, we assume that h 1 - h. We determine h by calculating the relative a priori probabilities given the distribution of the values of the object. The a priori probability of successfully guessing which parity a given user's object has is h. Now, if there exists some order of interaction, i which beats the a priori odds, then its number of correct guesses must be expressible as hn + for some &gt; 0. We can break the set of users whose policies for o M cannot meet down into a group of sets according to the values of the objects which are in inference components other than the one which contains o. We will define a set of sets, V G such that vg V G is a set of users all of which have the same values for all objects in all inference components other than the one which contains o. Now, let us consider the possibility of rearranging the public keys of members of this group. Because the strategies in use are defined to be deterministic with respect to the policies governing the attributes which distinguish the two configurations and because the opponent is defined to be deterministic: it follows that if we were to rearrange user's public keys from the original mapping to create a new mapping , the communication would be the same in both. Since the communication would be the same, it follows that the tactic would make the same guesses relative to the order of users because it is a deterministic machine and must produce the same output given the same input, the end result of which is that switching two users both of whom are members of the same value group will result in the guesses of the parity of those two users switching as well. We can then consider the set of all arrangements of public keys formed by switching principals around within their value groups, which we shall call I. So the question at hand, then, is whether or not the expected value of across all members of I is positive. If we can demonstrate that it is not, then no successful opponent can exist. Here we introduce another lemma. Proof of this lemma is now sufficient to establish our earlier lemma. Lemma 2. The expected value of across all public key mappings is less than or equal to zero. 44 If we have some quantity of extra correct guesses, , for some public key mapping i, then these guesses must be distributed over some set of value groups. If is to be positive on average, then at least some value groups must average a number of correct guesses above the a priori probability over all arrangements in I. Let us assume that we have one such group vg. Because the distributions of values of items in other inference components are defined to be precisely independent of o, we can know that in each group, there must be a fraction h of the members which have one parity and 1 - h which have the other. So, in vg there will be x = h|vg| principals with the first parity and y = (1 - h)|vg| principals with the second, and the a priori expected number of correct guesses would be x. If, for some mapping, i, the tactic is successful, then there must be some number of correct guesses x + where &gt; 0. We also know that y simply because the tactic is limited in total correct guesses to |vg| = x + y. As the number of correct guesses is x + , it must follow that the number of incorrect guesses is y - . Further, we need to note that the tactic must make some quantity of first parity guesses and some quantity of second parity guesses. Obviously, these quantities need to add up to |vg|, but need not match up with x and y. Every extra first parity or second parity guess guarantees at least one mistake, but even with several mistakes, it is quite possible to beat the a priori odds for some arrangements. So we define x + c to be the number of first parity guesses and y - c to be the number of second parity guesses. Now, we know that each increase of one in |c| guarantees at least one wrong guess, so we have a bound of + |c| y. Further, we know that since c is fixed (as it is not dependent on the arrangement, only the guesses which are unchanging ), the only way to gain a wrong guess is to swap a first parity principal with a second parity principal, which must necessarily create two wrong guesses. So we can quantify the number of wrong first parity guesses and the number of wrong second parity guesses using the terms we have set up. Specifically, there must be 1 2 (y - + c) incorrect first parity guesses, and 1 2 (y - - c) incorrect second parity guesses. Now we can determine the number of arrangements of principals which will create x + correct guesses. Specifically , we look at the total number of principals which are first parity and choose a way to arrange them to match up with incorrect second parity guesses and we look at the total number of principals which are second parity and choose a way to arrange them to match up with incorrect first parity guesses. Then we multiply that by the number of permutations of first parity principals and the number of permutations of second parity principals. And we arrive at ` x 1 2 (y--c) ` y 1 2 (y-+c) x!y!. Now, similarly, we can calculate the number of arrangements which will result in x - correct answers. And if for all there are at least as many arrangements which produce x - correct answers as produce x + of them then the average of cannot exceed 0. Now, if there are x - correct answers, then there must be y + incorrect ones. And we can use the same reasoning to establish that there must be 1 2 (y + + c) incorrect first parity guesses and 1 2 (y + - c) incorrect second parity guesses, and hence ` x 1 2 (y+-c) ` y 1 2 (y++c) x!y! arrangements which result in x correct guesses. So if we can prove that this is no less than the previous quantity then our proof will be complete. ` x 1 2 (y--c) ` y 1 2 (y-+c) x!y! ` x 1 2 (y+-c) ` y 1 2 (y++c) x!y! ` x 1 2 (y--c) ` y 1 2 (y-+c) ` x 1 2 (y+-c) ` y 1 2 (y++c) x! ( 1 2 (y--c))!(x-1 2 (y--c))! y! ( 1 2 (y-+c))!(y-1 2 (y-+c))! x! ( 1 2 (y+-c))!(x-1 2 (y+-c))! y! ( 1 2 (y++c))!(y-1 2 (y++c))! 1 ( 1 2 (y--c))!(x-1 2 (y--c))! 1 ( 1 2 (y-+c))!( 1 2 (y+-c))! 1 ( 1 2 (y+-c))!(x-1 2 (y+-c))! 1 ( 1 2 (y++c))!( 1 2 (y--c))! 1 (x-1 2 (y--c))!( 1 2 (y-+c))! 1 (x-1 2 (y+-c))!( 1 2 (y++c))! (x 1 2 (y - - c))!( 1 2 (y - + c))! (x 1 2 (y + - c))!( 1 2 (y + + c))! (x-1 2 (y--c))! (x-1 2 (y+-c))! ( 1 2 (y++c))! ( 1 2 (y-+c))! (x-1 2 (y--c))! (x-1 2 (y--c)-)! ( 1 2 (y++c))! ( 1 2 (y++c)-)! We define a function f (a, k) = a!/(a-k)!, i.e. the product starting from a going down k integers. And obviously a b f (a, k) f (b, k), b k 0. Then we can rewrite the last inequality as f (x 1 2 (y - c), ) f ( 1 2 (y + + c), ), which, noting that 0 and y + |c| y - c y + c + 2 1 2 (y + c + ) , is implied by x-1 2 (y - -c) 1 2 (y + +c) x-1 2 y 1 2 y x y h|vg| (1 - h)|vg| h (1 - h) which we know to be true from our assumption at the start of the proof. So we have proven lemma 2, and this completes the proof. B. PROOF OF THEOREM 2 We define n to be the number of users, |K|. Because we assume that this system is in a fixed state, every user k is in some configuration g k . Now let us examine some particular attribute, t. We know that a fraction h of users have that attribute and 1 - h do not. Let us define a set of policies L = {p|t T , k Kq s t Q such that p = q s t (k)}. We also need to know the fraction of users who have each policy in L. As the number of users grows towards infinity, the number of possible policies stays finite, so multiple users with the attribute will wind up sharing the same policy. For every member l L, we define f l to be the fraction of users with attribute t who have policy l. P lL f l = 1. We assume that as n approaches infinity, f l approaches some fixed quantity ^ f l for every l L. Essentially, what we are assuming is that there is a fixed fraction of users with the attribute who will chose any given policy. The particular number will vary at any given time, but over time, we will approach this fraction. We should then know that for some particular policy l, the odds of a user without the attribute drawing policy l are also f l because policies are handed out with the same distribution that they are submitted. The distribution which describes how many users we are actually going to have with this policy is a binomial distribution . The variance of a binomial distribution is 2 = n(1-h)f l (1-f l ). The difference between the actual and the ideal is the square root of the variance divided by the expected number of users who have a given policy, which is nf l . Hence, the expected difference between our practical system and the ideal system is n(1-h)f l (1-f l ) nf l = q (1-h)(1-f l ) nf l . 1 - h is a constant term, and f l will approach ^ f l , which is a fixed quantity. So lim ninf (1-h)(1-f l ) nf l = 0, and we have proven that our system approaches the ideal as the number of users goes to infinity. 45

Privacy;Trust Negotiation;Attribute-based Access Control

151

Probabilistic Author-Topic Models for Information Discovery

We propose a new unsupervised learning technique for extracting information from large text collections. We model documents as if they were generated by a two-stage stochastic process. Each author is represented by a probability distribution over topics, and each topic is represented as a probability distribution over words for that topic. The words in a multi-author paper are assumed to be the result of a mixture of each authors' topic mixture. The topic-word and author-topic distributions are learned from data in an unsupervised manner using a Markov chain Monte Carlo algorithm . We apply the methodology to a large corpus of 160,000 abstracts and 85,000 authors from the well-known CiteSeer digital library, and learn a model with 300 topics. We discuss in detail the interpretation of the results discovered by the system including specific topic and author models, ranking of authors by topic and topics by author, significant trends in the computer science literature between 1990 and 2002, parsing of abstracts by topics and authors and detection of unusual papers by specific authors. An online query interface to the model is also discussed that allows interactive exploration of author-topic models for corpora such as CiteSeer.

INTRODUCTION With the advent of the Web and various specialized digital libraries, the automatic extraction of useful information from text has become an increasingly important research area in data mining. In this paper we discuss a new algorithm that extracts both the topics expressed in large text document collections and models how the authors of documents use those topics. The methodology is illustrated using a sample of 160,000 abstracts and 80,000 authors from the well-known CiteSeer digital library of computer science research papers (Lawrence, Giles, and Bollacker, 1999). The algorithm uses a probabilistic model that represents topics as probability distributions over words and documents as being composed of multiple topics. A novel feature of our model is the inclusion of author models, in which authors are modeled as probability distributions over topics. The author-topic models can be used to support a variety of interactive and exploratory queries on the set of documents and authors, including analysis of topic trends over time, finding the authors who are most likely to write on a given topic, and finding the most unusual paper written by a given author. Bayesian unsupervised learning is used to fit the model to a document collection. Supervised learning techniques for automated categorization of documents into known classes or topics has received considerable attention in recent years (e.g., Yang, 1998). For many document collections, however, neither predefined topics nor labeled documents may be available. Furthermore , there is considerable motivation to uncover hidden topic structure in large corpora, particularly in rapidly changing fields such as computer science and biology, where predefined topic categories may not accurately reflect rapidly evolving content. Automatic extraction of topics from text, via unsupervised learning, has been addressed in prior work using a number of different approaches. One general approach is to represent the high-dimensional term vectors in a lower-dimensional space. Local regions in the lower-dimensional space can then be associated with specific topics. For example , the WEBSOM system (Lagus et al. 1999) uses nonlinear dimensionality reduction via self-organizing maps to represent term vectors in a two-dimensional layout. Linear projection techniques, such as latent semantic indexing (LSI), are also widely used (Berry, Dumais, and O' Brien, 1995). For example, Deerwester et al. (1990), while not using the term "topics" per se, state: Roughly speaking, these factors may be thought of as artificial concepts; they represent extracted common meaning components of many different words and documents. Research Track Paper 306 A somewhat different approach is to cluster the documents into groups containing similar semantic content, using any of a variety of well-known document clustering techniques (e.g., Cutting et al., 1992; McCallum, Nigam, and Ungar, 2000; Popescul et al., 2000). Each cluster of documents can then be associated with a latent topic (e.g., as represented by the mean term vector for documents in the cluster). While clustering can provide useful broad information about topics, clusters are inherently limited by the fact that each document is (typically) only associated with one cluster. This is often at odds with the multi-topic nature of text documents in many contexts. In particular, combinations of diverse topics within a single document are difficult to represent. For example, this present paper contains at least two significantly different topics: document modeling and Bayesian estimation. For this reason, other representations (such as those discussed below) that allow documents to be composed of multiple topics generally provide better models for sets of documents (e.g., better out of sample predictions , Blei, Ng, and Jordan (2003)). Hofmann (1999) introduced the aspect model (also referred to as probabilistic LSI, or pLSI) as a probabilistic alternative to projection and clustering methods. In pLSI, topics are modeled as multinomial probability distributions over words, and documents are assumed to be generated by the activation of multiple topics. While the pLSI model produced impressive results on a number of text document problems such as information retrieval, the parameterization of the model was susceptible to overfitting and did not provide a straightforward way to make inferences about new documents not seen in the training data. Blei, Ng, and Jordan (2003) addressed these limitations by proposing a more general Bayesian probabilistic topic model called latent Dirichlet allocation (LDA). The parameters of the LDA model (the topic-word and document-topic distributions) are estimated using an approximation technique known as variational EM, since standard estimation methods are intractable . Griffiths and Steyvers (2004) showed how Gibbs sampling, a Markov chain Monte Carlo technique, could be applied in this model, and illustrated this approach using 11 years of abstract data from the Proceedings of the National Academy of Sciences. Our focus here is to extend the probabilistic topic models to include authorship information. Joint author-topic modeling has received little or no attention as far as we are aware. The areas of stylometry, authorship attribution, and forensic linguistics focus on the problem of identifying what author wrote a given piece of text. For example, Mosteller and Wallace (1964) used Bayesian techniques to infer whether Hamilton or Madison was the more likely author of disputed Federalist papers. More recent work of a similar nature includes authorship analysis of a purported poem by Shakespeare (Thisted and Efron, 1987), identifying authors of software programs (Gray, Sallis, and MacDonell, 1997), and the use of techniques such as support vector machines (Diederich et al., 2003) for author identification. These author identification methods emphasize the use of distinctive stylistic features (such as sentence length) that characterize a specific author. In contrast, the models we present here focus on extracting the general semantic content of a document, rather than the stylistic details of how it was written. For example, in our model we omit common "stop" words since they are generally irrelevant to the topic of the document--however, the distributions of stop words can be quite useful in stylometry. While "topic" information could be usefully combined with stylistic features for author classification we do not pursue this idea in this particular paper. Graph-based and network-based models are also frequently used as a basis for representation and analysis of relations among scientific authors. For example, Newman (2001), Mutschke (2003) and Erten et al. (2003) use methods from bibliometrics, social networks, and graph theory to analyze and visualize co-author and citation relations in the scientific literature. Kautz, Selman, and Shah (1997) developed the interactive ReferralWeb system for exploring networks of computer scientists working in artificial intelligence and information retrieval, and White and Smyth (2003) used PageRank-style ranking algorithms to analyze co-author graphs. In all of this work only the network con-nectivity information is used--the text information from the underlying documents is not used in modeling. Thus, while the grouping of authors via these network models can implicitly provide indications of latent topics, there is no explicit representation of the topics in terms of the text content (the words) of the documents. The novelty of the work described in this paper lies in the proposal of a probabilistic model that represents both authors and topics, and the application of this model to a large well-known document corpus in computer science. As we will show later in the paper, the model provides a general framework for exploration, discovery, and query-answering in the context of the relationships of author and topics for large document collections. The outline of the paper is as follows: in Section 2 we describe the author-topic model and outline how the parameters of the model (the topic-word distributions and author-topic distributions) can be learned from training data consisting of documents with known authors. Section 3 illustrates the application of the model to a large collection of abstracts from the CiteSeer system, with examples of specific topics and specific author models that are learned by the algorithm. In Section 4 we illustrate a number of applications of the model, including the characterization of topic trends over time (which provides some interesting insights on the direction of research in computer science), and the characterization of which papers are most typical and least typical for a given author. An online query interface to the system is described in Section 5, allowing users to query the model over the Web--an interesting feature of the model is the coupling of Bayesian sampling and relational database technology to answer queries in real-time. Section 6 contains a brief discussion of future directions and concluding comments. AN OVERVIEW OF THE AUTHOR-TOPIC MODEL The author-topic model reduces the process of writing a scientific document to a simple series of probabilistic steps. The model not only discovers what topics are expressed in a document, but also which authors are associated with each topic. To simplify the representation of documents, we use a bag of words assumption that reduces each document to a Research Track Paper 307 x z w D K T d a Given the set of co-authors: N d 1. Choose an author 2. Choose a topic given the author 3. Choose a word given the topic Figure 1: The graphical model for the author-topic model using plate notation. vector of counts, where each vector element corresponds to the number of times a term appears in the document. Each author is associated with a multinomial distribution over topics. A document with multiple authors has a distribution over topics that is a mixture of the distributions associated with the authors. When generating a document, an author is chosen at random for each individual word in the document. This author picks a topic from his or her multinomial distribution over topics, and then samples a word from the multinomial distribution over words associated with that topic. This process is repeated for all words in the document. In the model, the authors produce words from a set of T topics. When T is kept relatively small relative to the number of authors and vocabulary size, the author-topic model applies a form of dimensionality reduction to documents ; topics are learned which capture the variability in word choice across a large set of documents and authors. In our simulations, we use 300 topics (see Rosen-Zvi et al. (2004) for an exploration of different numbers of topics). Figure 1 illustrates the generative process with a graphical model using plate notation. For readers not familiar with plate notation, shaded and unshaded variables indicate observed and latent variables respectively. An arrow indicates a conditional dependency between variables and plates (the boxes in the figure) indicate repeated sampling with the number of repetitions given by the variable in the bottom (see Buntine (1994) for an introduction). In the author-topic model, observed variables not only include the words w in a document but also the set of coauthors A d on each document d. Currently, the model does not specify the generative process of how authors choose to collaborate. Instead , we assume the model is provided with the authorship information on every document in the collection. Each author (from a set of K authors) is associated with a multinomial distribution over topics, represented by . Each topic is associated with a multinomial distribution over words, represented by . The multinomial distributions and have a symmetric Dirichlet prior with hyperparame-ters and (see Rosen-Zvi et al. (2004) for details). For each word in the document, we sample an author x uni-formly from A d , then sample a topic z from the multinomial distribution associated with author x and sample a word w from a multinomial topic distribution associated with topic z. This sampling process is repeated N times to form document d. 2.2 Bayesian Estimation of the Model Parameters The author-topic model includes two sets of unknown parameters--the K author-topic distributions , and the T topic distributions --as well as the latent variables corresponding to the assignments of individual words to topics z and authors x. The Expectation-Maximization (EM) algorithm is a standard technique for estimating parameters in models with latent variables, finding a mode of the posterior distribution over parameters. However, when applied to probabilistic topic models (Hofmann, 1999), this approach is susceptible to local maxima and computationally inefficient (see Blei, Ng, and Jordan, 2003). We pursue an alternative parameter estimation strategy, outlined by Griffiths and Steyvers (2004), using Gibbs sampling, a Markov chain Monte Carlo algorithm to sample from the posterior distribution over parameters. Instead of estimating the model parameters directly, we evaluate the posterior distribution on just x and z and then use the results to infer and . For each word, the topic and author assignment are sam-pled from: P (z i = j, x i = k|w i = m, z -i , x -i ) C W T mj + m C W T m j + V C AT kj + j C AT kj + T (1) where z i = j and x i = k represent the assignments of the ith word in a document to topic j and author k respectively , w i = m represents the observation that the ith word is the mth word in the lexicon, and z -i , x -i represent all topic and author assignments not including the ith word. Furthermore, C W T mj is the number of times word m is assigned to topic j, not including the current instance, and C AT kj is the number of times author k is assigned to topic j, not including the current instance, and V is the size of the lexicon. During parameter estimation, the algorithm only needs to keep track of a V T (word by topic) count matrix, and a K T (author by topic) count matrix, both of which can be represented efficiently in sparse format. From these count matrices, we can easily estimate the topic-word distributions and author-topic distributions by: mj = C W T mj + m C W T m j + V (2) kj = C AT kj + j C AT kj + T (3) where mj is the probability of using word m in topic j, and kj is the probability of using topic j by author k. These values correspond to the predictive distributions over new words w and new topics z conditioned on w and z. We start the algorithm by assigning words to random topics and authors (from the set of authors on the document). Each Gibbs sample then constitutes applying Equation (1) to every word token in the document collection. This sampling process is repeated for I iterations. In this paper we primarily focus on results based on a single sample so that specific topics can be identified and interpreted--in tasks involving prediction of words and authors one can average over topics and use multiple samples when doing so (Rosen-Zvi et al., 2004). Research Track Paper 308 WORD PROB. WORD PROB. WORD PROB. WORD PROB. PATTERNS 0.1965 USER 0.3290 MAGNETIC 0.0155 METHODS 0.5319 PATTERN 0.1821 INTERFACE 0.1378 STARS 0.0145 METHOD 0.1403 MATCHING 0.1375 USERS 0.1060 SOLAR 0.0135 TECHNIQUES 0.0442 MATCH 0.0337 INTERFACES 0.0498 EMISSION 0.0127 DEVELOPED 0.0216 TEXT 0.0242 SYSTEM 0.0434 MASS 0.0125 APPLIED 0.0162 PRESENT 0.0207 INTERACTION 0.0296 OBSERVATIONS 0.0120 BASED 0.0153 MATCHES 0.0167 INTERACTIVE 0.0214 STAR 0.0118 APPROACHES 0.0133 PAPER 0.0126 USABILITY 0.0132 RAY 0.0112 COMPARE 0.0113 SHOW 0.0124 GRAPHICAL 0.0092 GALAXIES 0.0105 PRACTICAL 0.0112 APPROACH 0.0099 PROTOTYPE 0.0086 OBSERVED 0.0098 STANDARD 0.0102 AUTHOR PROB. AUTHOR PROB. AUTHOR PROB. AUTHOR PROB. Navarro_G 0.0133 Shneiderman_B 0.0051 Falcke_H 0.0140 Srinivasan_A 0.0018 Amir_A 0.0099 Rauterberg_M 0.0046 Linsky_J 0.0082 Mooney_R 0.0018 Gasieniec_L 0.0062 Harrison_M 0.0025 Butler_R 0.0077 Owren_B 0.0018 Baeza-Yates_R 0.0048 Winiwarter_W 0.0024 Knapp_G 0.0067 Warnow_T 0.0016 Baker_B 0.0042 Ardissono_L 0.0021 Bjorkman_K 0.0065 Fensel_D 0.0016 Arikawa_S 0.0041 Billsus_D 0.0019 Kundu_M 0.0060 Godsill_S 0.0014 Crochemore_M 0.0037 Catarci_T 0.0017 Christensen-D_J 0.0057 Saad_Y 0.0014 Rytter_W 0.0034 St_R 0.0017 Mursula_K 0.0054 Hansen_J 0.0013 Raffinot_M 0.0032 Picard_R 0.0016 Cranmer_S 0.0051 Zhang_Y 0.0013 Ukkonen_E 0.0032 Zukerman_I 0.0016 Nagar_N 0.0050 Dietterich_T 0.0013 WORD PROB. WORD PROB. WORD PROB. WORD PROB. DATA 0.1622 PROBABILISTIC 0.0869 RETRIEVAL 0.1208 QUERY 0.1406 MINING 0.0657 BAYESIAN 0.0791 INFORMATION 0.0613 QUERIES 0.0947 DISCOVERY 0.0408 PROBABILITY 0.0740 TEXT 0.0461 DATABASE 0.0932 ATTRIBUTES 0.0343 MODEL 0.0533 DOCUMENTS 0.0385 DATABASES 0.0468 ASSOCIATION 0.0328 MODELS 0.0466 INDEXING 0.0369 DATA 0.0426 LARGE 0.0279 PROBABILITIES 0.0308 DOCUMENT 0.0316 RELATIONAL 0.0384 DATABASES 0.0257 INFERENCE 0.0306 QUERY 0.0261 JOIN 0.0188 KNOWLEDGE 0.0175 CONDITIONAL 0.0274 CONTENT 0.0256 PROCESSING 0.0165 PATTERNS 0.0174 PRIOR 0.0273 SEARCH 0.0174 SOURCES 0.0114 ITEMS 0.0173 POSTERIOR 0.0228 RELEVANCE 0.0171 OPTIMIZATION 0.0110 AUTHOR PROB. AUTHOR PROB. AUTHOR PROB. AUTHOR PROB. Han_J 0.0164 Koller_D 0.0104 Oard_D 0.0097 Levy_A 0.0092 Zaki_M 0.0089 Heckerman_D 0.0079 Hawking_D 0.0065 Naughton_J 0.0078 Liu_B 0.0071 Ghahramani_Z 0.0060 Croft_W 0.0057 Suciu_D 0.0075 Cheung_D 0.0066 Friedman_N 0.0060 Jones_K 0.0053 Raschid_L 0.0075 Shim_K 0.0051 Myllymaki_P 0.0057 Schauble_P 0.0052 DeWitt_D 0.0062 Mannila_H 0.0049 Lukasiewicz_T 0.0054 Voorhees_E 0.0050 Widom_J 0.0058 Rastogi_R 0.0049 Geiger_D 0.0045 Callan_J 0.0046 Abiteboul_S 0.0057 Ganti_V 0.0048 Muller_P 0.0044 Fuhr_N 0.0042 Chu_W 0.0055 Toivonen_H 0.0043 Berger_J 0.0044 Smeaton_A 0.0042 Libkin_L 0.0054 Liu_H 0.0043 Xiang_Y 0.0042 Sanderson_M 0.0041 Kriegel_H 0.0054 TOPIC 29 TOPIC 58 TOPIC 298 TOPIC 139 TOPIC 52 TOPIC 95 TOPIC 293 TOPIC 68 Figure 2: Eight example topics extracted from the CiteSeer database. Each is illustrated with the 10 most likely words and authors with corresponding probabilities. WORD PROB. WORD PROB. WORD PROB. WORD PROB. DATA 0.1468 PROBABILISTIC 0.0826 RETRIEVAL 0.1381 QUERY 0.1699 MINING 0.0631 BAYESIAN 0.0751 INFORMATION 0.0600 QUERIES 0.1209 DISCOVERY 0.0396 PROBABILITY 0.0628 INDEX 0.0529 JOIN 0.0258 ATTRIBUTES 0.0392 MODEL 0.0364 INDEXING 0.0469 DATA 0.0212 ASSOCIATION 0.0316 PROBABILITIES 0.0313 QUERY 0.0319 OPTIMIZATION 0.0171 RULES 0.0252 INFERENCE 0.0294 CONTENT 0.0299 PROCESSING 0.0162 PATTERNS 0.0210 MODELS 0.0273 BASED 0.0224 RELATIONAL 0.0131 LARGE 0.0207 CONDITIONAL 0.0262 SEARCH 0.0219 DATABASE 0.0128 ATTRIBUTE 0.0183 DISTRIBUTION 0.0261 RELEVANCE 0.0212 AGGREGATION 0.0117 DATABASES 0.0179 PRIOR 0.0259 SIMILARITY 0.0178 RESULT 0.0106 AUTHOR PROB. AUTHOR PROB. AUTHOR PROB. AUTHOR PROB. Han_J 0.0157 Koller_D 0.0109 Oard_D 0.0080 Naughton_J 0.0103 Zaki_M 0.0104 Heckerman_D 0.0079 Voorhees_E 0.0053 Suciu_D 0.0091 Liu_B 0.0080 Friedman_N 0.0076 Hawking_D 0.0053 Levy_A 0.0080 Cheung_D 0.0075 Ghahramani_Z 0.0060 Schauble_P 0.0051 DeWitt_D 0.0077 Hamilton_H 0.0058 Lukasiewicz_T 0.0053 Croft_W 0.0051 Wong_L 0.0071 Mannila_H 0.0056 Myllymaki_P 0.0053 Jones_K 0.0041 Ross_K 0.0067 Brin_S 0.0055 Poole_D 0.0050 Bruza_P 0.0041 Kriegel_H 0.0055 Ganti_V 0.0050 Xiang_Y 0.0048 Lee_D 0.0040 Mumick_I 0.0054 Liu_H 0.0050 vanderGaag_L 0.0047 Smeaton_A 0.0040 Raschid_L 0.0053 Toivonen_H 0.0049 Berger_J 0.0040 Callan_J 0.0039 Kossmann_D 0.0053 TOPIC 276 TOPIC 158 TOPIC 213 TOPIC 15 Figure 3: The four most similar topics to the topics in the bottom row of Figure 2, obtained from a different Markov chain run. AUTHOR-TOPICS FOR CITESEER Our collection of CiteSeer abstracts contains D = 162, 489 abstracts with K = 85, 465 authors. We preprocessed the text by removing all punctuation and common stop words. This led to a vocabulary size of V = 30, 799, and a total of 11, 685, 514 word tokens. There is inevitably some noise in data of this form given that many of the fields (paper title, author names, year, abstract ) were extracted automatically by CiteSeer from PDF or postscript or other document formats. We chose the simple convention of identifying authors by their first initial and second name, e.g., A Einstein, given that multiple first initials or fully spelled first names were only available for a relatively small fraction of papers. This means of course that for some very common names (e.g., J Wang or J Smith) there will be multiple actual individuals represented by a single name in the model. This is a known limitation of working with this type of data (e.g., see Newman (2001) for further discussion). There are algorithmic techniques that could be used to automatically resolve these identity problems-however , in this paper, we don't pursue these options and instead for simplicity work with the first-initial/last-name representation of individual authors. In our simulations, the number of topics T was fixed at 300 and the smoothing parameters and (Figure 1) were set at 0.16 and 0.01 respectively. We ran 5 independent Gibbs sampling chains for 2000 iterations each. On a 2GHz PC workstation, each iteration took 400 seconds, leading to a total run time on the order of several days per chain. 3.2 Author-Topic and Topic-Word Models for the CiteSeer Database We now discuss the author-topic and topic-word distributions learned from the CiteSeer data. Figure 2 illustrates eight different topics (out of 300), obtained at the 2000th iteration of a particular Gibbs sampler run. Each table in Figure 2 shows the 10 words that are most likely to be produced if that topic is activated, and the 10 authors who are most likely to have produced a word if it is known to have come from that topic. The words associated with each topic are quite intuitive and, indeed, quite precise in the sense of conveying a semantic summary of a particular field of research. The authors associated with each topic are also quite representative--note that the top 10 authors associated with a topic by the model are not necessarily the most well-known authors in that area, but rather are the authors who tend to produce the most words for that topic (in the CiteSeer abstracts). The first 3 topics at the top of Figure 2, topics #163, #87 and #20 show examples of 3 quite specific and precise topics on string matching, human-computer interaction, and astronomy respectively. The bottom four topics (#205, #209, #289, and #10) are examples of topics with direct relevance to data mining--namely data mining itself, probabilistic learning, information retrieval, and database querying and indexing. The model includes several other topics related to data mining, such as predictive modeling and neural networks , as well as topics that span the full range of research areas encompassed by documents in CiteSeer. The full list is available at http://www.datalab.uci.edu/author-topic. Topic #273 (top right Figure 2) provides an example of a Research Track Paper 309 topic that is not directly related to a specific research area. A fraction of topics, perhaps 10 to 20%, are devoted to "non-research -specific" topics, the "glue" that makes up our research papers, including general terminology for describing methods and experiments, funding acknowledgments and parts of addresses(which inadvertently crept in to the abstracts ), and so forth. We found that the topics obtained from different Gibbs sampling runs were quite stable. For example, Figure 3 shows the 4 most similar topics to the topics in the bottom row of Figure 2, but from a different run. There is some variability in terms of ranking of specific words and authors for each topic, and in the exact values of the associated probabilities, but overall the topics match very closely. APPLICATIONS OF THE AUTHOR-TOPIC MODEL TO CITESEER Of the original 162,489 abstracts in our data set, estimated years of publication were provided by CiteSeer for 130, 545 of these abstracts. There is a steady (and well-known) increase year by year in the number of online documents through the 1990's. From 1999 through 2002, however, the number of documents for which the year is known drops off sharply-the years 2001 and 2002 in particular are under-represented in this set. This is due to fact that it is easier for CiteSeer to determine the date of publication of older documents, e.g., by using citations to these documents. We used the yearly data to analyze trends in topics over time. Using the same 300 topic model described earlier, the documents were partitioned by year, and for each year all of the words were assigned to their most likely topic using the model. The fraction of words assigned to each topic for a given year was then calculated for each of the 300 topics and for each year from 1990 to 2002. These fractions provide interesting and useful indicators of relative topic popularity in the research literature in recent years. Figure 4 shows the results of plotting several different topics. Each topic is indicated in the legend by the five most probable words in the topic. The top left plot shows a steady increase (roughly three-fold) in machine learning and data mining topics. The top right plot shows a "tale of two topics": an increase in information-retrieval coupled to an apparent decrease in natural language processing. On the second row, on the left we see a steady decrease in two "classical" computer science topics, operating systems and programming languages. On the right, however, we see the reverse behavior, namely a corresponding substantial growth in Web-related topics. In the third row, the left plot illustrates trends within database research: a decrease in the transaction and concurrency-related topic, query-related research holding steady over time, and a slow but steady increase in integration-related database research. The plot on the right in the third row illustrates the changing fortunes of security-related research--a decline in the early 90's but then a seemingly dramatic upward trend starting around 1995. The lower left plot on the bottom row illustrates the somewhat noisy trends of three topics that were "hot" in the 1990's: neural networks exhibits a steady decline since the early 1990's (as machine learning has moved on to areas such as support vector machines), genetic algorithms appears to be relatively stable, and wavelets may have peaked in the 199498 time period. Finally, as with any large data set there are always some surprises in store. The final figure on the bottom right shows two somewhat unexpected "topics". The first topic consists entirely of French words (in fact the model discovered 3 such French language topics ). The apparent peaking of French words in the mid-1990s is likely to be an artifact of how CiteSeer preprocesses data rather than any indication of French research productivity. The lower curve corresponds to a topic consisting of largely Greek letters, presumably from more theoretically oriented papers--fans of theory may be somewhat dismayed to see that there is an apparent steady decline in the relative frequency of Greek letters in abstracts since the mid-1990s! The time-trend results above should be interpreted with some caution. As mentioned earlier, the data for 2001 and 2002 are relatively sparse compared to earlier years. In addition , the numbers are based on a rather skewed sample (online documents obtained by the CiteSeer system for which years are known). Furthermore, the fractions per year only indicate the relative number of words assigned to a topic by the model and make no direct assessment of the quality or importance of a particular sub-area of computer science. Nonetheless, despite these caveats, the results are quite informative and indicate substantial shifts in research topics within the field of computer science. In terms of related work, Popescul et al. (2000) investi-gated time trends in CiteSeer documents using a document clustering approach. 31K documents were clustered into 15 clusters based on co-citation information while the text information in the documents was not used. Our author-topic model uses the opposite approach. In effect we use the text information directly to discover topics and do not explic-itly model the "author network" (although implicitly the co-author connections are used by the model). A direct quantitative comparison is difficult, but we can say that our model with 300 topics appears to produce much more noticeable and precise time-trends than the 15-cluster model. 4.2 Topics and Authors for New Documents In many applications, we would like to quickly assess the topic and author assignments for new documents not contained in our subset of the CiteSeer collection. Because our Monte Carlo algorithm requires significant processing time for 160K documents, it would be computationally inefficient to rerun the algorithm for every new document added to the collection (even though from a Bayesian inference viewpoint this is the optimal approach). Our strategy instead is to apply an efficient Monte Carlo algorithm that runs only on the word tokens in the new document, leading quickly to likely assignments of words to authors and topics. We start by assigning words randomly to co-authors and topics. We then sample new assignments of words to topics and authors by applying Equation 1 only to the word tokens in the new document each time temporarily updating the count matrices C W T and C AT . The resulting assignments of words to authors and topics can be saved after a few iterations (10 iterations in our simulations). Figure 5 shows an example of this type of inference. Abstracts from two authors, B Scholkopf and A Darwiche were combined together into 1 "pseudo-abstract" and the docu-Research Track Paper 310 1990 1992 1994 1996 1998 2000 2002 1 2 3 4 5 6 7 8 x 10 -3 Year Fraction of Words Assigned to Topic 114:regression-variance-estimator -estimators-bias 153:classification-training-classifier -classifiers-generalization 205:data-mining-attributes-discovery -association 1990 1992 1994 1996 1998 2000 2002 2 3 4 5 6 7 8 x 10 -3 Year Fraction of Words Assigned to Topic 280:language-semantic-natural -linguistic-grammar 289:retrieval-text-documents -information-document 1990 1992 1994 1996 1998 2000 2002 2 3 4 5 6 7 8 9 10 11 x 10 -3 Year Fraction of Words Assigned to Topic 60:programming-language-concurrent -languages-implementation 139:system-operating-file -systems-kernel 1990 1992 1994 1996 1998 2000 2002 0 0.002 0.004 0.006 0.008 0.01 0.012 Year Fraction of Words Assigned to Topic 7:web-user-world-wide-users 80:mobile-wireless-devices -mobility-ad 275:multicast-multimedia-media -delivery-applications 1990 1992 1994 1996 1998 2000 2002 1 2 3 4 5 6 7 8 9 10 x 10 -3 Year Fraction of Words Assigned to Topic 10:query-queries-index-data-join 261:transaction-transactions -concurrency-copy-copies 194:integration-view-views-data -incremental 1990 1992 1994 1996 1998 2000 2002 1 2 3 4 5 6 7 8 9 x 10 -3 Year Fraction of Words Assigned to Topic 120:security-secure-access-key-authentication 240:key-attack-encryption-hash-keys 1990 1992 1994 1996 1998 2000 2002 1 2 3 4 5 6 7 x 10 -3 Year Fraction of Words Assigned to Topic 23:neural-networks-network-training-learning 35:wavelet-operator-operators-basis -coefficients 242:genetic-evolutionary-evolution-population-ga 1990 1992 1994 1996 1998 2000 2002 0 0.002 0.004 0.006 0.008 0.01 0.012 Year Fraction of Words Assigned to Topic 47:la-les-une-nous-est 157:gamma-delta-ff-omega-oe Figure 4: Topic trends for research topics in computer science. Research Track Paper 311 [ AUTH1=Scholkopf_B ( 69%, 31%)] [ AUTH2=Darwiche_A ( 72%, 28%)] A method 1 is described which like the kernel 1 trick 1 in support 1 vector 1 machines 1 SVMs 1 lets us generalize distance 1 based 2 algorithms to operate in feature 1 spaces usually nonlinearly related to the input 1 space This is done by identifying a class of kernels 1 which can be represented as norm 1 based 2 distances 1 in Hilbert spaces It turns 1 out that common kernel 1 algorithms such as SVMs 1 and kernel 1 PCA 1 are actually really distance 1 based 2 algorithms and can be run 2 with that class of kernels 1 too As well as providing 1 a useful new insight 1 into how these algorithms work the present 2 work can form the basis 1 for conceiving new algorithms This paper presents 2 a comprehensive approach for model 2 based 2 diagnosis 2 which includes proposals for characterizing and computing 2 preferred 2 diagnoses 2 assuming that the system 2 description 2 is augmented with a system 2 structure 2 a directed 2 graph 2 explicating the interconnections between system 2 components 2 Specifically we first introduce the notion of a consequence 2 which is a syntactically 2 unconstrained propositional 2 sentence 2 that characterizes all consistency 2 based 2 diagnoses 2 and show 2 that standard 2 characterizations of diagnoses 2 such as minimal conflicts 1 correspond to syntactic 2 variations 1 on a consequence 2 Second we propose a new syntactic 2 variation on the consequence 2 known as negation 2 normal form NNF and discuss its merits compared to standard variations Third we introduce a basic algorithm 2 for computing consequences in NNF given a structured system 2 description We show that if the system 2 structure 2 does not contain cycles 2 then there is always a linear size 2 consequence 2 in NNF which can be computed in linear time 2 For arbitrary 1 system 2 structures 2 we show a precise connection between the complexity 2 of computing 2 consequences and the topology of the underlying system 2 structure 2 Finally we present 2 an algorithm 2 that enumerates 2 the preferred 2 diagnoses 2 characterized by a consequence 2 The algorithm 2 is shown 1 to take linear time 2 in the size 2 of the consequence 2 if the preference criterion 1 satisfies some general conditions Figure 5: Automated labeling of a pseudo-abstract from two authors by the model. ment treated as if they had both written it. These two authors work in relatively different but not entirely unrelated sub-areas of computer science: Scholkopf in machine learning and Darwiche in probabilistic reasoning. The document is then parsed by the model. i.e., words are assigned to these authors. We would hope that the author-topic model, conditioned now on these two authors, can separate the combined abstract into its component parts. Figure 5 shows the results after the model has classified each word according to the most likely author. Note that the model only sees a bag of words and is not aware of the word order that we see in the figure. For readers viewing this in color, the more red a word is the more likely it is to have been generated (according to the model) by Scholkopf (and blue for Darwiche). For readers viewing the figure in black and white, the superscript 1 indicates words classified by the model for Scholkopf, and superscript 2 for Darwiche. The results show that all of the significant content words (such as kernel, support, vector, diagnoses, directed, graph) are classified correctly. As we might expect most of the "er-rors" are words (such as "based" or "criterion") that are not specific to either authors' area of research. Were we to use word order in the classification, and classify (for example) whole sentences, the accuracy would increase further. As it is, the model correctly classifies 69% of Scholkopf's words and 72% of Darwiche's. 4.3 Detecting the Most Surprising and Least Surprising Papers for an Author In Tables 1 through 3 we used the model to score papers attributed to three well-known researchers in computer science (Christos Faloutsos, Michael Jordan, and Tom Mitchell). For each document for each of these authors we calculate a perplexity score. Perplexity is widely used in language modeling to assess the predictive power of a model. It is a measure of how surprising the words are from the model's perspective, loosely equivalent to the effective branching factor . Formally, the perplexity score of a new unobserved document d that contains a set of words W d and conditioned on a topic model for a specific author a is: Perplexity(W d |a) = exp - log p(W d |a) |W d | where p(W d |a) is the probability assigned by the author topic model to the words W d conditioned on the single author a, and |W d | is the number of words in the document. Even if the document was written by multiple authors we evaluate the perplexity score relative to a single author in order to judge perplexity relative to that individual. Our goal here is not to evaluate the out-of-sample predictive power of the model, but to explore the range of perplexity scores that the model assigns to papers from specific authors. Lower scores imply that the words w are less surprising to the model (lower bounded by zero).In particular we are interested in the abstracts that the model considers most surprising (highest perplexity) and least surprising (lowest perplexity)--in each table we list the 2 abstracts with the highest perplexity scores, the median perplexity, and the 2 abstracts with the lowest perplexity scores. Table 1 for Christos Faloutsos shows that the two papers with the highest perplexities have significantly higher perplexity scores than the median and the two lowest perplexity papers. The high perplexity papers are related to "query by example" and the QBIC image database system, while the low perplexity papers are on high-dimensional indexing. As far as the topic model for Faloutsos is concerned, the indexing papers are much more typical of his work than the query by example papers. Tables 2 and 3 provide interesting examples in that the most perplexing papers (from the model's viewpoint) for each author are papers that the author did not write at all. As mentioned earlier, by combining all T Mitchell's and M Jordan's together, the data set may contain authors who are different from Tom Mitchell at CMU and Michael Jordan at Berkeley. Thus, the highest perplexity paper for T Mitchell is in fact authored by a Toby Mitchell and is on the topic of estimating radiation doses (quite different from the machine learning work of Tom Mitchell). Similarly, for Michael Jordan, the most perplexing paper is on software Research Track Paper 312 Table 1: Papers ranked by perplexity for C. Faloutsos, from 31 documents. Paper Title Perplexity Score MindReader: Querying databases through multiple examples 1503.7 Efficient and effective querying by image content 1498.2 MEDIAN SCORE 603.5 Beyond uniformity and independence: analysis of R-trees using the concept of fractal dimension 288.9 The TV-tree: an index structure for high-dimensional data 217.2 Table 2: Papers ranked by perplexity for M. Jordan, from 33 documents. Paper Title Perplexity Score Software configuration management in an object oriented database 1386.0 Are arm trajectories planned in kinematic or dynamic coordinates? An adaptation study 1319.2 MEDIAN SCORE 372.4 On convergence properties of the EM algorithm for Gaussian mixtures 180.0 Supervised learning from incomplete data via an EM approach 179.0 Table 3: Papers ranked by perplexity for T. Mitchell from 15 documents. Paper Title Perplexity Score A method for estimating occupational radiation dose to individuals, using weekly dosimetry data 2002.9 Text classification from labeled and unlabeled documents using EM 845.4 MEDIAN SCORE 411.5 Learning one more thing 266.5 Explanation based learning for mobile robot perception 264.2 configuration management and was written by Mick Jordan of Sun Microsystems. In fact, of the 7 most perplexing papers for M Jordan, 6 are on software management and the JAVA programming language, all written by Mick Jordan. However, the 2nd most perplexing paper was in fact coauthored by Michael Jordan, but in the area of modeling of motor planning, which is a far less common topic compared to the machine learning papers that Jordan typically writes. AN AUTHOR-TOPIC BROWSER We have built a JAVA-based query interface tool that supports interactive querying of the model 1 . The tool allows a user to query about authors, topics, documents, or words. For example, given a query on a particular author the tool retrieves and displays the most likely topics and their probabilities for that author, the 5 most probable words for each topic, and the document titles in the database for that author . Figure 6(a) (top panel) shows the result of querying on Pazzani M and the resulting topic distribution (highly-ranked topics include machine learning, classification, rule-based systems, data mining, and information retrieval). Mouse-clicking on one of the topics (e.g., the data mining topic as shown in the figure) produces the screen display to the left (Figure 6(b)). The most likely words for this topic and the most likely authors given a word from this topic are then displayed. We have found this to be a useful technique for interactively exploring topics and authors, e.g., which authors are active in a particular research area. Similarly, one can click on a particular paper (e.g., the paper A Learning Agent for Wireless News Access as shown in the lower screenshot (Figure 6(c)) and the display in the panel to the right is then produced. This display shows the words in the documents and their counts, the probability distribution over topics for the paper given the word counts 1 A prototype online version of the tool can be accessed at http://www.datalab.uci.edu/author-topic . (ranked by highest probability first), and a probability distribution over authors, based on the proportion of words assigned by the model to each topic and author respectively. The system is implemented using a combination of a relational database and real-time Bayesian estimation (a relatively rare combination of these technologies for a real-time query-answering system as far as we are aware). We use a database to store and index both (a) the sparse author-topic and topic-word count matrices that are learned by our algorithm from the training data, and (b) various tables describing the data such as document-word, document-author, and document-title tables. For a large document set such as CiteSeer (and with 300 topics) these tables can run into the hundred's of megabytes of memory--thus, we do not load them into main memory automatically but instead issue SQL commands to retrieve the relevant records in real-time. For most of the queries we have implemented to date the queries can be answered by simple table lookup followed by appropriate normalization (if needed) of the stored counts to generate conditional probabilities. For example, displaying the topic distribution for a specific author is simply a matter of retrieving the appropriate record. However, when a document is the basis of a query (e.g., as in the lower screenshot, Figure 6(c)) we must compute in real-time the conditional distribution of the fraction of words assigned to each topic and author, a calculation that cannot be computed in closed form. This requires retrieving all the relevant word-topic counts for the words in the document via SQL, then executing the estimation algorithm outlined in Section 4.2 in real-time using Gibbs sampling, and displaying the results to the user. The user can change adjust the burn-in time, the number of samples and the lag time in the sampling algorithm--typically we have found that as few as 10 Gibbs samples gives quite reasonable results (and takes on the order of 1 or 2 seconds depending on the machine being used other factors). Research Track Paper 313 (b) (a) (c) Figure 6: Examples of screenshots from the interactive query browser for the author-topic model with (a) querying on author Pazzani M, (b) querying on a topic (data mining) relevant to that author, and (c) querying on a particular document written by the author. Research Track Paper 314 CONCLUSIONS We have introduced a probabilistic algorithm that can that can automatically extract information about authors, topics, and documents from large text corpora. The method uses a generative probabilistic model that links authors to observed words in documents via latent topics. We demon-strated that Bayesian estimation can be used to learn such author-topic models from very large text corpora, using CiteSeer abstracts as a working example. The resulting CiteSeer author-topic model was shown to extract substantial novel "hidden" information from the set of abstracts, including topic time-trends, author-topic relations, unusual papers for specific authors and so forth. Other potential applications not discussed here include recommending potential reviewers for a paper based on both the words in the paper and the names of the authors. Even though the underlying probabilistic model is quite simple, and ignores several aspects of real-world document generation (such as topic correlation, author interaction, and so forth), it nonetheless provides a useful first step in understanding author-topic structure in large text corpora. Acknowledgements We would like to thank Steve Lawrence, C. Lee Giles, and Isaac Council for providing the CiteSeer data used in this paper. We also thank Momo Alhazzazi, Amnon Meyers, and Joshua O'Madadhain for assistance in software development and data preprocessing. The research in this paper was supported in part by the National Science Foundation under Grant IRI-9703120 via the Knowledge Discovery and Dissemination (KD-D) program. References Blei, D. M., Ng, A. Y., and Jordan, M. I., (2003) Latent Dirichlet allocation, Journal of Machine Learning Research 3, pp. 9931022. Buntine, W.L. (1994) Operations for learning with graphical models, Journal of Artificial Intelligence Research 2, pp. 159-225. Cutting, D., Karger, D. R., Pederson, J., and Tukey, J. W. (1992) Scatter/Gather: a cluster-based approach to browsing large document collections, in Proceedings of the 15th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 318329. Deerwester, S. C., Dumais, S. T., Landauer, T. K., Furnas, G. W., and Harshman, R. A. (1990) Indexing by latent semantic analysis, Journal of the American Society of Information Science, 41(6), pp. 391407. Diederich, J., Kindermann, J., Leopold, E., and Paass, G. (2003) Authorship attribution with support vector machines , Applied Intelligence 19 (1). Erten, C., Harding, P. J., Kobourov, S. G., Wampler, K., and Yee, G. (2003) Exploring the computing literature using temporal graph visualization, Technical Report, Department of Computer Science, University of Arizona . Gray, A., Sallis, P., MacDonell, S. (1997) Software forensics : Extending authorship analysis techniques to computer programs, Proceedings of the 3rd Biannual Conference of the International Association of Forensic Linguists (IAFL), Durham NC. Griffiths, T. L., and Steyvers , M. (2004) Finding scientific topics, Proceedings of the National Academy of Sciences, 101 (suppl. 1), 52285235. Hofmann, T. (1999) Probabilistic latent semantic indexing , in Proceedings of the 22nd International Conference on Research and Development in Information Retrieval (SIGIR'99). Kautz, H., Selman, B., and Shah, M. (1997) Referral Web: Combining social networks and collaborative filtering, Communications of the ACM, 3, pp. 6365. Lagus, K, Honkela, T., Kaski, S., and Kohonen, T. (1999) WEBSOM for textual data mining, Artificial Intelligence Review, 13 (56), pp. 345364. Lawrence, S., Giles, C. L., and Bollacker, K. (1999) Digital libraries and autonomous citation indexing, IEEE Computer, 32(6), pp. 6771. McCallum, A., Nigam, K., and Ungar, L. (2000) Efficient clustering of high-dimensional data sets with application to reference matching, in Proceedings of the Sixth ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 169178. Mosteller, F., and Wallace, D. (1964) Applied Bayesian and Classical Inference: The Case of the Federalist Papers, Springer-Verlag. Mutschke, P. (2003) Mining networks and central entities in digital libraries: a graph theoretic approach applied to co-author networks, Intelligent Data Analysis 2003, Lecture Notes in Computer Science 2810, Springer Verlag, pp. 155166 Newman, M. E. J. (2001) Scientific collaboration networks: I. Network construction and fundamental results, Physical Review E, 64, 016131. Popescul, A., Flake, G. W., Lawrence, S., Ungar, L. H., and Giles, C. L. (2000) Clustering and identifying temporal trends in document databases, IEEE Advances in Digital Libraries, ADL 2000, pp. 173182. Rosen-Zvi, M., Griffiths, T., Steyvers, M., Smyth, P. (2004) The author-topic model for authors and documents, Proceedings of the 20th UAI Conference, July 2004. Thisted, B., and Efron, R. (1987) Did Shakespeare write a newly discovered poem?, Biometrika, pp. 445455. White, S. and Smyth, P. (2003) Algorithms for estimating relative importance in networks, in Proceedings of the Ninth ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 266275. Yang, Y. (1999) An evaluation of statistical approaches to text categorization, Information Retrieval, 1, pp. 69 90. Research Track Paper 315

Gibbs sampling;text modeling;unsupervised learning

152

Proportional Search Interface Usability Measures

Speed, accuracy, and subjective satisfaction are the most common measures for evaluating the usability of search user interfaces. However, these measures do not facilitate comparisons optimally and they leave some important aspects of search user interfaces uncovered. We propose new, proportional measures to supplement the current ones. Search speed is a normalized measure for the speed of a search user interface expressed in answers per minute. Qualified search speed reveals the trade-off between speed and accuracy while immediate search accuracy addresses the need to measure success in typical web search behavior where only the first few results are interesting. The proposed measures are evaluated by applying them to raw data from two studies and comparing them to earlier measures. The evaluations indicate that they have desirable features.

INTRODUCTION In order to study the usability of search user interfaces we need proper measures. In the literature, speed, accuracy and subjective satisfaction measures are common and reveal interesting details. They have, however, a few shortcomings that call for additional measures. First, comparing results even within one experiment--let alone between different experiments--is hard because the measures are not typically normalized in the research reports but multiple raw numbers (like answers found and time used) are reported. Of course, unbiased comparison between studies will always be difficult as the test setup has a big effect on the results, but the problem is compounded by the presentation of multiple task dependent measures. A good measure would be as simple as possible, yet it must not discard relevant information. Second, the current measures do not reveal the sources of speed differences. In particular, the relation between speed and accuracy may be hard to understand since the current measures for those dimensions are completely separate. For example, it is essential to know if the increase in speed is due to careless behavior or better success. Third, in the web environment, a typical goal for a search is to find just a few good enough answers to a question. This is demonstrated by studies that show that about half of the users only view one or two result pages per query [11]. Current search user interface usability measures do not capture the success of such a behavior very well. In order to address these problems, we present three new proportional, normalized usability measures. The new measures are designed for the result evaluation phase of the search process [10] where real users are involved. Search speed is a normalized speed measure expressed in answers per minute. It makes within study comparisons simple and between studies bit more feasible. Qualified search speed is a combination of speed and accuracy measures that reveals the tradeoff between speed and accuracy. It shows the source of speed differences in terms of accuracy and is also measured in answers per minute. Immediate search accuracy is a measure that captures the success of result evaluation when only the first few hits are interesting. These new measures are evaluated by applying them to data from real experiments and comparing them to conventional measures. RELATED WORK In usability evaluations, the measurements are typically based on the three major components of usability: effectiveness, efficiency, and satisfaction [3, 4]. International ISO 9241-11 standard [4] defines effectiveness as the "accuracy and completeness with which the users achieve specified goals" and efficiency as "resources expended in relation to the accuracy and completeness with which users achieve goals". According Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. NordiCHI '04, October 23-27, 2004 Tampere, Finland Copyright 2004 ACM 1-58113-857-1/04/10... $5.00 365 to the standard, efficiency measure divides the effectiveness (achieved results) by the resources used (e.g. time, human effort, or cost). In this work, we will leave satisfaction measures out of the discussion and concentrate on objective quantitative measures. Usability measurements are strongly domain dependent. In the search user interface domain effectiveness is typically measured in terms of accuracy (which is recognized as an example measure in the ISO standard as well). Time (speed of use) is typically used as the critical resource when calculating the efficiency. In the following we will discuss measuring practices in typical studies evaluating search user interfaces. Note that although almost every study in the information retrieval community deals with searching, they tend to focus on system performance [8] and thus only a few studies are mentioned here. Speed Measures The basic approach for measuring the speed is simply to measure the time required for performing a task, but the actual implementation differs from study to study. In early evaluations of the Scatter/Gather system by Pirolli et al. [6], times were recorded simply on a task basis. In the results they reported how many minutes it took, on average, to complete a task. In the study by Dumais et al. [2], roughly the same method was used, except that the times were divided into categories according to the difficulty of the task. Sebrechts et al. [9] used a different categorization method where task execution times were divided into categories according to the subject's computer experience. Time measurements can also be recorded in a somewhat reversed manner as Pratt and Fagan [7] did. They reported how many results users found in four minutes. This is close to measuring speed (achievement / time), but this normalization to four minutes is arbitrary and does not facilitate comparisons optimally. In a study by Dennis et al. [1], the time to bookmark a result page was measured and only one page was bookmarked per task. This setup makes the comparison fairly easy since the reported time tells how much time it takes to find a result with the given user interface. However, this desirable feature was caused by the setup where only one result was chosen, and other types of tasks were not considered. Accuracy Measures Accuracy measures are based on the notion of relevance which is typically determined by independent judges in relation to a task. In information retrieval studies, accuracy is typically a combination of two measures: recall and precision. Recall describes the amount of relevant results found in a search in a relation to all the relevant results in the collection. As a perfect query in terms of recall could return all the entries in the collection, it is counterbalanced with the precision measure. Precision describes how clean the result set is by describing the density of relevant results in it. Precision, like recall, is expressed by a percentage number which states the proportion of relevant targets in the result set. Recall and precision measures are designed for measuring the success of a query. In contrast, when the success of the result evaluation process is studied, the users need to complete the process by selecting the interesting results. Measures are then based on analyzing the true success of the selections. Recall and precision measures are used here too, but the calculation is different. In these cases recall describes the amount of relevant results selected in relation to the amount of them in the result set. Precision, on the other hand, describes the density of relevant results among the selected results. Veerasamy and Heikes [13] used such measures (called interactive recall and interactive precision) in their study of a graphical display of retrieval studies. They asked participants to judge the relevance of the results in order to get the users' idea of the document relevance. Pirolli et al. [6] used only the precision measure in their test of the Scatter/Gather system. The selection of the results was implemented by a save functionality. Dennis et al. [1] used an approach where they reported the average relevance of the results found with a given user interface. Relevant results were indicated by bookmarking them. Further variations of the measures where user interaction is taken into account in accuracy evaluation were proposed and used by Veerasamy and Belkin [12]. Information Foraging Theory Stuart Card, Peter Pirolli and colleagues have made extensive research on information foraging theory [5] in Xerox Parc and the results are relevant here as well. In its conventional form information foraging theory states that the rate of gain of valuable information (R) can be calculated using the formula: W B T T G R + = (1) In the formula, G is the amount of gained information, T B is the total time spent between information patches and T W is the total time spent within an information patch [5]. An information patch is understood to mean a collection of information such as a document collection, a search result collection or even a single document that can be seen to be a collection of information that requires some actions for digesting the information. In the information foraging process, the forager navigates first between patches and then finds actual meaningful information within a patch. The process is then started over by seeking a new patch. If we discard the separation of two different types of activities (between and within patches) for simplicity, equation 1 states the information gain rate in terms of time unit. This matches with common practices in the field and is the basis for our proposed measurements as well. 366 The gap that is left in the information foraging theory in a relation to making concrete measurements, is the definition of information gain. The gap is well justified as the definition would unnecessarily reduce the scope of the theory. On the other hand, when we deal with concrete problems, we can be more specific and thus obtain preciseness. This is our approach here: we apply the basic relationships stated in the information foraging theory and provide meaningful ways of measuring the gain. All this is done in the context of evaluating search user interfaces in the search result evaluation phase. We will get back to this topic in the discussions of the new measures to see their relationship to the information foraging theory in more detail. EXPERIMENT We will evaluate the proposed measures using data from an experiment of ours. This experiment was conducted to evaluate a new search user interface idea by comparing it to the de facto standard solution. Our proposed user interface used automatically calculated categories for facilitating the result access (Figure 1, left). As the categories we used the most common words and phrases found within the result titles and text summaries (snippets). Stop word list and a simple stemmer were used for improving the quality of the categories (e.g. discarding very common words such as `and' or `is'). As the category word (or phrase) selection was based solely on the word frequencies, the categories were neither exclusive nor exhaustive. There was a special built-in category for accessing all the results as one long list. The hypothesis behind the category user interface was such that it would allow users to identify and locate interesting results easier and faster than the conventional solution. The calculated categories were presented to the user as a list beside the actual result list. When a category was selected from the list, the result listing was filtered to display only those result items that contained the selected word or phrase. There were a total of 150 results that the user could access and from which the categories were computed. Participants There were 20 volunteer participants (8 male, 12 female) in the experiment. Their average age was 35 years varying from 19 to 57 years and they were recruited from the local university. Almost all of the participants can be regarded as experienced computer users, but none of them was an information technology professional. Apparatus There were two user interfaces to access the search results: 1. The category interface (category UI, Figure 1, left) presented the users with a list of 15 automatically generated categories on the left side of the user interface. When the user selected a category, the corresponding results were shown on the right side of the user interface much like in popular e-mail clients. 2. The reference interface (reference UI, Figure 1, right) was a Google web search engine imitation showing results in separate pages, ten results per page. The order of the results was defined by the search engine (Google). In the bottom of the window, there were controls to browse the pages in order (Previous and Figure 1. Compared user interfaces in our experiment. Category user interface on the left, reference user interface on the right. 367 Next buttons) or in random order (a radio button for each page). There were 15 pages so that the participants could access a total of 150 results. Design and Procedure The experiment had search user interface as the only independent variable with two values: category UI and reference UI. The values of the independent variable were varied within the subjects and thus the analysis was done using repeated measures tools. As dependent variables we measured: 1) time to accomplish a task in seconds, 2) number of results selected for a task, 3) relevance of selected result in a three step scale (relevant, related, not relevant), and 4) subjective attitudes towards the systems. The experiments were carried out in a usability laboratory. One experiment lasted approximately 45 minutes and contained 18 (9+9) information seeking tasks in two blocks: one carried out with the category interface and the other using the reference interface. The order of the blocks and the tasks were counterbalanced between the participants. For each task, there was a ready-made query and users did not (re)formulate the queries themselves. This kind of restriction in the setup was necessary to properly focus on measuring the success in the result evaluation phase of the search. The actual task of the participant was to "collect as many relevant results for the information seeking task as possible as fast as you can". The participants collected results by using check boxes that were available beside each result item (see Figure 1). In the test situation there were two windows in the computer desktop. The task window displayed information seeking tasks for the participants who were instructed to first read the task description, then push the `Start' button in the task window and promptly proceed to accomplish the task in the search window. Upon task completion (participant's own decision or time-out), the participants were instructed to push the `Done' button in the task window. The time between `Start' and `Done' button presses was measured as the total time for the task. This timing scheme was explained to the participants. Time for each task was limited to one minute. Accuracy measures are based on ratings done by the experimenter (one person). The rating judgments were made based solely on the task description and the very same result title and summary texts that the participants saw in the experiment. Actual result pages were not used because it would have added an extra variable into the design (result summary vs. page relation), which we did not wish. All the tasks had at least two defining concepts like in "Find pictures of planet Mars". For relevant results, all of the concepts was required to be present in some form (different wording was of course allowed). Related results were those where only the most dominant concept was present (e.g. planet Mars). Rest of the results was considered to be not relevant. RESULTS For comparing the proposed measures we present here the results of our experiment using the conventional measures: time, number of results, and precision. The time measure did not reveal very interesting results, because the test setup limited the total time for one task to one minute. Thus the mean times for conditions were close to each other: 56.6 seconds (sd = 5.5) for the category UI and 58.3 seconds (sd = 3.5) for the reference UI. The difference is not statistically significant as repeated measures analysis of variance (ANOVA) gives F(1,19) = 3.65, ns. In contrast, number of results revealed a difference. When using the category UI the participants were able to find on average 5.1 (sd = 2.1) results per task whereas using the reference UI yielded on average 3.9 (sd = 1.2) selections. The difference is significant since ANOVA gives F(1,19) = 9.24, p &lt; .01. Precision measure gave also a statistically significant difference. When using the category UI on average 65% (sd = 13) of the participants' selections were relevant in a relation to the task. The corresponding number for the reference UI was 49% (sd = 15). ANOVA gave F(1,19) = 14.49, p &lt; .01. The results are compatible with other studies done with similar categorizing search user interfaces. For example, Pratt and Fagan [7] have also reported similar results in favor of categorizing user interface. When categories work, they enhance the result evaluation process by reducing the number of items that need to be evaluated. Users find interesting looking categories and evaluate only the results within those categories. Concentration of relevant documents in the interesting categories is higher than in the whole result set. SEARCH SPEED In order to make the comparison of speed measures easier, we suggest a proportional measure. When the search time and number of results are combined into one measure, just like in measuring physical speed by kilometers or miles per hour, we get a search user interface search speed measure expressed in answers per minute (APM). It is calculated by dividing the number of answers found by the time it took to find them: searched minutes found ans/wers speed search = (2) In relation to the ISO-9241-11 standard this is an efficiency measure whereas the plain number of answers is an (simple) effectiveness measure. In terms of information foraging theory, we replace the G term in equation 1 with number of results found and the time is normalized to minutes. This concretizes the rate (R) in equation 1 to be answers per minute. The structure of equations 1 and 2 is essentially the same. 368 Whenever two (or more) measures are reduced into one, there is a risk of loosing relevant information. This is the case here as well. The proposed measure does not make the distinction between a situation where one answer is found in 10 seconds and a situation where four answers are found in 40 seconds. In both cases the speed is 6 answers per minute and the details of the situation are lost in the measurement. However, we feel that speed measure is nevertheless correct also in this case. The situation can be compared to driving 50 km/h for 10 or 40 minutes. The traveled distance is different, but the speed is the same. This means that proposed speed measure does not apply in every situation and attention must be paid in measurement selection. The problem of reducing two measures into one has also been thoroughly discussed by Shumin Zhai [14] in the context of input devices. He points out that reduction of two Fitts' law variables (a and b) in calculating throughput of an input device leads to a measure that is dependent of the task. The same problem does not apply here as our situation is not related to Fitts' law. However, our measure is dependent on the task, but it is not dependent of the used time or the number of results collected like previous measures. Evaluation In order to evaluate the suggested measure it was applied to the results of Scatter/Gather evaluation by Pirolli et al. [6]. In their experiment the task was to find relevant documents for a given topic. The table below summarizes the results (SS = similarity search, SG = scatter/gather): Measurement SS SG Original Time used in minutes 10.10 30.64 Number of answers 16.44 12.26 Search speed Answers per minute 1.62 0.40 The first two rows show the actual numbers reported in the paper while the third row shows the same results in answers per minute. It is arguably easier to understand the relationship between the two user interfaces from the normalized search speed measure. It communicates that the SS condition was roughly four times faster than the SG condition. The relation is hard to see from the original results. In addition, measurements can be easily related to one's own experiences with similar user interfaces because of the normalization. In the second table below, the search speed measure is applied to the data from our own experiment. Here the difference between raw numbers and normalized measure is not as large as in the previous example because the time used for the tasks is roughly the same in both cases due to the test setup. Nevertheless, the suggested measure makes the comparison easier. Note also that the fairly large difference with the speeds in the experiment by Pirolli et al. is presumably due to experiment set-up (tasks, conditions, equipment, etc.). Measurement Category UI Reference UI Raw numbers Time used in minutes 0.94 0.97 Number of answers 5.1 3.9 Search speed Answers per minute 5.4 4.0 When an analysis of variance is calculated on the answers per minute measure, we see a bit stronger result compared to the conventional measures where just the number of results revealed significant difference. Here ANOVA gives F(1,19) = 11.3, p &lt; .01. Slight increase in the F statistic is due to the combination of two measures that both have a difference in the same direction. In summary, search speed measures the same phenomena as the previously used measures (it is calculated from the same numbers) and it can make distinctions between the measured objects. QUALIFIED SEARCH SPEED Previously used recall and precision measures do not directly tell where possible speed differences come from or what the relation between speed and accuracy is. The suggested qualified search speed measure refines the search speed measure with categories of relevance to address this shortcoming. To keep the measure understandable and robust, we use only two or three categories of relevance. Like the previous measure, the qualified search speed is also measured in answers per minute, with a distinction that the speed is calculated separately for each relevance category according to the equation 3. There RCi stands for relevance category i (typical categories are e.g. relevant and irrelevant). searched minutes found answers speed search qualified RCi RCi = (3) Note that the sum over all relevance categories equals to the normal search speed. When qualified search speed is described in information foraging terminology, we can see that the gain is now defined more precisely than with search speed. While search speed takes into account only the number of results, qualified search speed adds the quality of the results into the equation. In essence, this gives us a more accurate estimate of the gain of information, and thus a more accurate rate of information gain. Note that this shows also in the rate magnitude: rate is now stated in (e.g.) number of relevant results per minute. Evaluation When the qualified search speed measure is applied to the data of our experiment and compared to the simple measure of precision, a few observations can be made. First, the proposed measure preserves the statistically significant 369 difference that was observed with the conventional precision measure. ANOVA for the speed of acquiring relevant results gives F(1,19) = 32.4, p &lt; .01. Second, both measures (Figure 2) convey roughly the same information about the precision of the user interfaces including: 1) with the category UI more than half of the selected results were relevant whereas with the reference UI about half of the results were relevant, and 2) using the category UI participants were more successful in terms of precision. However, with the suggested qualified search speed measure, the amplitude of difference in precision is not obvious and thus the new measure cannot replace the old one. Third, in addition to what can be seen in the precision chart, the qualified search speed chart (Figure 2) reveals some interesting data. It shows that the improvement in speed is due to the fact that participants have been able to select more relevant results while the proportion of not relevant results decreased a bit. The same information could surely be acquired by combining conventional speed and precision measures, but when the information is visible in one figure it is arguably easier to find such a relationship. Note also that although the new measure is mainly concerned about the accuracy of use, it informs the reader simultaneously about the speed of use as well. Figure 3 makes a comparison between the new measure and the original precision measure using the data collected in the Scatter/Gather experiment [6]. Here it is worthwhile to note that even though precision measures are close to those in the previous example, the qualified search speed measure reveals large differences between the conditions. Qualified search speed seems to reveal the tradeoff between accuracy and speed convincingly in this case. We can also notice that both conditions here are much slower than those in Figure 2 as the qualified search speed is normalized just like the simpler search speed. It is notable that qualified search speed does not measure the same phenomena as precision and thus they are not replaceable. We can image a situation where high qualified speed is associated with low precision and vice versa. In reality this could happen when users try to be very precise in one condition and very fast in another. On the other hand, we saw that qualified evaluation speed can make clear distinctions between user interfaces, which is a compulsory quality for a useful measure. IMMEDIATE ACCURACY The last suggested measure captures the success of typical web search behavior. In such a task, the user wants to find a piece of information that would be good enough for an information need and overall speed and accuracy are not as important as quick success. The measure is called immediate accuracy and it is expressed as a success rate. The success rate states the proportion of cases where at least one relevant result is found by the n th selection. For applying the measure, the order of each result selection must be stored and the relevance of them must be judged against the task. The selections for each task and participant are then gone through in the order they were made and the frequency of first relevant result finding is calculated for each selection (first, second, and so on). When this figure is divided by the total number of observations (number of participants * number of tasks) we get the percentage of first relevant result found per each selection. Equation 4 shows the calculation more formally, there n stands for n th selection. ns observatio of number total n results relevant first of number n accuracy immediate = (4) When the figures calculated with equation 4 are plotted into a cumulative line chart (Figure 4) we can see when at least one relevant result is found on average. For example, (in Figure 4) after the second selection in 79 % of the cases at least one relevant result is found when using the category user interface. Notice also that the lines do not reach the 100 % value. This means that in some of the cases the users were not able to find any relevant results. When looking back to information foraging theory, this measure takes us to a different approach compared to the previous ones. This measure abandons time as the limiting Figure 3. Qualified search speed measure compared to precision measure in the Scatter/Gather study [4]. Figure 2. Qualified search speed measure compared to precision measure of data gathered in our own study. 370 resource against which the gain is compared and replaces it by selection ordinal (remember that ISO standard leaves the choice of resource up to the domain). As this new resource is discrete in nature, the expression of the measure as a single figure (rate) becomes hard and thus, for example, a cumulative chart is preferred for easily interpretable figures. From another perspective of information foraging theory, we can say that immediate accuracy is a measure for estimating the beginning of the within patch gain slope. Note, that it is only an estimation of the beginning of the slope as all subsequent relevant selections are discarded in this measure. In this view, we define an information patch to be a search result set. Evaluation The evaluation is based only on our own data because the measure requires information that is typically not reported in the publications. Figure 4 shows that the user orientates faster while using the category UI as the first selection produces already a relevant result in 56 % of the cases. In contrast, the reference UI produces a relevant result in 40 % of the first selections. By the second selection, the difference is bit greater since in 79 % of the cases the users have found at least one relevant result with the category UI, while the corresponding number for the reference UI is 62 %. In the analysis of cumulative data, the most interesting points are those where the difference between compared measurements changes. Change points are most important because cumulative figures will preserve the difference if not further changes happen. In our case the difference is made at the first selection and remains virtually the same afterwards. This difference is statistically significant as ANOVA gives F(1,19) = 12.5, p &lt; .01 and it is preserved throughout the selections (F(1,19) 10.4, p &lt; .01 for all subsequent selections). Findings of Spink et al. [11] stated that users only select one or two results per query. Immediate accuracy allows us to see the success of the studied user interface in such a case. We can focus on a given selection and quickly see the success rate at that point. Note that this kind of information is not available using the conventional accuracy measures and straightforward speed measures. Immediate Success Speed Another fairly simple and obvious way for measuring immediate success would be to record the time to the first relevant result. We did try this measure as well, but found a problem. In our experiment, the average time to find the first relevant result was practically the same in both cases (20 and 21 seconds for category and reference UI respectively) and there was no statistically significant difference. This could, of course, be the true situation, but the amount of relevant results suggested the opposite. The problem comes from the fact that the first relevant result is not always found. With the category UI users were not able to find a single relevant result for a task in 10% of the cases whereas the same number for reference UI was 21%. We felt that this is a big difference and that it should be visible in the measurement as well. However, we were not able to come up with a reasonable solution for normalizing the time measurement in this respect and thus the measurement is not promoted as such. In addition, the results of Spink et al. [11] suggest that the time to first relevant result is not very important for the search process. Since searchers tend to open only one or two results, the time does not seem to be the limiting factor, but the number of result selections is. This supports also the choice of immediate accuracy over the time to the first relevant result. DISCUSSION Our goal was to provide search user interface designers, researchers, and evaluators with additional measures that would complement the current ones. The first problem with them is that result comparison is hard, even within one experiment. Proportional measures makes within study comparisons easy and in addition they let readers relate their previous experience better to the presented results. We proposed normalized search speed measure that is expressed in answers per minute. As the measure combines two figures (number of answers and time searched) into one proportional number, it makes the comparisons within an experiment easy and between experiments bit more feasible. The second shortcoming of the current measures is the fact that it is difficult to see the tradeoff between speed and accuracy. To address this problem, we proposed the qualified search speed measure that divides the search Immediate Accuracy 0 % 20 % 40 % 60 % 80 % 100 % 1. 2. 3. 4. 5. 6. 7. 8. selection success rate Category UI Reference UI Figure 4. Immediate accuracy of category UI and reference UI. The measure shows the proportion of the cases where a relevant result have been found at n th selection. 371 speed measure into relevance categories. The measure allows readers to see what the source of speed in terms of accuracy is. In the evaluation we showed that conventional measures may only tell the half of the story. For instance, in the case of the Scatter/Gather experiment the precision measure showed only moderate difference between the systems whereas qualified speed revealed a vast difference in the gain of relevant results. Combining speed and accuracy measures is particularly effective in such a case as it eliminates the need to mentally combine the two measures (speed and accuracy). The third weakness of the current measures is their inability to capture users' success in typical web search behavior where the first good enough result is looked for. We proposed the immediate accuracy measure to solve this flaw. Immediate accuracy shows the proportion of the cases where the users are able to find at least one relevant result per n th result selection. It allows readers to see how well and how fast the users can orient themselves to the task with the given user interface. As the measurements are made based on finding the first relevant result, the reader can compare how well different solutions support users' goal of finding the first relevant answer (and presumably few others as well) to the search task. The proposed measures are not intended to replace the old measures, but rather to complement them. They lessen the mental burden posed to the reader as important information of different type (e.g. speed, accuracy) is combined into one proportional measure. In summary, the proposed measures capture important characteristics of search user interface usability and communicate them effectively. The issue of making comparisons between experiments is not completely solved by these new measures. We feel that the problem is not in the properties of the new measures but in the nature of the phenomena to be measured. In the context of search user interfaces the test settings have a huge effect on the results that cannot be solved simply with new measures. One solution for the problem could be test setup standardization. In the TREC interactive track such an effort have been taken, but it seems that the wide variety of research questions connected to searching cannot be addressed with a single standard test setup. ACKNOWLEDGMENTS This work was supported by the Graduate School in User Centered Information Technology (UCIT). I would like to thank Scott MacKenzie, Kari-Jouko Rih, Poika Isokoski, and Natalie Jhaveri for invaluable comments and encouragement. REFERENCES 1. Dennis, S., Bruza, P., McArthur, R. Web Searching: A Process-Oriented Experimental Study of Three Interactive Search Paradigms. Journal of the American Society for Information Science and Technology, Vol. 53, No. 2, 2002, 120-133. 2. Dumais, S., Cutrell, E., Chen, H. Optimizing Search by Showing Results in Context. Proceedings of ACM CHI'01 (Seattle, USA), ACM Press, 2001, 277-284. 3. Frkjr, E., Hertzum, M., Hornbk, K. Measuring Usability: Are Effectiveness, Efficiency, and Satisfaction Really Correlated? Proceedings of ACM CHI'2000 (The Hague, Netherlands), ACM Press, 2000, 345-352. 4. ISO 9241-11: Ergonomic requirements for office work with visual display terminals (VDTs) - Part 11: Guidance on usability, International Organization for Standardization, March 1998. 5. Pirolli, P. and Card, S. Information Foraging. Psychological Review, 1999, Vol. 106, No. 4, 643-675. 6. Pirolli, P., Schank, P., Hearst, M., Diehl, C. Scatter/Gather Browsing Communicates the Topic Structure of a Very Large Text Collection. Proceedings of ACM CHI'96 (Vancouver, Canada), ACM Press, 1996, 213-220. 7. Pratt, W., Fagan, L. The Usefulness of Dynamically Categorizing Search Results. Journal of the American Medical Informatics Association, Vol. 7, No. 6, Nov/Dec 2000, 605-617. 8. Saracevic, T. Evaluation of Evaluation in Information Retrieval. Proceedings of ACM SIGIR'95 (Seattle, USA), ACM Press, 1995, 138-146. 9. Sebrechts, M., Vasilakis, J., Miller, M., Cugini, J., Laskowski, S. Visualization of Search Results: A Comparative Evaluation of Text, 2D, and 3D Interfaces. Proceedings of ACM SIGIR'99 (Berkeley, USA), ACM Press, 1999. 10. Shneiderman, B., Byrd, D., Croft, B. Clarifying Search: A User-Interface Framework for Text Searches. D-Lib Magazine, January 1997. 11. Spink, A., Wolfram, D., Jansen, M., and Saracevic, T.: Searching the Web: The Public and Their Queries. Journal of the American Society for Information Science and Technology, 2001, Vol. 52, No. 6, 226-234. 12. Veerasamy, A., Belkin, N. Evaluation of a Tool for Visualization of Information Retrieval Results. Proceedings of ACM SIGIR'96 (Zurich, Switzerland), ACM Press, 1996, 85-92. 13. Veerasamy, A., Heikes, R. Effectiveness of a Graphical Display of Retrieval Results. Proceedings of ACM SIGIR'97 (Philadelphia, USA), ACM Press, 1997, 236-244 . 14. Zhai, S. On the validity of Throughput as a Characteristic of Computer Input. IBM Research Report, RJ 10253, IBM Research Division. August 2002. 372

usability evaluation;Search user interface;speed;usability measure;accuracy

153

Protected Interactive 3D Graphics Via Remote Rendering

Valuable 3D graphical models, such as high-resolution digital scans of cultural heritage objects, may require protection to prevent piracy or misuse, while still allowing for interactive display and manipulation by a widespread audience. We have investigated techniques for protecting 3D graphics content, and we have developed a remote rendering system suitable for sharing archives of 3D models while protecting the 3D geometry from unauthorized extraction . The system consists of a 3D viewer client that includes low-resolution versions of the 3D models, and a rendering server that renders and returns images of high-resolution models according to client requests. The server implements a number of defenses to guard against 3D reconstruction attacks, such as monitoring and limiting request streams, and slightly perturbing and distorting the rendered images. We consider several possible types of reconstruction attacks on such a rendering server, and we examine how these attacks can be defended against without excessively compromising the interactive experience for non-malicious users.

Protecting digital information from theft and misuse, a subset of the digital rights management problem, has been the subject of much research and many attempted practical solutions. Efforts to protect software, databases, digital images, digital music files, and other content are ubiquitous, and data security is a primary concern in the design of modern computing systems and processes. However, there have been few technological solutions to specifically protect interactive 3D graphics content. The demand for protecting 3D graphical models is significant. Contemporary 3D digitization technologies allow for the reliable and efficient creation of accurate 3D models of many physical objects, and a number of sizable archives of such objects have been created. The Stanford Digital Michelangelo Project [Levoy et al. 2000], for example, has created a high-resolution digital archive of 10 large statues of Michelangelo, including the David. These statues represent the artistic patrimony of Italy's cultural institutions, and the contract with the Italian authorities permits the distribution of the 3D models only to established scholars for non-commercial use. Though all parties involved would like the models to be widely available for constructive purposes, were the digital 3D model of the David to be distributed in an unprotected fashion, it would soon be pirated, and simulated marble replicas would be manufactured outside the provisions of the parties authorizing the creation of the model. Digital 3D archives of archaeological artifacts are another example of 3D models often requiring piracy protection. Curators of such artifact collections are increasingly turning to 3D digitization as a way to preserve and widen scholarly usage of their holdings, by allowing virtual display and object examination over the Internet, for example. However, the owners and maintainers of the artifacts often desire to maintain strict control over the use of the 3D data and to guard against theft. An example of such a collection is [Stanford Digital Forma Urbis Project 2004], in which over one thousand fragments of an ancient Roman map were digitized and are being made available through a web-based database, providing that the 3D models can be adequately protected. Other application areas such as entertainment and online commerce may also require protection for 3D graphics content. 3D character models developed for use in motion pictures are often repurposed for widespread use in video games and promotional materials. Such models represent valuable intellectual property, and solutions for preventing their piracy from these interactive applications would be very useful. In some cases, such as 3D body scans of high profile actors, content developers may be reluctant to distribute the 3D models without sufficient control over reuse. In the area of online commerce, a number of Internet content developers have reported an unwillingness of clients to pursue 3D graphics projects specifically due to the lack of ability to prevent theft of the 3D content [Ressler 2001]. Prior technical research in the area of intellectual property protections for 3D data has primarily concentrated on 3D digital watermarking techniques. Over 30 papers in the last 7 years describe steganographic approaches to embedding hidden information into 3D graphical models, with varying degrees of robustness to attacks that seek to disable watermarks through alterations to the 3D shape or data representation. Many of the most successful 3D watermarking schemes are based on spread-spectrum frequency domain transformations, which embed watermarks at multiple scales by introducing controlled perturbations into the coordinates of the 3D model vertices [Praun et al. 1999; Ohbuchi et al. 2002]. Complementary technologies search collections of 3D models and examine them for the presence of digital watermarks, in an effort to detect piracy. We believe that for the digital representations of highly valuable 3D objects such as cultural heritage artifacts, it is not sufficient to detect piracy after the fact; we must instead prevent it. The computer industry has experimented with a number of techniques for preventing unauthorized use and copying of computer software and digital data. These techniques have included physical dongles, software access keys, node-locked licensing schemes, copy prevention software, program and data obfuscation, and encryption with embedded keys. Most such schemes are either broken or bypassed by determined attackers, and cause undue inconvenience and expense for non-malicious users. High-profile data and software is particularly susceptible to being quickly targeted by attackers. Fortunately, 3D graphics data differs from most other forms of digital media in that the presentation format, 2D images, is fundamen-tally different from the underlying representation (3D geometry). Usually, 3D graphics data is displayed as a projection onto a 2D display device, resulting in tremendous information loss for single views. This property supports an optimistic view that 3D graphics systems can be designed that maintain usability and utility, while not being as vulnerable to piracy as other types of digital content. In this paper, we address the problem of preventing the piracy of 3D models, while still allowing for their interactive display and manipulation . Specifically, we attempt to provide a solution for maintainers of large collections of high-resolution static 3D models, such as the digitized cultural heritage artifacts described above. The methods we develop aim to protect both the geometric shape of the 3D models, as well as their particular geometric representation, such as the 3D mesh vertex coordinates, surface normals, and connectivity information. We accept that the coarse shape of visible objects can be easily reproduced regardless of our protection efforts, so we concentrate on defending the high-resolution geometric details of 3D models, which may have been most expensive to model or measure (perhaps requiring special access and advanced 3D digitizing technology), and which are most valuable in exhibiting fidelity to the original object. In the following paper sections, we first examine the graphics pipeline to identify its possible points of attack, and then propose several possible techniques for protecting 3D graphics data from such attacks. Our experimentation with these techniques led us to conclude that remote rendering provides the best solution for protecting 3D graphical models, and we describe the design and implementation of a prototype system in Section 4. Section 5 describes some types of reconstruction attacks against such a remote rendering system and the initial results of our efforts to guard against them. Possible Attacks in the Graphics Pipeline Figure 1 shows a simple abstraction of the graphics pipeline for purposes of identifying possible attacks to recover 3D geometry. We note several places in the pipeline where attacks may occur: 3D model file reverse-engineering. Fig. 1(a). 3D graphics models are typically distributed to users in data streams such as files in common file formats. One approach to protecting the data is to obfuscate or encrypt the data file. If the user has full access to the data file, such encryptions can be reverse-engineered and broken, and the 3D geometry data is then completely unprotected. Tampering with the viewing application. Fig. 1(b). A 3D viewer application is typically used to display the 3D model and allow for its manipulation. Techniques such program tracing, memory dumping , and code replacement are practiced by attackers to obtain access to data in use by application programs. Graphics driver tampering. Fig. 1(c). Because the 3D geometry usually passes through the graphics driver software on its way to the GPU, the driver is vulnerable to tampering. Attackers can replace graphics drivers with malicious or instrumented versions to capture streams of 3D vertex data, for example. Such replacement drivers are widely distributed for purposes of tracing and debugging graphics programs. Reconstruction from the framebuffer. Fig. 1(d). Because the framebuffer holds the result of the rendered scene, its contents can be used by sophisticated attackers to reconstruct the model geometry , using computer vision 3D reconstruction techniques. The Figure 1: Abstracted graphics pipeline showing possible attack locations (a-e). These attacks are described in the text. framebuffer contents may even include depth values for each pixel, and attackers may have precise control over the rendering parameters used to create the scene (viewing and projection transformations , lighting, etc.). This potentially creates a perfect opportunity for computer vision reconstruction, as the synthetic model data and controlled parameters do not suffer from the noise, calibration, and imprecision problems that make robust real world vision with real sensors very difficult. Reconstruction from the final image display. Fig. 1(e). Regardless of whatever protections a graphics system can guarantee throughout the pipeline, the rendered images finally displayed to the user are accessible to attackers. Just as audio signals may be recorded by external devices when sound is played through speakers , the video signals or images displayed on a computer monitor may be recorded with a variety of video devices. The images so gathered may be used as input to computer vision reconstruction attacks such as those possible when the attacker has access to the framebuffer itself, though the images may be of degraded quality, unless a perfect digital video signal (such as DVI) is available. Techniques for Protecting 3D Graphics In light of the possible attacks in the graphics pipeline as described in the previous section, we have considered a number of approaches for sharing and rendering protected 3D graphics. Software-only rendering. A 3D graphics viewing system that does not make use of hardware acceleration may be easier to protect from the application programmer's point of view. Displaying graphics with a GPU can require transferring the graphics data in precisely known and open formats, through a graphics driver and hardware path that is often out of the programmer's control. A custom 3D viewing application with software rendering allows the 3D content distributor to encrypt or obfuscate the data in a specific manner, all the way through the graphics pipeline until display. Hybrid hardware/software rendering. Hybrid hardware and software rendering schemes can be used to take at least some advantage of hardware accelerated rendering, while benefiting from software rendering's protections as described above. In one such scheme, a small but critically important portion of a protected model's geometry (such as the nose of a face) is rendered in software, while the rest of the model is rendered normally with the accelerated GPU hardware. This technique serves as a deterrent to attackers tampering with the graphics drivers or hardware path, but the two-phase drawing with readback of the color and depth buffers can incur a 696 performance hit, and may require special treatment to avoid artifacts on the border of the composition of the two images. In another hybrid rendering scheme, the 3D geometry is transformed and per-vertex lighting computations are performed in software . The depth values computed for each vertex are distorted in a manner that still preserves the correct relative depth ordering, while concealing the actual model geometry as much as possible. The GPU is then used to complete rendering, performing rasteri-zation , texturing, etc. Such a technique potentially keeps the 3D vertex stream hidden from attackers, but the distortions of the depth buffer values may impair certain graphics operations (fog computation , some shadow techniques), and the geometry may need to be coarsely depth sorted so that Z-interpolation can still be performed in a linear space. Deformations of the geometry. Small deformations in large 2D images displayed on the Internet are sometimes used as a defense against image theft; zoomed higher resolution sub-images with varying deformations cannot be captured and easily reassembled into a whole. A similar idea can be used with 3D data: subtle 3D deformations are applied to geometry before the vertices are passed to the graphics driver. The deformations are chosen so as to vary smoothly as the view of the model changes, and to prohibit recovery of the original coordinates by averaging the deformations over time. Even if an attacker is able to access the stream of 3D data after it is deformed, they will encounter great difficulty reconstructing a high-resolution version of the whole model due to the distortions that have been introduced. Hardware decryption in the GPU. One sound approach to providing for protected 3D graphics is to encrypt the 3D model data with public-key encryption at creation time, and then implement custom GPUs that accept encrypted data, and perform on-chip decryption and rendering. Additional system-level protections would need to be implemented to prevent readback of framebuffer and other video memory, and to place potential restrictions on the command stream sent to the GPU, in order to prevent recovery of the 3D data. Image-based rendering. Since our goal is to protect the 3D geometry of graphic models, one technique is to distribute the models using image-based representations, which do not explicitly include the complete geometry data. Examples of such representations include light fields and Lumigraphs [Levoy and Hanrahan 1996; Gortler et al. 1996], both of which are highly amenable to interactive display. Remote rendering. A final approach to secure 3D graphics is to retain the 3D model data on a secure server, under the control of the content owner, and pass only 2D rendered images of the models back to client requests. Very low-resolution versions of the models, for which piracy is not a concern, can be distributed with special client programs to allow for interactive performance during manipulation of the 3D model. This method relies on good network bandwidth between the client and server, and may require significant server resources to do the rendering for all client requests, but it is vulnerable primarily only to reconstruction attacks. Discussion. We have experimented with several of the 3D model protection approaches described above. For example, our first protected 3D model viewer was an encrypted version of the "QS-plat" [Rusinkiewicz and Levoy 2000] point-based rendering system , which omits geometric connectivity information. The 3D model files were encrypted using a strong symmetric block cipher scheme, and the decryption key was hidden in a heavily obfuscated 3D model viewer program, using modern program obfuscation techniques [Collberg and Thomborson 2000]. Vertex data was decrypted on demand during rendering, so that only a very small portion of the decrypted model was ever in memory, and only software rendering modes were used. Unfortunately, systems such as this ultimately rely on "security through obfuscation," which is theoretically unsound from a computer security point of view. Given enough time and resources, an attacker will be able to discover the embedded encryption key or otherwise reverse-engineer the protections on the 3D data. For this reason, any of the 3D graphics protection techniques that make the actual 3D data available to potential attackers in software can be broken [Schneier 2000]. It is possible that future "trusted comput-ing" platforms for general purpose computers will be available that make software tampering difficult or impossible, but such systems are not widely deployed today. Similarly, the idea of a GPU with decryption capability has theoretical merit, but it will be some years before such hardware is widely available for standard PC computing environments, if ever. Thus, for providing practical, robust, anti-piracy protections for 3D data, we gave strongest consideration to purely image-based representations and to remote rendering. Distributing light fields at the high resolutions necessary would involve huge, unwieldy file sizes, would not allow for any geometric operations on the data (such as surface measurements performed by archaeologists), and would still give attackers unlimited access to the light field for purposes of performing 3D reconstruction attacks using computer vision algorithms. For these reasons, we finally concluded that the last technique, remote rendering, offers the best solution for protecting interactive 3D graphics content. Remote rendering has been used before in networked environments for 3D visualization, although we are not aware of a system specifically designed to use remote rendering for purposes of security and 3D content protection. Remote rendering systems have been previously implemented to take advantage of off-site specialized rendering capabilities not available in client systems, such as intensive volume rendering [Engel et al. 2000], and researchers have developed special algorithmic approaches to support efficient distribution of rendering loads and data transmission between rendering servers and clients [Levoy 1995; Yoon and Neumann 2000]. Remote rendering of 2D graphical content is common for Internet services such as online map sites; only small portions of the whole database are viewed by users at one time, and protection of the entire 2D data corpus from theft via image harvesting may be a factor in the design of these systems. Remote Rendering System To test our ideas for providing controlled, protected interactive access to collections of 3D graphics models, we have implemented a remote rendering system with a client-server architecture, as described below. 4.1 Client Description Users of our protected graphics system employ a specially-designed 3D viewing program to interactively view protected 3D content . This client program is implemented as an OpenGL and wxWindows-based 3D viewer, with menus and GUI dialogs to control various viewing and networking parameters (Figure 2). The client program includes very low-resolution, decimated versions of the 3D models, which can be interactively rotated, zoomed, and re-lit by the user in real-time. When the user stops manipulating the low-resolution model, detected via a "mouse up" event, the client program queries the remote rendering server via the network for a 697 Figure 2: Screenshot of the client program. high-resolution rendered image corresponding to the selected rendering parameters. These parameters include the 3D model name, viewpoint position and orientation, and lighting conditions. When the server passes the rendered image back to the client program, it replaces the low-resolution rendering seen by the user (Figure 3). On computer networks with reasonably low latencies, the user thus has the impression of manipulating a high-resolution version of the model. In typical usage for cultural heritage artifacts, we use models with approximately 10,000 polygons for the low resolution version, whereas the server-side models often contain tens of millions polygons. Such low-resolution model complexities are of little value to potential thieves, yet still provide enough clues for the user to navigate. The client viewer could be further extended to cache the most recent images returned from the server and projec-tively texture map them onto the low-resolution model as long as they remain valid during subsequent rotation and zooming actions. 4.2 Server Description The remote rendering server receives rendering requests from users' client programs, renders corresponding images, and passes them back to the clients. The rendering server is implemented as a module running under the Apache 2.0 HTTP Server; as such, the module communicates with client programs using the standard HTTP protocol, and takes advantage of the wide variety of access protection and monitoring tools built into Apache. The rendering server module is based upon the FastCGI Apache module, and allows for multiple rendering processes to be spread across any number of server hardware nodes. As render requests are received from clients, the rendering server checks their validity and dispatches the valid requests to a GPU for OpenGL hardware-accelerated rendering. The rendered images are read back from the framebuffer, compressed using JPEG compression , and returned to the client. If multiple requests from the same client are pending (such as if the user rapidly changes views while on a slow network), earlier requests are discarded, and only the most recent is rendered. The server uses level-of-detail techniques to speed the rendering of highly complex models, and lower level-of -detail renderings can be used during times of high server load to maintain high throughput rates. In practice, an individual server node with a Pentium 4 CPU and an NVIDIA GeForce4 video card can handle a maximum of 8 typical client requests per second; the Figure 3: Client-side low resolution (left) and server-side high resolution (right) model renderings. bottlenecks are in the rendering and readback (about 100 milliseconds ), and in the JPEG compression (approximately 25 milliseconds ). Incoming request sizes are about 700 bytes each, and the images returned from our deployed servers average 30 kB per request . 4.3 Server Defenses In Section 2, we enumerated several possible places in the graphics pipeline that an attacker could steal 3D graphics data. The benefit of using remote rendering is that it leaves only 3D reconstruction from 2D images in the framebuffer or display device as possible attacks. General 3D reconstruction from images constitutes a very difficult computer vision problem, as evidenced by the great amount of research effort being expended to design and build robust computer vision systems. However, synthetic 3D graphics renderings can be particularly susceptible to reconstruction because the attacker may be able to exactly specify the parameters used to create the images, there is a low human cost to harvest a large number of images, and synthetic images are potentially perfect, with no sensor noise or miscalibration errors. Thus, it is still necessary to defend the remote rendering system from reconstruction attacks; below, we describe a number of such defenses that we have implemented in combination for our server. Session-based defenses. Client programs that access the remote rendering system are uniquely identified during the course of a usage session. This allows the server to monitor and track the specific sequence of rendering requests made by each client. Automatic analysis of the server logs allows suspicious request streams to be classified, such as an unusually high number of requests per unit time, or a particular pattern of requests that is indicative of an image harvesting program. High quality computer vision reconstructions often require a large number of images that densely sample the space of possible views, so we are able to effectively identify such access patterns and terminate service to those clients. We can optionally require recurrent user authentication in order to further deter some image harvesting attacks, although a coalition of users mounting a low-rate distributed attack from multiple IP addresses could still defeat such session-based defenses. Obfuscation. Although we do not rely on obfuscation to protect the 3D model data, we do use obfuscation techniques on the client side of the system to discourage and slow down certain attacks. The low-resolution models that are distributed with the client viewer program are encrypted using an RC4-variant stream cipher, and the keys are embedded in the viewer and heavily obfuscated. The rendering request messages sent from the client to the server are also encrypted with heavily obfuscated keys. These encryptions simply serve as another line of defense; even if they were broken, attackers would still not be able to gain access to the high resolution 3D data except through reconstruction from 2D images. 698 Limitations on valid rendering requests. As a further defense, we provide the capability in our client and remote server to constrain the viewing conditions. Some models may have particular "stayout" regions defined that disallow certain viewing and lighting angles, thus keeping attackers from being able to reconstruct a complete model. For the particular purpose of defending against the enumeration attacks described in Section 5.1, we put restrictions on the class of projection transformations allowed to be requested by users (requiring a perspective projection with particular fixed field of view and near and far planes), and we prevent viewpoints within a small offset of the model surface. Perturbations and distortions. Passive 3D computer vision reconstructions of real-world objects from real-world images are usually of relatively poor quality compared to the original object. This failure inspires the belief that we can protect our synthetically rendered models from reconstruction by introducing into the images the same types of obstacles known to plague vision algorithms. The particular perturbations and distortions that we use are described below; we apply these defenses to the images only to the degree that they do not distract the user viewing the models. Additionally, these defenses are applied in a pseudorandomly generated manner for each different rendering request, so that attackers cannot systematically determine and reverse their effects, even if the specific form of the defenses applied is known (such as if the source code for the rendering server is available). Rendering requests with identical parameters are mapped to the same set of perturbations, in order to deter attacks which attempt to defeat these defenses by averaging multiple images obtained under the same viewing conditions. Perturbed viewing parameters We pseudorandomly introduce subtle perturbations into the view transformation matrix for the images rendered by the server; these perturbations have the effect of slightly rotating, translating, scaling, and shearing the model. The range of these distortions is bounded such that no point in the rendered image is further than either m object space units or n pixels from its corresponding point in an unperturbed view. In practice, we generally set m proportional to the size of the model's geometry being protected, and use values of n = 15 pixels, as experience has shown that users can be distracted by larger shifts between consecutively displayed images. Perturbed lighting parameters We pseudorandomly introduce subtle perturbations into the lighting parameters used to render the images; these perturbations include modifying the lighting direction specified in the client request, as well as addition of randomly changing secondary lighting to illuminate the model. Users are somewhat sensitive to shifts in the overall scene intensity and shading, so the primary light direction perturbations used are generally fairly small (maximum of 10 for typical models, which are rendered using the OpenGL local lighting model). High-frequency noise added to the images We introduce two types of high-frequency noise artifacts into the rendered images. The first, JPEG artifacts, are a convenient result of the compression scheme applied to the images returned from the server. At high compression levels (we use a maximum libjpeg quality factor of 50), the quantization of DCT coefficients used in JPEG compression creates "blocking" discontinuities in the images, and adds noise in areas of sharp contrast. These artifacts create problems for low-level computer vision image processing algorithms, while the design of JPEG compression specifically seeks to minimize the overall perceptual loss of image quality for human users. Additionally, we add pseudorandomly generated monochromatic Gaussian noise to the images, implemented efficiently by blending noise textures during hardware rendering on the server. The added noise defends against computer vision attacks by making background segmentation more difficult, and by breaking up the highly regular shading patterns of the synthetic renderings. Interestingly, users are not generally distracted by the added noise, but have even commented that the rendered models often appear "more realistic" with the high-frequency variations caused by the noise. One drawback of the added noise is that the increased entropy of the images can result in significantly larger compressed file sizes; we address this in part by primarily limiting the application of noise to the non-background regions of the image via stenciled rendering. Low-frequency image distortions Just as real computer vision lens and sensor systems sometimes suffer from image distortions due to miscalibration, we can effectively simulate and extend these distortions in the rendering server. Subtle non-linear radial distortions, pinching, and low-frequency waves can be efficiently implemented with vertex shaders, or with two-pass rendering of the image as a texture onto a non-uniform mesh, accelerated with the "render to texture" capabilities of modern graphics hardware. Due to the variety of random perturbations and distortions that are applied to the images returned from the rendering server, there is a risk of distracting the user, as the rendered 3D model exhibits changes from frame to frame, even when the user makes very minor adjustments to the view. However, we have found that the brief switch to the lower resolution model in between display of the high resolution perturbed images, inherent to our remote rendering scheme, very effectively masks these changes. This masking of changes is attributed to the visual perception phenomenon known as change blindness [Simons and Levin 1997], in which significant changes occurring in full view are not noticed due to a brief disruption in visual continuity, such as a "flicker" introduced between successive images. Reconstruction Attacks In this section we consider several classes of attacks, in which sets of images may be gathered from our remote rendering server to make 3D reconstructions of the model, and we analyze their efficacy against the countermeasures we have implemented. 5.1 Enumeration Attacks The rendering server responds to rendering requests from users specifying the viewing conditions for the rendered images. This ability for precise specification can be exploited by attackers, as they can potentially explore the entire 3D model space, using the returned images to discover the location of the 3D model to any arbitrary precision. In practice, these attacks involve enumerating many small cells in a voxel grid, and testing each such voxel to determine intersection with the remote high-resolution model's surface; thus we term them enumeration attacks. Once this enumeration process is complete, occupied cells of the voxel grid are exported as a point cloud and then input to a surface reconstruction algorithm. In the plane sweep enumeration attack, the view frustum is specified as a rectangular, one-voxel-thick "plane," and is swept over the model (Figure 4(a)). Each requested image represents one slice of the model's surface, and each pixel of each image corresponds to a single voxel. A simple comparison of each image pixel against the expected background color is performed to determine whether that 699 (a) (b) Figure 4: Enumeration Attacks: (a) the plane sweep enumeration attack sweeps a one-voxel thick orthographic view frustum over the model, (b) the near plane sweep enumeration attack sweeps the viewpoint over the model, marking voxels where the model surface is clipped by the near plane. pixel is a model surface or background pixel. Sweeps from multiple view angles (such as the six faces of the voxels) are done to catch backfacing polygons that may not be visible from a particular angle. These redundant multiple sweeps also allow the attacker to be liberal about ignoring questionable background pixels that may occur, such as if low-amplitude background noise or JPEG compression is being used as a defense on the server. Our experiments demonstrate that the remote model can be efficiently reconstructed against a defenseless server using this attack (Figure 5(b)). Perturbing viewing parameters can be an effective defense against this attack; the maximum reconstruction resolution will be limited by the maximum relative displacement that an individual model surface point undergoes. Figure 5(c) shows the results of a reconstruction attempt against a server pseudorandomly perturbing the viewing direction by up to 0.3 in the returned images . Since plane sweep enumeration relies on the correspondence between image pixels and voxels, image warps can also be effective as a defense. The large number of remote image requests required for plane sweep enumeration (O(n) requests for an n n n voxel grid) and the unusual request parameters may look suspicious and trigger the rendering server log analysis monitors. Plane sweep enumeration attacks can be completely nullified by limiting user control of the view frustum parameters, which we implement in our system and use for valuable models. Another enumeration attack, near plane sweep enumeration, involves sweeping the viewpoint (and thus the near plane) over the model, checking when the model surface is clipped by the near plane and marking voxels when this happens (Figure 4(b)). The attacker knows that the near plane has clipped the model when a pixel previously containing the model surface begins to be classified as the background. In order to determine which voxel each image pixel corresponds to, the attacker must know two related parameters : the distance between the viewpoint position and the near plane, and the field of view. These parameters can be easily discovered. The near plane distance can be determined by first obtaining the exact location of one feature point on the model surface through triangulation of multiple rendering requests and then moving the viewpoint slowly toward that point on the model. When the near plane clips the feature point, the distance between that point and the view position equals the near plane distance. The horizontal and vertical field of view angles can be obtained by moving the viewpoint slowly toward the model surface, stopping when any surface point becomes clipped by the near plane. The viewpoint is then moved a small amount perpendicular to its original direction of motion such that the clipped point moves slightly relative to the view but stays on the new image (near plane). Since the near plane distance has already been (a) (b) (c) (d) Figure 5: 3D reconstruction results from enumeration attacks: (a) original 3D model, (b) plane sweep attack against defenseless server (6 passes, 3,168 total rendered images), (c) plane sweep attack against 0.3 viewing direction perturbation defense (6 passes, 3,168 total rendered images), (d) near plane sweep attack against defenseless server (6 passes, 7,952 total rendered images). obtained, the field of view angle (horizontal or vertical depending on direction of motion) can be obtained from the relative motion of the clipped point across the image. Because the near plane is usually small compared to the dimensions of the model, many sweeps must be tiled in order to attain full coverage . Sweeps must also be made in several directions to ensure that all model faces are seen. Because this attack relies on seeing the background to determine when the near plane has clipped a surface , concave model geometries will present a problem for surface detection. Although sweeps from multiple directions will help, this problem is not completely avoidable. Figure 5(d) illustrates this problem, showing a case in which six sweeps have not fully captured all the surface geometry. Viewing parameter perturbations and image warps will nearly destroy the effectiveness of near plane sweep enumeration attacks, as they can make it very difficult to determine where the surface lies and where it does not near silhouette edges (pixels near these edges will change erratically between surface and background). The most solid defense against this attack is to prevent views within a certain small offset of the model surface. This defense, which we use in our system to protect valuable models, prevents the near plane from ever clipping the model surface and thereby completely nullifies this attack. 5.2 Shape-from-silhouette Attacks Shape-from-silhouette [Slabaugh et al. 2001] is one well studied, robust technique for extracting a 3D model from a set of images. The method consists of segmenting the object pixels from the background in each image, then intersecting in space their resulting extended truncated silhouettes, and finally computing the surface of the resulting shape. The main limitation of this technique is that only a visual hull [Laurentini 1994] of the 3D shape can be recovered ; the line-concave parts of the model are beyond the capabilities of the reconstruction. Thus, the effectiveness of this attack depends on the specific geometric characteristics of the object; the high-resolution 3D models that we target often have many concavities that are difficult or impossible to fully recover using shape-from-silhouette . However, this attack may also be of use to attackers 700 Figure 6: The 160 viewpoints used to reconstruct the model with a shape-from-silhouette attack; results are shown in Figure 7. to obtain a coarse, low-resolution version of the model, if they are unable to break through the obfuscation protections we use for the low-resolution models distributed with the client. To measure the potential of a shape-from-silhouette attack against our protected graphics system, we have conducted reconstruction experiments on a 3D model of the David as served via the rendering server, using a shape-from-silhouette implementation described in [Tarini et al. 2002]. With all server defenses disabled, 160 images were harvested from a variety of viewpoints around the model (Figure 6); these viewpoints were selected incrementally, with later viewpoints chosen to refine the reconstruction accuracy as measured during the process. The resulting 3D reconstruction is shown in Figure 7(b). Several of the perturbation and distortion defenses implemented in our server are effective against the shape-from-silhouette attack. Results from experiments showing the reconstructed model quality with server defenses independently enabled are shown in Figures 7(c-g). Small perturbations in the viewing parameters were particularly effective at decreasing the quality of the reconstructed model, as would be expected; Niem [1997] performed an error analysis of silhouette-based modeling techniques and showed the linear relationship between error in the estimation of the view position and error in the resulting reconstruction. Perturbations in the images returned from the server, such as radial distortion and small random shifts, were also effective. Combining the different perturbation defenses, as they are implemented in our remote rendering system, makes for further deterioration of the reconstructed model quality (Figure 7(h)). High frequency noise and JPEG defenses in the server images can increase the difficulty of segmenting the object from the background . However, shape-from-silhouette software implementa-tions with specially tuned image processing operations can take the noise characteristics into account to help classify pixels accurately. The intersection stage of shape-from-silhouette reconstruction algorithms makes them innately robust with respect to background pixels misclassified as foreground. 5.3 Stereo Correspondence-based Attacks Stereo reconstruction is another well known 3D computer vision technique. Stereo pairs of similarly neighborhooded pixels are detected , and the position of the corresponding point on the 3D surface is found via the intersection of epipolar lines. Of particular relevance to our remote rendering system, Debevec et al. [1996] showed that the reconstruction task can be made easier and more accurate if an approximate low resolution model is available, by warping the images over it before performing the stereo matching. (a) E = 0 (b) E = 4.5 (c) E = 13.5 (d) E = 45.5 (e) E = 11.6 (f) E = 9.3 (g) E = 16.2 (h) E = 26.6 Figure 7: Performance of shape-from-silhouette reconstructions against various server defenses. Error values (E) measure the mean surface distance (mm) from the 5m tall original model. Top row: (a) original model, (b) reconstruction from defenseless server, reconstruction with (c) 0.5 and (d) 2.0 perturbations of the view direction. Bottom row: (e) reconstruction with a random image offset of 4 pixels, with (f) 1.2% and (g) 2.5% radial image distortion, and (h) reconstruction against combined defenses (1.0 view perturbation , 2 pixel random offset, and 1.2% radial image distortion). Ultimately, however, stereo correspondence techniques usually rely on matching detailed, high-frequency features in order to yield high-resolution reconstruction results. The smoothly shaded 3D computer models generated by laser scanning that we share via our remote rendering system thus present significant problems to basic two-frame stereo matching algorithms. When we add in the server defenses such as image-space high frequency noise, and slight perturbations in the viewing and lighting parameters, the stereo matching task becomes even more ill-posed. Other stereo research such as [Scharstein and Szeliski 2002] also reports great difficulty in stereo reconstruction of noise-contaminated, low-texture synthetic scenes. Were we to distribute 3D models with high resolution textures applied to their surfaces, stereo correspondence methods may be a more effective attack. 5.4 Shape-from-shading Attacks Shape-from-shading attacks represent another family of computer vision techniques for reconstructing the shape of a 3D object (see [Zhang et al. 1999] for a survey). The primary attack on our remote rendering system that we consider in this class involves first 701 (a) E = 0 (b) E = 1.9 (c) E = 1.0 (d) E = 1.1 (e) E = 1.7 (f) E = 2.0 Figure 8: Performance of shape-from-shading reconstruction attacks . Error values (E) measure the mean surface distance (mm) from the original model. Top row: (a) original model, (b) low-resolution base mesh, (c) reconstruction from defenseless server. Bottom row: reconstruction results against (d) high-frequency image noise, (e) complicated lighting model (3 lights), and (f) viewing angle perturbation (up to 1.0 ) defenses. obtaining several images from the same viewpoint under varying, known lighting conditions. Then, using photometric stereo methods , a normal is computed for each pixel by solving a system of rendering equations. The resulting normal map can be registered and applied to an available approximate 3D geometry, such as the low-resolution model used by the client, or one obtained from another reconstruction technique such as shape-from-silhouette. This coarse normal-mapped model itself may be of value to some attackers: when rendered it will show convincing 3D high frequency details that can be shaded under new lighting conditions, though with artifacts at silhouettes. However, the primary purpose of our system is to protect the high-resolution 3D geometry, which if stolen could be used maliciously for shape analysis or to create replicas. Thus, a greater risk is posed if the normal map is integrated by the attacker to compute a displacement map, and the results are used to displace a refined version of the low-resolution model mesh. Following this procedure with images harvested from a defenseless remote rendering server and using a low-resolution client model, we were able to successfully reconstruct a high-resolution 3D model. The results shown in Figure 8(c) depict a reconstruction of the David's head produced from 200 1600x1114 pixel images taken from 10 viewpoints, with 20 lighting positions used at each viewpoint, assuming a known, single-illuminant OpenGL lighting model and using a 10,000 polygon low-resolution model (Fig. 8(b)) of the whole statue. Some of the rendering server defenses, such as adding high-frequency noise to the images, can be compensated for by attackers by simply adding enough input images to increase the robustness of the photometric stereo solution step (although harvesting too many images will eventually trigger the rendering server log analysis monitors). Figure 8(d) shows the high quality reconstruction result possible when only random Gaussian noise is used as a defense. More effective defenses against shape-from-shading attacks include viewing and lighting perturbations and low-frequency image distortions, which can make it difficult to precisely register images onto the low-resolution model, and can disrupt the photometric stereo solution step without a large number of aligned input images. Figure 8(e) shows a diminished quality reconstruction when the rendering server complicates the lighting model by using 3 perturbed light sources with a Phong component unknown to the attacker, and Figure 8(f) shows the significant loss of geometric detail in the reconstruction when the server randomly perturbs the viewing direction by up to 1.0 (note that the reconstruction error exceeds that of the starting base mesh). The quality of the base mesh is an important determinant in the success of this particular attack. For example, repeating the experiment of Figure 8 with a more accurate base mesh of 30,000 polygons yields results of E = 0.8, E = 0.6, and E = 0.7 for the conditions of Figures 8(b), 8(c), and 8(e), respectively. This reliance on an accurate low-resolution base mesh for the 3D model reconstruction is a potential weak point of the attack; attackers may be deterred by the effort required to reverse-engineer the protections guarding the low-resolution model or to reconstruct an acceptable base mesh from harvested images using another technique. 5.5 Discussion Because we know of no single mechanism for guaranteeing the security of 3D content delivered through a rendering server, we have instead taken a systems-based approach, implementing multiple defenses and using them in combination. Moreover, we know of no formalism for rigorously analyzing the security provided by our defenses ; the reconstruction attacks that we have empirically considered here are merely representative of the possible threats. Of the reconstruction attacks we have experimented with so far, the shape-from-shading approach has yielded the best results against our defended rendering server. Enumeration attacks are easily foiled when the user's control over the viewpoint and view frustum is constrained, pure shape-from-silhouette methods are limited to reconstructing a visual hull, and two-frame stereo algorithms rely on determining accurate correspondences which is difficult with the synthetic, untextured models we are attempting to protect. Attackers could improve the results of the shape-from-shading algorithm against our perturbation defenses by explicitly modeling the distortions and trying to take them into account in the optimization step, or alternatively by attempting to align the images by interactively establishing point to point correspondences or using an automatic technique such as [Lensch et al. 2001]. Such procedures for explicitly modeling the server defenses, or correcting for them via manual specification of correspondences, are applicable to any style of reconstruction attempt. To combat these attacks, we must rely on the combined discouraging effect of multiple defenses running simultaneously, which increases the number of degrees of freedom of perturbation to a level that would be difficult and time-consuming to overcome. Some of our rendering server defenses, such as the lighting model and non-linear image distortions , can be increased arbitrarily in their complexity. Likewise, the magnitude of server defense perturbations can be increased with a corresponding decrease in the fidelity of the rendered images. Ultimately, no fixed set of defenses is bulletproof against a sophisticated , malicious attacker with enough resources at their disposal , and one is inevitably led to an "arms race" between attacks and countermeasures such as we have implemented. As the expense required to overcome our remote rendering server defenses becomes greater, determined attackers may instead turn to reaching their piracy goals via non-reconstruction-based methods beyond the scope of this paper, such as computer network intrusion or exploitation of non-technical human factors. 702 Results and Future Work A prototype of our remote rendering system (ScanView, available at http://graphics.stanford.edu/software/scanview/ ) has been deployed to share 3D models from a major cultural heritage archive, the Digital Michelangelo Project [Levoy et al. 2000], as well as other collections of archaeological artifacts that require protected usage. In the several months since becoming publically available, more than 4,000 users have installed the client program on their personal computers and accessed the remote servers to view the protected 3D models. The users have included art students, art scholars , art enthusiasts, and sculptors examining high-resolution artworks , as well as archaeologists examining particular artifacts. Few of these individuals would have qualified under the strict guidelines required to obtain completely unrestricted access to the models, so the protected remote rendering system has enabled large, entirely new groups of users access to 3D graphical models for professional study and enjoyment. Reports from users of the system have been uniformly positive and enthusiastic. Fetching high-resolution renderings over intercontinental broadband Internet connections takes less than 2 seconds of latency, while fast continental connections generally experience latencies dominated by the rendering server's processing time (around 150 ms). The rendering server architecture can scale up to support an arbitrary number of requests per second by adding additional CPU and GPU nodes, and rendering servers can be installed at distributed locations around the world to reduce intercontinental latencies if desired. Our log analysis defenses have detected multiple episodes of system users attempting to harvest large sets of images from the server for purposes of later 3D reconstruction attempts, though these incidents were determined to be non-malicious. In general, the monitoring capabilities of a remote rendering server are useful for reasons beyond just security, as the server logs provide complete accounts of all usage of the 3D models in the archive, which can be valuable information for archive managers to gauge popularity of individual models and understand user interaction patterns. Our plans for future work include further investigation of computer vision techniques that address 3D reconstruction of synthetic data under antagonistic conditions, and analysis of their efficacy against the various rendering server defenses. More sophisticated extensions to the basic vision approaches described above, such as multi-view stereo algorithms, and robust hybrid vision algorithms which combine the strengths of different reconstruction techniques, can present difficult challenges to protecting the models. Another direction of research is to consider how to allow users a greater degree of geometric analysis of the protected 3D models without further exposing the data to theft; scholarly and professional users have expressed interest in measuring distances and plotting profiles of 3D objects for analytical purposes beyond the simple 3D viewing supported in the current system. Finally, we are continuing to investigate alternative approaches to protecting 3D graphics, designing specialized systems which make data security a priority while potentially sacrificing some general purpose computing platform capabilities. The GPU decryption scheme described herein, for example , is one such idea that may be appropriate for console devices and other custom graphics systems. Acknowledgements We thank Kurt Akeley, Sean Anderson, Jonathan Berger, Dan Boneh, Ian Buck, James Davis, Pat Hanrahan , Hughes Hoppe, David Kirk, Matthew Papakipos, Nick Triantos, and the anonymous reviewers for their useful feedback, and Szymon Rusinkiewicz for sharing code. This work has been supported in part by NSF contract IIS0113427, the Max Planck Center for Visual Computing and Communication, and the EU IST-2001 -32641 ViHAP3D Project. References C OLLBERG , C., AND T HOMBORSON , C. 2000. Watermarking, tamper-proofing , and obfuscation: Tools for software protection. Tech. Rep. 170, Dept. of Computer Science, The University of Auckland. D EBEVEC , P., T AYLOR , C., AND M ALIK , J. 1996. Modeling and rendering architecture from photographs: A hybrid geometry- and image-based approach. In Proc. of ACM SIGGRAPH 96, 1120. E NGEL , K., H ASTREITER , P., T OMANDL , B., E BERHARDT , K., AND E RTL , T. 2000. Combining local and remote visualization techniques for interactive volume rendering in medical applications. In Proc. of IEEE Visualization 2000 , 449452. G ORTLER , S., G RZESZCZUK , R., S ZELISKI , R., AND C OHEN , M. F. 1996. The lumigraph. In Proc. of ACM SIGGRAPH 96, 4354. L AURENTINI , A. 1994. The visual hull concept for silhouette-based image understanding. IEEE Trans. on Pattern Analysis and Machine Intelligence 16 , 2, 150162. L ENSCH , H. P., H EIDRICH , W., AND S EIDEL , H.-P. 2001. A silhouette-based algorithm for texture registration and stitching. Graphical Models 63 , 245262. L EVOY , M., AND H ANRAHAN , P. 1996. Light field rendering. In Proc. of ACM SIGGRAPH 96 , 3142. L EVOY , M., P ULLI , K., C URLESS , B., R USINKIEWICZ , S., K OLLER , D., P EREIRA , L., G INZTON , M., A NDERSON , S., D AVIS , J., G INSBERG , J., S HADE , J., AND F ULK , D. 2000. The digital michelangelo project. In Proc. of ACM SIGGRAPH 2000, 131144. L EVOY , M. 1995. Polygon-assisted jpeg and mpeg compression of synthetic images. In Proc. of ACM SIGGRAPH 95, 2128. N IEM , W. 1997. Error analysis for silhouette-based 3d shape estimation from multiple views. In International Workshop on Synthetic-Natural Hybrid Coding and 3D Imaging . O HBUCHI , R., M UKAIYAMA , A., AND T AKAHASHI , S. 2002. A frequency-domain approach to watermarking 3d shapes. Computer Graphics Forum 21 , 3. P RAUN , E., H OPPE , H., AND F INKELSTEIN , A. 1999. Robust mesh watermarking . In Proc. of ACM SIGGRAPH 99, 4956. R ESSLER , S., 2001. Web3d security discussion. Online article: http://web3d.about.com/library/weekly/aa013101a.htm . R USINKIEWICZ , S., AND L EVOY , M. 2000. QSplat: A multiresolution point rendering system for large meshes. In Proc. of ACM SIGGRAPH 2000 , 343352. S CHARSTEIN , D., AND S ZELISKI , R. 2002. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. International Journal of Computer Vision 47 , 13, 742. S CHNEIER , B. 2000. The fallacy of trusted client software. Information Security (August). S IMONS , D., AND L EVIN , D. 1997. Change blindness. Trends in Cognitive Sciences 1 , 7, 261267. S LABAUGH , G., C ULBERTSON , B., M ALZBENDER , T., AND S CHAFER , R. 2001. A survey of methods for volumetric scene reconstruction from photographs. In Proc. of the Joint IEEE TCVG and Eurographics Workshop (VolumeGraphics-01) , Springer-Verlag, 81100. S TANFORD D IGITAL F ORMA U RBIS P ROJECT , 2004. http://formaurbis.stanford.edu. T ARINI , M., C ALLIERI , M., M ONTANI , C., R OCCHINI , C., O LSSON , K., AND P ERSSON , T. 2002. Marching intersections: An efficient approach to shape-from-silhouette. In Proceedings of the Conference on Vision, Modeling, and Visualization (VMV 2002) , 255262. Y OON , I., AND N EUMANN , U. 2000. Web-based remote rendering with IBRAC. Computer Graphics Forum 19, 3. Z HANG , R., T SAI , P.-S., C RYER , J. E., AND S HAH , M. 1999. Shape from shading: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 21 , 8, 690706. 703

digital rights management;remote rendering;security;3D models

154

Providing the Basis for Human-Robot-Interaction: A Multi-Modal Attention System for a Mobile Robot

In order to enable the widespread use of robots in home and office environments, systems with natural interaction capabilities have to be developed. A prerequisite for natural interaction is the robot's ability to automatically recognize when and how long a person's attention is directed towards it for communication. As in open environments several persons can be present simultaneously, the detection of the communication partner is of particular importance. In this paper we present an attention system for a mobile robot which enables the robot to shift its attention to the person of interest and to maintain attention during interaction. Our approach is based on a method for multi-modal person tracking which uses a pan-tilt camera for face recognition, two microphones for sound source localization, and a laser range finder for leg detection. Shifting of attention is realized by turning the camera into the direction of the person which is currently speaking. From the orientation of the head it is decided whether the speaker addresses the robot. The performance of the proposed approach is demonstrated with an evaluation. In addition, qualitative results from the performance of the robot at the exhibition part of the ICVS'03 are provided.

INTRODUCTION A prerequisite for the successful application of mobile service robots in home and office environments is the development of systems with natural human-robot-interfaces. Much research focuses Figure 1: Even in crowded situations (here at the ICVS'03) the mobile robot BIRON is able to robustly track persons and shift its attention to the speaker. on the communication process itself, e.g. speaker-independent speech recognition or robust dialog systems. In typical tests of such human-machine interfaces, the presence and position of the communication partner is known beforehand as the user either wears a close-talking microphone or stands at a designated position. On a mobile robot that operates in an environment where several people are moving around, it is not always obvious for the robot which of the surrounding persons wants to interact with it. Therefore, it is necessary to develop techniques that allow a mobile robot to automatically recognize when and how long a user's attention is directed towards it for communication. For this purpose some fundamental abilities of the robot are required . First of all, it must be able to detect persons in its vicinity and to track their movements over time in order to differentiate between persons. In previous work, we have demonstrated how tracking of persons can be accomplished using a laser range finder and a pan-tilt color camera [6]. As speech is the most important means of communication for humans, we extended this framework to incorporate sound source information for multi-modal person tracking and attention control. This enables a mobile robot to detect and localize sound sources in the robot's surroundings and, therfore, to observe humans and to shift its attention to a person that is likely to communicate with the robot. The proposed attention system is part of a larger research effort aimed at building BIRON the Bielefeld Robot Companion. BIRON has already performed attention control successfully during several demonstrations. Figure 1 depicts a typical situation during the exhibition of our mobile robot at the International Conference on Computer Vision Systems (ICVS) 2003 in Graz. The paper is organized as follows: At first we discuss approaches that are related to the detection of communication partners in section 2. Then, in section 3 the robot hardware is presented. Next, multi-modal person tracking is outlined in section 4, followed by the explanation of the corresponding perceptual systems in section 5. This is the basis of our approach for the detection of communication partners explained in section 6. In section 7 an extensive evaluation of the system is presented. The paper concludes with a short summary in section 8. RELATED WORK As long as artificial systems interact with humans in static setups the detection of communication partners can be achieved rather easily . For the interaction with an information kiosk the potential user has to enter a definite space in front of it (cf. e.g. [14]). In intelligent rooms usually the configuration of the sensors allows to monitor all persons involved in a meeting simultaneously (cf. e.g. [18]). In contrast to these scenarios a mobile robot does not act in a closed or even controlled environment. A prototypical application of such a system is its use as a tour guide in scientific laboratories or museums (cf. e.g. [3]). All humans approaching or passing the robot have to be considered to be potential communication partners . In order to circumvent the problem of detecting humans in an unstructured and potentially changing environment, in the approach presented in [3] a button on the robot itself has to be pushed to start the interaction. Two examples for robots with advanced human-robot interfaces are SIG [13] and ROBITA [12] which currently demonstrate their capabilities in research labs. Both use a combination of visual face recognition and sound source localization for the detection of a person of interest. SIG's focus of attention is directed towards the person currently speaking that is either approaching the robot or standing close to it. In addition to the detection of talking people, ROBITA is also able to determine the addressee of spoken utterances . Thus, it can distinguish speech directed towards itself from utterances spoken to another person. Both robots, SIG and ROBITA , can give feedback which person is currently considered to be the communication partner. SIG always turns its complete body towards the person of interest. ROBITA can use several combinations of body orientation, head orientation, and eye gaze. The multi-modal attention system for a mobile robot presented in this paper is based on face recognition, sound source localization and leg detection. In the following related work on these topics will be reviewed. For human-robot interfaces tracking of the user's face is indispensable . It provides information about the user's identity, state, and intent. A first step for any face processing system is to detect the locations of faces in the robot's camera image. However, face detection is a challenging task due to variations in scale and position within the image. In addition, it must be robust to different lighting conditions and facial expressions. A wide variety of techniques has been proposed, for example neural networks [15], deformable templates [23], skin color detection [21], or principle component analysis (PCA), the so-called Eigenface method [19]. For an overview the interested reader is referred to [22, 9]. In current research on sound or speaker localization mostly microphone arrays with at least 3 microphones are used. Only a few approaches employ just one pair of microphones. Fast and robust techniques for sound (and therefore speaker) localization are e.g. the Generalized Cross-Correlation Method [11] or the Cross-Powerspectrum Phase Analysis [8], which both can be applied for microphone-arrays as well as for only one pair of microphones. More complex algorithms for speaker localization like spectral separation and measurement fusion [2] or Linear-Correction Least-Squares [10] are also very robust and can additionally estimate the distance and the height of a speaker or separate different audio sources. Such complex algorithms require more than one pair of microphones to work adequately and also require substantial processing power. In mobile robotics 2D laser range finders are often used, primarily for robot localization and obstacle avoidance. A laser range finder can also be applied to detect persons. In the approach presented in [16] for every object detected in a laser scan features like diameter, shape, and distance are extracted. Then, fuzzy logic is used to determine which of the objects are pairs of legs. In [17] local minima in the range profile are considered to be pairs of legs. Since other objects (e.g. trash bins) produce patterns similar to persons , moving objects are distinguished from static objects, too. ROBOT HARDWARE Figure 2: The mobile robot BIRON. The hardware platform for BIRON is a Pioneer PeopleBot from ActivMedia (Fig. 2) with an on-board PC (Pentium III, 850 MHz) for controlling the motors and the on-board sensors and for sound processing. An additional PC (Pentium III, 500 MHz) inside the robot is used for image processing. The two PC's running Linux are linked with a 100 Mbit Ethernet and the controller PC is equipped with wireless Ethernet to enable remote control of the mobile robot. For the interaction with a user a 12" touch screen display is provided on the robot. A pan-tilt color camera (Sony EVI-D31 ) is mounted on top of the robot at a height of 141 cm for acquiring images of the upper body part of humans interacting with the robot. Two AKG far-field microphones which are usually used for hands free telephony are located at the front of the upper platform at a height of 106 cm, right below the touch screen display. The distance between the microphones is 28.1 cm. A SICK laser range finder is mounted at the front at a height of approximately 30 cm. For robot navigation we use the ISR (Intelligent Service Robot) control software developed at the Center for Autonomous Systems, KTH, Stockholm [1]. MULTI-MODAL PERSON TRACKING In order to enable a robot to direct its attention to a specific person it must be able to distinguish between different persons. Therefore , it is necessary for the robot to track all persons present as robustly as possible. Person tracking with a mobile robot is a highly dynamic task. As both, the persons tracked and the robot itself might be moving the sensory perception of the persons is constantly changing. Another difficulty arises from the fact that a complex object like a person 29 usually cannot be captured completely by a single sensor system alone. Therefore, we use the sensors presented in section 3 to obtain different percepts of a person: The camera is used to recognize faces. From a face detection step the distance, direction, and height of the observed person are extracted, while an identification step provides the identity of the person if it is known to the system beforehand (see section 5.1). Stereo microphones are applied to locate sound sources using a method based on Cross-Powerspectrum Phase Analysis [8]. From the result of the analysis the direction relative to the robot can be estimated (see section 5.2). The laser range finder is used to detect legs. In range readings pairs of legs of a human result in a characteristic pattern that can be easily detected [6]. From detected legs the distance and direction of the person relative to the robot can be extracted (see section 5.3). The processes which are responsible for processing the data of these sensors provide information about the same overall object: the person. Consequently, this data has to be fused. We combine the information from the different sensors in a multi-modal framework which is described in the following section. 4.1 Multi-Modal Anchoring In order to solve the problem of person tracking we apply multi-modal anchoring [6]. This approach extends the idea of standard anchoring as proposed in [4]. The goal of anchoring is defined as establishing connections between processes that work on the level of abstract representations of objects in the world (symbolic level) and processes that are responsible for the physical observation of these objects (sensory level). These connections, called anchors, must be dynamic, since the same symbol must be connected to new percepts every time a new observation of the corresponding object is acquired. Therefore, in standard anchoring at every time step , an anchor contains three elements: a symbol, which is used to denote an object , a percept of the same object, generated by the corresponding perceptual system, and a signature, which is meant to provide an estimate for the values of the observable properties of the object. If the anchor is grounded at time , it contains the percept perceived at as well as the updated signature. If the object is not observable at and therefore the anchor is ungrounded, then no percept is stored in the anchor but the signature still contains the best available estimate. Because standard anchoring only considers the special case of connecting one symbol to the percepts acquired from one sensor, the extension to multi-modal anchoring was necessary in order to handle data from several sensors. Multi-modal anchoring allows to link the symbolic description of a complex object to different types of percepts, originating from different perceptual systems. It enables distributed anchoring of individual percepts from multiple modalities and copes with different spatio-temporal properties of the individual percepts. Every part of the complex object which is captured by one sensor is anchored by a single component anchoring process. The composition of all component anchors is realized by a composite anchoring process which establishes the connection between the symbolic description of the complex object and the percepts from the individual sensors. In the domain of person tracking the person itself is the composite object while its components are face, speech, and legs, respectively. In addition Signature list Fusion Motion Composition Face region Sound source Laser legs Anchoring Anchoring Anchoring person face speech legs ... name, height, t 2 t 0 t 1 Anchor position, etc. Anchoring of composite object Person models Symbols Percepts Anchoring of component objects Figure 3: Multi-modal anchoring of persons. to standard anchoring, the composite anchoring module requires a composition model, a motion model, and a fusion model: The composition model defines the spatial relationships of the components with respect to the composite object. It is used in the component anchoring processes to anchor only those percepts that satisfy the composition model. The motion model describes the type of motion of the complex object, and therefore allows to predict its position. Using the spatial relationships of the composition model, the position of percepts can be predicted, too. This information is used by the component anchoring processes in two ways: 1. If multiple percepts are generated from one perceptual system the component anchoring process selects the percept which is closest to the predicted position. 2. If the corresponding perceptual system receives its data from a steerable sensor with a limited field of view (e.g. pan-tilt camera), it turns the sensor into the direction of the predicted position. The fusion model defines how the perceptual data from the component anchors has to be combined. It is important to note, that the processing times of the different perceptual systems may differ significantly. Therefore, the perceptual data may not arrive at the composite anchoring process in chronological order. Consequently, the composite anchor provides a chronologically sorted list of the fused perceptual data. New data from the component anchors is inserted in the list, and all subsequent entries are updated. The anchoring of a single person is illustrated in Figure 3. It is based on anchoring the three components legs, face, and speech. For more details please refer to [6]. 4.2 Tracking Multiple Persons If more than one person has to be tracked simultaneously, several anchoring processes have to be run in parallel. In this case, multi-modal anchoring as described in the previous section may lead to the following conflicts between the individual composite anchoring processes: The anchoring processes try to control the pan-tilt unit of the camera in a contradictory way. A percept is selected by more than one anchoring process. 30 In order to resolve these problems a supervising module is required, which grants the access to the pan-tilt camera and controls the selection of percepts. The first problem is handled in the following way: The supervising module restricts the access to the pan-tilt unit of the camera to only one composite anchoring process at a time. How access is granted to the processes depends on the intended application. For the task of detecting communication partners which is presented in this paper, only the anchoring process corresponding to the currently selected person of interest controls the pan-tilt unit of the camera (see section 6). In order to avoid the second problem, the selection of percepts is implemented as follows. Instead of selecting a specific percept de-terministically , every component anchoring process assigns scores to all percepts rating the proximity to the predicted position. After all component anchoring processes have assigned scores, the supervising module computes the optimal non-contradictory assignment of percepts to component anchors. Percepts that are not assigned to any of the existing anchoring processes are used to establish new anchors. Additionally, an anchor that was not updated for a certain period of time will be removed by the supervising module. PERCEPTUAL SYSTEMS In order to supply the anchoring framework presented in 4.1 with sensory information about observed persons, three different perceptual systems are used. These are outlined in the following subsec-tions . 5.1 Face Recognition In our previous work [6], face detection was realized using a method which combines adaptive skin-color segmentation with face detection based on Eigenfaces [7]. The segmentation process reduces the search space, so that only those sub-images which are located at skin colored regions have to be verified with the Eigenface method. In order to cope with varying lighting conditions the model for skin-color is continuously updated with pixels extracted from detected faces. This circular process requires initialization, which is realized by performing face detection using Eigenfaces on the whole image, since initially no suitable model for skin-color is available. This method has two major drawbacks: It is very sensitive to false positive detections of faces, since then the skin-model may adapt to a wrong color. In addition, initialization is computa-tionally very expensive. In our current system presented in this paper, the detection of faces (in frontal view) is based on the framework proposed by Viola and Jones [20]. This method allows to process images very rapidly with high detection rates for the task of face detection. Therefore, neither a time consuming initialization nor the restriction of the search using a model of skin color is necessary. The detection is based on two types of features (Fig. 4), which are the same as proposed in [24]. A feature is a scalar value which is computed by the weighted sum of all intensities of pixels in rectangular regions. The computation can be realized very efficiently using integral images (see [20]). The features have six degrees of freedom for two-block features ( ) and seven degrees of freedom for three-block features ( ). With restrictions to the size of the rectangles and their distances we obtain about 300.000 different features for sub-windows of a size of pixels. Classifiers are constructed by selecting a small number of important features using AdaBoost [5]. A cascade of classifiers of increasing complexity (increasing number of features) forms the over-all face detector (Fig. 5). For face detection an image is scanned, and every sub-image is classified 2 1 dy 2 dx 1 dy dx 2 w h w dx h x y x y Figure 4: The two types of features used for face detection. Each feature takes a value which is the weighted sum of all pixels in the rectangles. ..... Non-face Non-face No No Input Sub-Window Yes Yes Non-face Face No C n C 2 C 1 Yes Figure 5: A cascade of classifiers of increasing complexity enables fast face detection. with the first classifier of the cascade. If classified as non-face, the process continues with the next sub-image. Otherwise the current sub-image is passed to the next classifier ( ) and so on. The first classifier of the cascade is based on only two features, but rejects approximately 75 % of all sub-images. Therefore, the detection process is very fast. The cascade used in our system consists of 16 classifiers based on 1327 features altogether. In order to update the multi-modal anchoring process the position of the face is extracted: With the orientation of the pan-tilt camera, the angle of the face relative to the robot is calculated. The size of the detected face is used to estimate the distance of the person : Assuming that sizes of heads of adult humans only vary to a minor degree, the distance is proportional to the reciprocal of the size. The height of the face above the ground is also extracted by using the distance and the camera position. Since the approach presented so far does not provide face identification , a post-processing step is is required. Therefore, we use a slightly enhanced version of the Eigenface method [19]. Each individual is represented in face space by a mixture of several Gaussians with diagonal covariances. Practical experiments have shown that the use of four to six Gaussians leads to a satisfying accuracy in discriminating between a small set of known persons. 5.2 Sound Source Localization In order to detect speaking persons, we realize the localization of sound sources using a pair of microphones. Given a sound source in 3D space, the distances and between and the two microphones and generally differ by the amount of (see Fig. 6). This difference results in a time delay of the received signal between the left and the right channel (microphone). Based on Cross-Powerspectrum Phase Analysis [8] we first calculate a spectral correlation measure (1) where and are the short-term power spectra of the left and the right channel, respectively (calculated within a 43 ms window from the signal sampled at 48 kHz). If only a single sound 31 d 10 cm d =0 c m d =5c m d = 1 0c m d = 2 0 c m d = 25 cm d = 1 5 c m s m 2 m 1 d 1 d 2 b = 28.1 cm Figure 6: The distances and between the sound source and the two microphones and differ by the amount of . All sound events with identical are located on one half of a two-sheeted hyperboloid (gray). source is present the time delay will be given by the argument that maximizes the spectral correlations measure : (2) Taking into account also local maxima delivered by equation (1), we are able to detect several sound sources simultaneously. Even in the planar case, where all sound sources are on the same level as the microphones, the position of can be estimated only if its distance is known or additional assumptions are made. In a simplified geometry the microphone distance is considered suf-ficiently small compared to the distance of the source. Therefore, the angles of incidence of the signals observed at the left or right microphone, respectively, will be approximately equal and can be calculated directly from . In the 3D-case the observed time delay not only depends on the direction and distance but also on the relative elevation of the source with respect to the microphones. Therefore, given only the problem is under-determined. All sound events which result in the same are located on one half of a two-sheeted hyperboloid, given by (3) where is the position of the sound source given in Cartesian coordinates. The axis of symmetry of the hyperboloid coincides with the axis on which the microphones are located (y-axis). Figure 6 shows the intersections of hyperboloids for different with the plane spanned by , , and . Consequently, the localization of sound sources in 3D using two microphones requires additional information. As in our scenario sound sources of interest correspond to persons talking, the additional spatial information necessary can be obtained from the other perceptual systems of the multi-modal anchoring framework. Leg detection and face recognition provide information about the direction, distance, and height of a person with respect to the local coordinate system of the robot. Even if no face was detected at all, the height of a person can be estimated as the standard size of an adult. In order to decide whether a sound percept can be assigned to a specific person, the sound source has to be located in 3D. For this purpose it is assumed that the sound percept originates from 100 cm Robot Figure 7: A sample laser scan. The arrow marks a pair of legs. the person and is therefore located at the same height and same distance . Then, the corresponding direction of the sound source can be calculated from equation (3) transformed to cylindric coordinates. Depending on the difference between this direction and the direction in which the person is located, the sound percept is assigned to the person's sound anchor. Similar to other component anchors, the direction of the speech is also fused with the position of the person. Note that the necessity of positional attributes of a person for the localization of speakers implies that speech can not be anchored until the legs or the face of a person have been anchored. In conclusion, the use of only one pair of microphones is sufficient for feasible speaker localization in the multi-modal anchoring framework. 5.3 Leg Detection The laser range finder provides distance measures within a field of view at leg-height. The angular resolution is resulting in 361 reading points for a single scan (see Fig. 7 for an example ). Usually, human legs result in a characteristic pattern which can be easily detected. This is done as follows: At first, neighboring reading points with similar distance values are grouped into segments. Then, these segments are classified as legs or non-legs based on a set of features (see [6]). Finally, legs with a distance that is below a threshold are grouped into pairs of legs. FOCUSING THE ATTENTION For the detection of a person of interest from our mobile robot we apply multi-modal person tracking, as described in section 4. Every person in the vicinity of the robot is anchored and, therefore, tracked by an individual anchoring process, as soon as the legs or the face can be recognized by the system. If the robot detects that a person is talking, this individual becomes the person of interest and the robot directs its attention towards it. This is achieved by turning the camera into the direction of the person. The anchoring process corresponding to the person of interest maintains access to the pan-tilt camera and keeps the person in the center of the camera's field of view. If necessary, the entire robot basis is turned in the direction of the person of interest. If this person moves to far away from the robot, the robot will start to follow the person. This behavior ensures that the sensors of the robot do not loose track of this person. Moreover, the person can guide the robot to a specific place. As long as the speech of the person of interest is anchored, other people talking are ignored. This allows the person of interest to take breath or make short breaks while speaking without loosing the robots attention. When the person of interest stops talking for more than two seconds, the person of interest looses its speech anchor. Now, another person can become the person of interest. If no other person is speaking in the vicinity of the robot, the person which 32 (2) (3) (4) (1) P 1 P 1 P 1 P 1 P 2 P 2 P 2 P 2 R R R R Figure 8: Sample behavior with two persons and standing near the robot : In (1) is considered as communication partner, thus the robot directs its attention towards . Then stops speaking but remains person of interest (2). In (3) begins to speak. Therefore the robot's attention shifts to by turning its camera (4). Since is facing the robot, is considered as new communication partner. is in the focus of attention of the robot remains person of interest. Only a person that is speaking can take over the role of the person of interest. Notice, that a person which is moving fast in front of the robot is considered as a passer-by, and hence is definitely no person of interest even if this person is speaking. In addition to the attention system described so far, which enables the robot to detect the person of interest and to maintain its attention during interaction, the robot decides whether the person of interest is addressing the robot and, therefore, is considered as communication partner. This decision is based on the orientation of the person's head, as it is assumed that humans face their addressees for most of the time while they are talking to them. Whether a tracked person faces the robot or not is derived from the face recognition system. If the face of the person of interest is detected for more than 20 % of the time the person is speaking, this person is considered to be the communication partner. A sample behavior of the robot is depicted in Figure 8. SYSTEM PERFORMANCE In order to analyze the performance of the proposed approach, we present results from three different types of evaluation. At first, we study the accuracy of sound source localization independently . The second part deals with a quantitative evaluation of our approach for a multi-modal attention system. Finally, qualitative results from a performance of the robot at the exhibition part of the ICVS'03 are presented. 7.1 Evaluation of Sound Source Localization The objective of this evaluation was to analyze the accuracy of locating speakers with a pair of microphones using the method described in section 5.2 independently from the multi-modal anchoring framework. In order to be able to estimate the arrival angle relative to the microphones, the setup for the experiment was arranged such that the sound source (mouth of the speaker) was always at the same height as the microphones. Therefore, the simplified geometric model mentioned in section 5.2 can be used. The experiments were carried out with five subjects. Every subject was positioned at six different angles ( , , , , , and ), and at two different distances (100 cm and 200 cm), respectively , resulting in 12 positions altogether. At every position a subject had to read out one specific sentence which took about 8 seconds. During every utterance the position of the speaker was calculated every 50 ms. Based on the angles estimated by our localization algorithm we calculated the mean angle and the variance for every speaker. It is important to note, that in our setup it is almost impossible to position the test speaker accurately on the target angle. For this reason, Distance between speaker and robot Angle 100 cm 200 cm 0 -0.9 0.56 -0.3 0.81 10 9.1 0.34 9.2 0.37 20 18.9 0.21 19.3 0.27 40 38.2 0.50 38.8 0.22 60 57.7 0.40 57.5 0.64 80 74.0 2.62 73.3 2.18 Table 1: Averaged estimated speaker positions and averaged variances for the acoustic speaker localization. we used the mean estimated angle for every speaker instead of the target angle to calculate the variance. Following the calculation of mean angle and variance for every speaker we averaged for every position the mean angle and the variance across all speakers. Table 1 shows the results of our experiments. First, the results suggest that the robot was not correctly aligned, because especially for small angles (0 to 20 ) the averaged angle differs constantly from the target angle about 1 . Under this justifiable assumption the speaker localization works very well for angles between 0 and 40 . The estimated angle is nearly equivalent to the actual angle and the variance is also very low. Furthermore, the acoustic position estimation works equally well for 100 cm and for 200 cm. For angles greater than 40 the difference between estimated angle and target angle as well as the variance increases. This means that the accuracy of the acoustic position estimation decreases with an increasing target angle. The main reason for this behavior is the directional characteristic of the microphones. However, the evaluation has shown that the time delay estimation works reasonably well. Thus the sound source localization provides important information for the detection and localization of the current person of interest. 7.2 Evaluation of the Attention System The objective of this evaluation was to analyze the performance of the attention system presented in this paper. On the one hand, the capability of the robot to successfully shift its attention on the speaker, and to recognize when it was addressed was investigated. On the other hand, details about the perceptual sub-systems were of interest. The experiment was carried out in an office room (Fig. 9). Four persons were standing around the robot at designated positions. In reference to the local coordinate system of the robot, person was located at a distance of cm and an angle of , where is defined as the direction ahead of the robot. Person was located at cm , person at cm , and person at cm . The subjects were asked to speak for about 10 seconds, one after another. They had to either address the robot or one of the other persons by turning their heads into the corresponding direction. There were no restrictions on how to stand. The order in which the persons were speaking was predetermined (see Table 2). The experiment was carried out three times with nine different subjects altogether. The following results were achieved: The attention system was always able to determine the correct person of interest within the time the person was speaking . However, in some situations either the reference to the 33 P2 R P3 P4 P1 Figure 9: Setup for the evaluation of the attention system. Person Step 1 2 3 4 5 6 7 8 9 11 12 10 P 4 P 3 P 2 P 1 Table 2: Order in which the persons were speaking, either to the robot (steps 14 and 912) or to another person (steps 58). last person of interest was sustained too long or an incorrect person of interest was selected intermediately. A diagram of the robot's focus of attention is shown in Figure 10. The erroneous behavior occurred in 4 of the 36 time slices: In these cases, the robot shifted its attention to a person which was currently not speaking (see column 5 in all experiments and column 4 in the last experiment in Fig. 10). Note that in all failure cases person , which was located in front of the robot, was selected as person of interest. In addition, there were two shifts which were correct but had a very long delay (eighth time slice of the first and the third experiment). Again, the person in front of the robot ( ) was involved. All errors occurred because a sound source was located in the direction of , although person was not speaking. This can be explained with the noise of the robot itself, which is interpreted as a sound source in the corresponding direction. This error could be suppressed using voice activity detection, which distinguishes speech from noise. This will be part of our future work. As the diagram in Fig. 10 shows, every shift of attention had a delay of approximately 2 seconds. This results from the anchoring framework: The anchor for the sound source is removed after a period of 2 seconds with no new assigned percepts. Now, if another person is talking it becomes the person of interest. The decision whether the current person of interest was addressing the robot or not was made as described in section 6. It was correct for all persons in all runs. This means that the robot always determined himself as addressee in steps 14 and 912, and never in steps 58. These results prove that the presented approach for a multi-modal attention system on a mobile robot is capable to identify communication partners successfully. P 1 P 3 P 4 P 2 1 2 3 4 5 6 7 8 9 10 11 12 1st P 1 P 2 P 3 P 4 3rd P 1 P 2 P 3 P 4 2nd Figure 10: Diagram for the three runs of the experiment. Every person is assigned a track (light-gray) which is shaded while the person was speaking. The solid line shows which person was in focus of the robot's attention. In addition the following measurements concerning the anchoring framework were extracted during the experiments: The attention system and the face recognition were running on one PC (Pentium III, 500 MHz), while the sound source localization and the robot control software were running on the other PC (Pentium III, 850 MHz). Face recognition was performed on images of a size of at a rate of 9.6 Hz. Localization of sound sources was running at a rate of 5.5 Hz. The laser range finder provided new data at a rate of 4.7 Hz while the processing time for the detection of legs was negligible. The anchoring processes of the persons which were currently speaking to the robot were updated with percepts at a rate of 15.4 Hz. Face percepts were assigned to the corresponding anchor at 71.4 % of the time. Note, that after a new person of interest is selected it takes up to approximately 1 second until the camera is turned and the person is in the field of view. During this time, no face percept for the person of interest can be generated. Sound percepts were assigned at 69.5 % of the time, and leg percepts at 99.9 % of the time. The multi-modal anchoring framework was able to quantify the body heights of all subjects with an accuracy of at least 5 cm, which was sufficient to precisely locate sound sources in 3D (see section 5.2). 7.3 Performance at an Exhibition In the beginning of April 2003 our robot was presented at the exhibition part of the International Conference on Computer Vision Systems (ICVS) in Graz. There we were able to demonstrate the robot's capabilities in multi-modal person tracking, and also in following people. BIRON was continuously running without any problems. On the two exhibition days, the robot was running 9:20 hours and 6:30 hours, respectively, tracking about 2240 persons on the first day, and about 1400 persons on the second day. The large amount of persons tracked results from the following condition: Every person which came in the vicinity of the robot was counted once. However, if a person left the observed area and came back later, it was counted again as a new person. Since the coffee breaks of the conference took place in the exhibition room, there were extremely busy phases. Even then, the robot was able to track up to 10 persons simultaneously. Despite the high noise level, the sound source localization worked reliably, even though it was necessary to talk slightly louder to attract the robot's attention. 34 SUMMARY In this paper we presented a multi-modal attention system for a mobile robot. The system is able to observe several persons in the vicinity of the robot and to decide based on a combination of acoustic and visual cues whether one of these is willing to engage in a communication with the robot. This attentional behavior is realized by combining an approach for multi-modal person tracking with the localization of sound sources and the detection of head orientation derived from a face recognition system. Note that due to the integration of cues from multiple modalities it is possible to verify the position of a speech source in 3D space using a single pair of microphones only. Persons that are observed by the robot and are also talking are considered persons of interest. If a person of interest is also facing the robot it will become the current communication partner. Otherwise the robot assumes that the speech was addressed to another person present. The performance of our approach and its robustness even in real world situations were demonstrated by quantitative evaluations in our lab and a qualitative evaluation during the exhibition of the mobile robot system at the ICVS'03. ACKNOWLEDGMENTS This work has been supported by the German Research Foundation within the Collaborative Research Center 'Situated Artificial Communicators' and the Graduate Programs 'Task Oriented Com-munication' and 'Strategies and Optimization of Behavior'. REFERENCES [1] M. Andersson, A. Oreback, M. Lindstrom, and H. I. Christensen. ISR: An intelligent service robot. In H. I. Christensen, H. Bunke, and H. Noltmeier, editors, Sensor Based Intelligent Robots; International Workshop Dagstuhl Castle, Germany, September/October 1998, Selected Papers, volume 1724 of Lecture Notes in Computer Science, pages 287310. Springer, New York, 1999. [2] B. Berdugo, J. Rosenhouse, and H. Azhari. Speakers' direction finding using estimated time delays in the frequency domain. Signal Processing, 82:1930, 2002. [3] W. Burgard, A. B. Cremers, D. Fox, D. Hahnel, G. Lakemeyer, D. Schulz, W. Steiner, and S. Thrun. The interactive museum tour-guide robot. In Proc. Nat. Conf. on Artificial Intelligence (AAAI), pages 1118, Madison, Wisconsin, 1998. [4] S. Coradeschi and A. Saffiotti. Perceptual anchoring of symbols for action. In Proc. Int. Conf. on Artificial Intelligence, pages 407412, Seattle, WA, 2001. [5] Y. Freund and R. E. Shapire. A decision-theoretic generalization of on-line learning and an application to boosting. Computational Learning Theory: Eurocolt '95, pages 2327, 1995. [6] J. Fritsch, M. Kleinehagenbrock, S. Lang, T. Plotz, G. A. Fink, and G. Sagerer. Multi-modal anchoring for human-robot-interaction. Robotics and Autonomous Systems, Special issue on Anchoring Symbols to Sensor Data in Single and Multiple Robot Systems, 43(23):133147, 2003. [7] J. Fritsch, S. Lang, M. Kleinehagenbrock, G. A. Fink, and G. Sagerer. Improving adaptive skin color segmentation by incorporating results from face detection. In Proc. IEEE Int. Workshop on Robot and Human Interactive Communication (ROMAN), pages 337343, Berlin, Germany, 2002. [8] D. Giuliani, M. Omologo, and P. Svaizer. Talker localization and speech recognition using a microphone array and a cross-powerspectrum phase analysis. In Proc. Int. Conf. on Spoken Language Processing, volume 3, pages 12431246, Yokohama, Japan, 1994. [9] E. Hjelmas and B. K. Low. Face detection: A survey. Computer Vision and Image Understanding, 83(3):236274, 2001. [10] Y. Huang, J. Benesty, G. W. Elko, and R. M. Mersereau. Real-time passiv source localization: A practical linear-correction least-square approach. IEEE Trans. on Speech and Audio Processing, 9(8):943956, 2001. [11] C. H. Knapp and G. C. Carter. The generalized correlation method for estimation of time delay. IEEE Trans. on Acoustics, Speech and Signal Processing, ASSP-24(4):320327, 1976. [12] Y. Matsusaka, S. Fujie, and T. Kobayashi. Modeling of conversational strategy for the robot participating in the group conversation. In Proc. European Conf. on Speech Communication and Technology, pages 21732176, Aalborg, Denmark, 2001. [13] H. G. Okuno, K. Nakadai, and H. Kitano. Social interaction of humanoid robot based on audio-visual tracking. In Proc. Int. Conf. on Industrial and Engineering Applications of Artificial Intelligence and Expert Systems, Cairns, Australia, 2002. [14] V. Pavlovic, A. Garg, J. Rehg, and T. Huang. Multimodal speaker detection using error feedback dynamic bayesian networks. In Proc. Int. Conf. on Computer Vision and Pattern Recognition, pages 3443, Los Alamitos, CA, 2000. [15] H. Rowley, S. Baluja, and T. Kanade. Neural network-based face detection. IEEE Trans. on Pattern Analysis and Machine Intelligence, 20(1):2338, 1998. [16] R. D. Schraft, B. Graf, A. Traub, and D. John. A mobile robot platform for assistance and entertainment. Industrial Robot, 28(1):2934, 2001. [17] D. Schulz, W. Burgard, D. Fox, and A. B. Cremers. Tracking multiple moving objects with a mobile robot. In Proc. Int. Conf. on Computer Vision and Pattern Recognition, volume 1, pages 371377, Kauwai, Hawaii, 2001. [18] R. Stiefelhagen, J. Yang, and A. Waibel. Estimating focus of attention based on gaze and sound. In Workshop on Perceptive User Interfaces, Orlando, FL, 2001. [19] M. Turk and A. Pentland. Eigenfaces for recognition. Journal of Cognitive Neuro Science, 3(1):7186, 1991. [20] P. Viola and M. Jones. Robust real-time object detection. In Proc. IEEE Int. Workshop on Statistical and Computational Theories of Vision, Vancouver, Canada, 2001. [21] J. Yang and A. Waibel. A real-time face tracker. In Proc. IEEE Workshop on Applications of Computer Vision, pages 142147, Sarasota, Florida, 1996. [22] M. H. Yang, D. J. Kriegman, and N. Ahuja. Detecting faces in images: A survey. IEEE Trans. on Pattern Analysis and Machine Intelligence, 24(1):3458, 2002. [23] A. L. Yuille, P. W. Hallinan, and D. S. Cohen. Feature extraction from faces using deformable templates. Int. Journal of Computer Vision, 8(2):99111, 1992. [24] Z. Zhang, L. Zhu, S. Z. Li, and H. Zhang. Real-time multi-view face detection. In Proc. Int. Conf. on Automatic Face and Gesture Recognition, Washington, DC, 2002. 35

Multi-modal person tracking;Attention;Human-robot-interaction

155

Psychologically Targeted Persuasive Advertising and Product Information in E-Commerce

In this paper, we describe a framework for a personalization system to systematically induce desired emotion and attention related states and promote information processing in viewers of online advertising and e-commerce product information. Psychological Customization entails personalization of the way of presenting information (user interface, visual layouts, modalities, structures) per user to create desired transient psychological effects and states, such as emotion, attention, involvement, presence, persuasion and learning. Conceptual foundations and empiric evidence for the approach are presented.

INTRODUCTION Advertising and presentation of product information is done both to inform people about new products and services and to persuade them into buying them. Persuasion can be thought of as influencing peoples attitudes and behavior. Advertising done in a mass medium, such as television or magazines can be segmented to desired audiences. However, there is a possibility to personalize or mass customize advertising in the internet and for instance in mobile phones. Similarly, in the internet also product information of various items for sale can be personalized to desired users. These two areas are introduced here together as they represent interesting opportunities for personalization. Consequently, personalization may turn out to be an important driver for future commercial applications and services in a one-to-one world in which automatic and intelligent systems tailor the interactions of users, contexts and systems in real-time. This paper describes the foundations of information personalization systems that may facilitate desired psychological states in individual users of internet based advertising and product information presentation in e-commerce thereby creating psychologically targeted messages for users of such systems. It is preliminarily hypothesized that such personalization may be one way to more efficient persuasion. When perceiving information via media and communications technologies users have a feeling of presence. In presence, the mediated information becomes the focused object of perception, while the immediate, external context, including the technological device, fades into the background [8, 36, 37]. Empirical studies show that information experienced in presence has real psychological effects on perceivers, such as emotional responses based on the events described or cognitive processing and learning from the events [see 51]. It is likely that perceivers of advertisements and product information experience presence that may lead to various psychological effects. For instance, an attitude may be hold with greater confidence the stronger the presence experience. Personalization and customization entails the automatic or semiautomatic adaptation of information per user in an intelligent way with information technology [see 33, 68]. One may also vary the form of information (modality for instance) per user profile, which may systematically produce, amplify, or shade different psychological effects [56, 57, 58, 59, 60, 61, 62, 63]. Media- and communication technologies as special cases of information technology may be considered as consisting of three layers [6]. At the bottom is a physical layer that includes the physical technological device and the connection channel that is used to transmit communication signals. In the middle is a code layer that consists of the protocols and software that make the physical layer run. At the top is a content layer that consists of multimodal information. The content layer includes both the substance and the form of multimedia content [7, 56]. Substance refers to the core message of the information. Form implies aesthetic and expressive ways of organizing the substance, such as using different modalities and structures of information [56]. With the possibility of real-time customization and adaptation of information for different perceivers it is hypothesized that one may vary the form of information within some limits per the same substance of information. For instance, the same substance can be expressed in different modalities, or with different ways of interaction with the user and technology. This may produce a certain psychological effect in some perceivers; or shade or amplify a certain effect. In Figure 1 the interaction of media and communications technology and the user in context with certain types of tasks is seen as producing transient psychological effects, thereby creating various "archetypal technologies" that systematically facilitate desired user experiences [see 55, 56]. Media and communication technology is divided into the physical, code and content layers. The user is seen as consisting of various different psychological profiles, such as individual differences related to cognitive style, personality, cognitive ability, previous knowledge (mental models related to task) and other differences, such as pre-existing mood. [49, 56, 58, 59] Media- and communication technologies may be called Mind-Based if they simultaneously take into account the interaction of three different key components: i) the individual differences and/or user segment differences of perceptual processing and sense making ii) the elements and factors inherent in information and technology that may produce psychological effects (physical, code and content layers), and iii) the consequent transient psychological effects emerging based on perception and processing of information at the level of each individual. [see 63] This definition can be extended to include both context and at least short-term behavioral consequences. Regarding context, a Mind-Based system may alter its functionalities depending on type of task of the user, physical location, social situation or other ad-hoc situational factors that may have a psychological impact. Behavioral consequences of using a Mind-Based system may be thought of especially in the case of persuasion as facilitating desired instant behaviors such as impulse buying. Of course, if a Mind-Based system builds a positive image and schema of a product over longer periods of time reflected in product and brand awareness that may the influence user behaviors later on. As the task of capturing and predicting users psychological state in real time is highly complex, one possible realization for capturing users psychological state is to have the user linked to a sufficient number of measurement channels of various i) psychophysiological signals (EEG, EMG, GSR, cardiovascular activity, other), ii) eye-based measures (eye blinks, pupil dilation, eye movements) and iii) behavioral measures (response speed, response quality, voice pitch analysis etc.). An index based on these signals then would verify to the system whether a desired psychological effect has been realized. Fig. 1. Mind-Based Technologies as a framework for producing psychological effects. Adapted from [56]. Another approach would be to conduct a large number of user studies on certain tasks and contexts with certain user groups, psychological profiles and content-form variations and measure various psychological effects as objectively as possible. Here, both subjective methods (questionnaires and interviews) and objective measures (psychophysiological measures or eye-based methods) may be used as well interviews [for a review on the use of psychophysiological methods in media research, see 46]. This would constitute a database of design-rules for automatic adaptations of information per user profile to create similar effects in highly similar situations with real applications. Naturally, a hybrid approach would combine both of these methods for capturing and facilitating the users likely psychological state. Capturing context and short-term user behavior is a challenge. Computational approach to context utilizes a mass of sensors that detect various signals in an environment. AI-based software then massively computes from the signal flow significant events either directly or with the help of some simplifying rules and algorithms. Capturing user behavior in context is easier if the user is using an internet browser to buy an item, for instance. In this case action, or behavior can be captured by the system as the user clicks his mouse to buy an item. If the user is wondering around in a supermarket with a mobile phone that presented a persuasive message to buy the item on aisle 7 it may be difficult to verify this other than cross-reference his checkout bill with the displayed adverts inside the store. However, it is beyond the scope of this paper to fully elaborate on the contextual and behavioral dimensions of Mind-Based Technologies. 2.2 Description of a Psychological Customization System Psychological Customization is one possible way of implementing of Mind-Based Technologies in system design. It can be applied to various areas of HCI, such as Augmentation Systems (augmented and context sensitive financial news), Notification Systems (alerts that mobilize a suitable amount of attention per task or context of use), Affective Computing (emotionally adapted games), Collaborative Filtering (group-focused information presentation), Persuasive Technology Media and Communications Technology Substance Form - modalities - visual layout - structure - other Type of device Ways of interaction User interface Technologies influencing subjective experience -Emotion Technologies -Flow Technologies -Presence Technologies Technologies influencing subjective experience -Emotion Technologies -Flow Technologies -Presence Technologies Technologies influencing change of knowledge - Knowledge Technologies Technologies influencing change of knowledge - Knowledge Technologies Mind Psychological profiles - cognitive style - personality - mental models - other Results of processing Results of processing Physical Content Code Context and Task 246 246 (advertising for persuasion, e-commerce persuasion), Computer Mediated Social Interaction Systems (collaborative work, social content creation templates), Messaging Systems (emotionally adapted mobile multimedia messaging and email) and Contextually Sensitive Services (psychologically efficient adaptation of presentation of information sensitive to physical, social or situational context, such as available menus to control a physical space, available information related to a particular situation, such as social interaction or city navigation with a mobile device). It can be hypothesized that the selection and manipulation of substance of information takes place through the technologies of the various application areas of Psychological Customization. Underlying the application areas is a basic technology layer for customizing design. This implies that within some limits one may automatically vary the form of information per a certain category of substance of information. The design space for Psychological Customization is formed in the interaction of a particular application area and the possibilities of the technical implementation of automated design variation Initially, Psychological Customization includes modeling of individuals, groups, and communities to create psychological profiles and other profiles based on which customization may be conducted. In addition, a database of design rules is needed to define the desired cognitive and emotional effects for different types of profiles. Once these components are in place, content management technologies can be extended to cover variations of form and substance of information based on psychological profiles and design rules to create the desired psychological effects. [see 63] At the technically more concrete level, a Psychological Customization System is a new form of middleware between applications, services, content management systems and databases. It provides an interface for designing desired psychological effects and user experiences for individual users or user groups. The most popular framework for building customized Web-based applications is Java 2 Enterprise Edition. J2EE-based implementation of the Psychological Customization System for Web-based applications is depicted in Figure 2. The basic J2EE three-tiered architecture consisting of databases, application servers, and presentation servers has been extended with three middleware layers: content management layer, customer relationship management layer, and psychological customization layer. The profiles of the users and the communities are available in the profile repository. [see 69] The Content Management System is used to define and manage the content repositories. This typically is based on metadata descriptions of the content assets. The metadata of the content repositories is matched against the user and community profiles by the Customer Relationship Management (CRM) system. The CRM system includes tools for managing the logic behind content, application and service customization. Rules can be simple matching rules or more complex rule sets. A special case of a rule set are scenarios, which are rule sets involving sequences of the interactions on the Web site. The Customer Relationship Management layer also includes functionality for user and community modeling. This layer can also perform automated customer data analysis, such as user clustering. [see 69] The Psychological Customization System layer performs the optimization of the form of the content as selected by the Customer Relationship Management layer. This functionality can be considered similar to the device adaptation by using content transformation rules (for example XSL-T). In the case of the psychological customization, the transformation rules are produced based on the design rules for content presentation variation and the contents of the psychological profile of the user. After this optimization, the content is passed to the Web presentation layer. Community Profiles Content Rules and Scenarios Application Server User User User User profile Community profile Content Management System Customer Relationship Management Web Presentation Layer Psychological Customization System Figure 2. J2EE implementation of the Psychological Customization System [69] Even though a working prototype of Psychological Customization has not been built yet, several empirical studies support the feasibility of a user-experience driven system that matches the form of information to the psychologically relevant properties and other profile factors of individual users and user groups. For instance, there are individual differences in cognitive processes such as attention, working memory capacity, general intelligence, perceptual-motor skills and language abilities. These individual differences have a considerable effect on computer-based performance and may product sometimes quite large variance in the intensity or type of psychological effects, such as depth of learning, positive emotion, persuasion, presence, social presence and other types of psychological states and effects as well as consequent behavior [13, 14, 18, 56, 57, 58, 59, 60, 61, 62, 63, 70]. There is considerable evidence in literature and in our own experimental research that varying the form of information, such as modality, layouts, background colors, text types, emotionality of the message, audio characteristics, presence of image motion and subliminality creates for instance emotional, cognitive and attentional effects [9, 25, 27, 28, 29, 30, 31, 32, 33, 34, 48]. Some of these effects are produced in interaction with individual differences, such as cognitive style, personality, age and gender [21, 46, 47], or pre-existing mood [49]. The role of hardware should not be neglected. A device with a large screen or a 247 247 portable device with smaller screen with user-changeable covers may also influence the emerging effects [e.g. 30]. Table 1. Key factors influencing psychological effects. Adapted from [56]. Layer of technology Key factors Physical Hardware - large or small vs. human scale - mobile or immobile - close or far from body (intimate-personal -social distance) Interaction - degree of user vs. system control and proactivity through user interface Code Visual-functional aspects - way of presenting controls in an interface visually and functionally Substance - the essence of the event described - type of substance (factual/imaginary; genre, other) - narrative techniques used by authors Content Form 1. Modalities - text, video, audio, graphics, animation, etc. 2. Visual layout - ways of presenting various shapes, colours, font types, groupings and other relationships or expressive properties of visual representations - ways of integrating modalities into the user interface 3. Structure - ways of presenting modalities, visual layout and other elements of form and their relationships over time - linear and/or non-linear structure (sequential vs. parallel; narrative techniques, hypertextuality) This empiric evidence partly validates the possibility for Psychological Customization Systems at least with mobile devices and user interface prototypes used in our own research. Typical experiments we have conducted on the influence of form of information on psychological effects have included such manipulations as animation and movement (for orientation response), fonts of text, layout of text, background colors of text, user interface navigation element shapes (round vs. sharp), user interface layout directions, adding background music to reading text, use of subliminal affective priming in the user interface (emotionally loaded faces) and use of different modalities, for instance. Table 1 addresses the key factors that may influence psychological effects of processing mediated information. APPLICATION AREAS The focus of this paper is on persuasion in advertising and product information presentation in e-commerce. The key application area to realize this with Psychological Customization is Persuasive Technology. It refers to human-computer interaction in which there is an underlying goal to non-coercively change the attitudes, motivation and/or behavior of the user [15, 16]. For instance, one may motivate users to quit smoking via motivating games. However, it is clear that how much people allocate resources to processing a particular persuasive message has to be taken into account. Further, it may be that there is not so much freedom to manipulate persuasive messages to produce effects and the effects themselves may be sometimes small. Despite this, empiric evidence in personalization, as discussed, suggests that statistically significant effects perhaps in the range of a few percentages to even tens of percents exist in the area of Psychological Customization, such as emotion, presence and efficiency of information processing. Hence, it can be at least preliminarily assumed that with persuasion similar level effects may be achievable also. Persuasion in human computer interaction has been researched from the point of view of seeing computers as tools (increasing capabilities of the user), as a medium (providing experiences) and as a social actor (creating a relationship) [15, 16]. For the purposes of this article, technology used in Psychological Customization for presentation of e-commerce product information and online advertising is seen mostly as a medium and partly as a social actor. How then to model and explain persuasion in more detail? Evidently no universal theory of what is the process of persuasion has been created yet [17]. Candidates for explaining and modeling persuasion include i) learning theory (operant conditioning), ii) functional paradigm theory (similarity-attraction, pleasure seeking), iii) cognitive consistency theory (new information creates tension that needs to be relieved by adopting schemata), iv) congruity principle theory (interpretations of new information tend to be congruent with existing schemata), v) cognitive dissonance theory (certain actions and information produce tension that needs to be relieved by adopting mental structures or behavior), counter-attitudinal advocacy (belief-discrepant messages are persuasive), vi) inoculation theory (combining supportive and refutational information to achieve better persuasion) and vii) attribution theory (people make simple models to predict events of the world and behaviors of other people). [for a review, see 50].Some contemporary models of persuasion are i) social learning theory (environmental learning is the source of persuasion, such as social relationships), ii) the elaboration likelihood model (a specific and limited model on how a piece of information may influence the attitudes of the receiver) and iii) the communication/persuasion model (the source, the message content and form, the channel, the properties of the receiver and the immediacy of the communication influence persuasion). [2, 39, 42] The latter approach partly resembles the approach of Mind-Based Technologies as a way of finding out the values of relevant parameters in the layers of technology, the user and the transient results of processing, such as emotion and cognition. Other frameworks have also been presented. Meyers-Levy and Malaviya (1998) have presented a framework introducing several 248 248 strategies to process persuasive messages. Each strategy represents a different amount of cognitive resources employed during processing and may influence the level of persuasion. [38] The position of the authors is that while various theories and models for persuasion have been presented, within the context of personalized information presentation especially by varying both the substance of the message and the form of the message it is difficult to know what types of persuasive effects may emerge. This is partly due to the fact that especially the perception of form of information is most likely not a conscious process involving in-depth processing and cognitive appraisal but a rather automatic and non-cognitive process. Hence, if one influences the conditions of perceptual processing or some early-level cognitive processing of multimodal information, no clear models are available for explaining and predicting persuasion. Also, the exact influence of the amount of cognitive resources employed during early and later processing of a persuasive message remains unknown. It is most evident that case studies with particular application are needed to verify such effects. However, the authors present one possible way of seeing persuasion mostly via a link to transient emotional states and moods immediately before, during and after processing information presented through a Psychological Customization system. Yet, based on this approach the claim for more efficient persuasion in each application area, such as using Psychological Customization in advertising or e-commerce product information remains a complex task. Despite this difficulty, we now present a possible selection of relevant psychological principles related to perceptual processing and persuasion of advertising and e-commerce product information. First, a similarity-attraction process may arise between the presented information and the personality of the user that may lead to the information being processed more fully [i.e., trait-congruency hypothesis; see 54]. That is, users are likely to be attracted to information with content and formal characteristics manifesting a personality similar to their own [see e.g., 21]. Second, the decrease of cognitive load in perceptual processing (i.e., high processing fluency) may induce a feeling of pleasantness that may label the information processed [for a review, see 65]. That is, fluent stimuli are associated with increased liking and positive affective responses as assessed by facial EMG, for example. Third, the creation of specific emotional reactions and moods varying in valence and arousal may label the information processed; here the effects may depend on the type of emotion. For instance, mood-congruency may provide more intensive engagement with the information presented when the mood induced by the information processed matches a pre-existing mood of the user [see 49]. Fourth, the emotional reactions may induce increased attention that may lead to more in-depth processing of information [e.g., 26]. Fifth, as suggested by excitation transfer theory, arousal induced by a processed stimuli influences the processing of subsequent stimuli [see 71, 72]. Sixth, according to selective-exposure theory, individuals are motivated to make media choices in order to regulate their affective state [i.e., to maintain excitatory homeostasis; 73]. This may mean that people use also e-commerce product information to manage their moods, i.e. neutralize an unwanted mood, such as depression by engaging with exciting and positive product information. Users may also intensify an existing mood by selecting product information content that may add to the present mood. Seventh, affective priming research indicates that the valence of subliminally exposed primes (e.g., facial expressions) influences the affective ratings of subsequent targets [40, 43], including video messages presented on a small screen [48]. Eighth, the perceived personal relevance of the particular information to the user exerts a robust influence on message processing and involvement [64]. This means that if the user is interested in the product described in the information presented, he will be quite involved when processing the information and hence his memory of the product will be enhanced. Consequently, it has been shown that information tailored to the needs and contexts of users often increases the potential for attitude and behavior change [5, 11, 41, 66, 67]. Further, there is quite a lot of research indicating that, when compared to video form, text has a greater capacity to trigger active or systematic message processing and is perceived as more involving [see 44]; this depends on the mood of the user, however [48]. Ninth, some emotional states and moods lead to secondary effects related to decision-making, judgment and behavior [4, 10, 20]. It then seems that indeed a relevant area to focus on regarding persuasion with Psychological Customization is emotion (arousal and valence) immediately before, during and right after viewing product information and ads. One may focus on &quot;primitive&quot; emotional responses or emotions requiring more extensive cognitive appraisal and processing. Both of these types of emotions can be linked to various psychological consequences. Consequently, with emotionally loaded personalized information products one may focus on i) creating immediate and primitive emotional responses, ii) creating mood and iii) indirectly influencing secondary effects of emotion and mood, such as attention, memory, performance and judgment. Known psychological mechanisms used to create desired emotions or moods would be for instance similarity attraction (trait congruency), decrease of cognitive load (high processing fluency), mood congruency, excitation transfer, mood management and affective priming. These mechanisms are not without problems as they may have also opposite effects. For instance mood congruency may decrease attention and hence lessen the mobilization of cognitive resources in processing a persuasive message. Also, even though emotion is good candidate to look for a strong link to persuasion, the exact nature of this link is unclear. The key idea of using emotion as a hypothesized gateway to persuasion would be that more in-depth processing of information caused by arousal, valence, attention or involvement may lead to increased memory and perceived trustworthiness of information and also influence attitudes towards the product in question [e.g. 65]. This in turn may lead to instant behavior, such as buying online, clicking through an ad or purchasing the item later in a department store based on long-term memory schemata. It should be noted that this view is based on empiric evidence of the psychological effects and their consequences in general, but they have not yet been validated with the use of e-commerce systems that personalize the form of information for persuasion. Hence, it would be most beneficial to capture the users emotional states or mood before the user starts to browse a particular piece of product information to be able to automatically realize various effects with adaptation of the form of information and track the changes of the online behavior of the user. 249 249 3.2 Persuasive Advertising The effectiveness of persuasion in advertisements in general is a complex issue. Subliminal priming, use of commonly known symbols, matching the advertisement to basic biological needs, such as food, shelter and sex, maximizing the credibility of the message, telling a compelling story, creating a desirable image of the perceiver with the product, placing TV-ads immediately after emotional (arousal) and attentional peaks of TV-programming and other approaches have been widely used. However, research into effectiveness of the form of presentation of advertising is not widely available in the scientific community. Moreover, little research has been done to understand the psychological effectiveness of online advertising. It should be also noted that advertising may be mostly a creative and design-driven high-speed field of industrial production in which various types of authors and artists collaborate like in film-production to make the advertisement rather than a field filled with scientists attempting to analyze the advertisements and their effects in great detail. In internet-based advertising the advertisement is typically presented on a web page as a banner. The banner is embedded in editorial content, such as the front page of a magazine or online newspaper. The banners are often placed according to the number and demography of the visitors on a particular section of editorial content. This means a best guess is taken as to what may be the most efficient and visible way of placing the banner based on previous knowledge of the behavior of desired user segments on the website. It seems that ads placed in context work best also online. This means that an ad that is related to the editorial content it is displayed with is more efficiently persuasive [3]. Another issue is that text-based ads online may work better than only graphics. This implies that most ads contain text based over a graphical surface. Overall, very simple principles (larger is better etc.) seem to guide people's choices: e.g., larger ads are thought to be more appealing and affective. Further, in mobile contexts personalized advertising has been studied from the point of view of targeting users by emotions in addition to location and other relevant factors [19]. However, the exact transient psychological influence of a particular piece of editorial content the online advertisement is displayed with remains unknown. It is possible that the editorial content repels the user and the advertisement is also labeled by this emotion. It is also possible that the editorial content induces a positive emotion and the advertisement gets an advantage based on this emotional state. The emotional tone of the advertisements and editorial content may also interact. For example, Kamins, Marks, and Skinner (1991) showed that commercials that are more consistent in emotional tone with the TV program perform better as measured by likeability and purchase intention ratings than those that are inconsistent in tone. Sometimes advertisements are changed in real-time per type of user as the system recognizes a user segment to which a certain banner has been targeted. However, what is lacking here is i) more detailed information of the type of user (such as what may be the most efficient way to influence him psychologically) and ii) what may be the psychological impact on the same user of the editorial content within which the banner is placed. [22] With a Psychological Customization system some of these gaps may be at least indirectly addressed as presented in Table 2. Table 2. Technological possibilities of persuasive advertising with Psychological Customization. Layer of Technology Adaptations of Advertising Banners 1. Physical -multimedia PC or mobile device -The advertisement substance and form may be matched to the technology used by lifestyle segments or other means of segmentation (hip ads for mobile phones etc.) -Mobile device: user changeable covers in colors and shapes that facilitate emotion 2. Code -Windows-type user interface -Mouse, pen, speech, -The user interface elements (background color, forms, shapes, directions of navigation buttons etc.) may be varied in real-time per page per user in which a certain advertisement is located to create various emotions and ease of perceptual processing -audio channel may be used to create emotional effects (using audio input/output sound, varying pitch, tone, background music, audio effects etc.). 3. Content A. Substance - Fixed multimedia content -The editorial content may be matched with the ad -The content of the ad may be matched to the users based on various factors (interests, use history, demography, personality etc.) -Adding subliminal extra content to create emotion B. Form Modality -Multimedia -Modality may be matched to cognitive style or preexisting mood of the enable easier processing. -Background music, audio effects or ringing tones may be used as a separate modality to facilitate desired emotions and moods. -Animated text can be used to create more efficient processing of text facilitate some emotional effects. Visual presentation -Emotionally evaluated and positioned layout designs and templates for ads (colors, shapes and textures) may be utilized per type of user segment Structure -temporal, other -Offering emotionally evaluated and positioned narrative templates for creating emotionally engaging stories. Based on Table 2 a Psychological Customization system may operate by trying to optimize desired emotional effects that may be related to persuasion. The content provider, such as a media company, is able to set desired effects per type of user group and advertiser need by using a Psychological Customization system. Also, the placement of ads within desired editorial contexts may be utilized with a more developed system. When a user logs in with his profile already to the database of the content provider the system will start real-time personalization of form of information. As the user has logged in, the front page of the service may be altered for him according to advertiser needs. As the user navigates the system and consumes information the system follows ready-set effects to be realized to the user. It is clear that such a scenario is difficult, but if it is done in a simple enough manner it may be that the persuasive efficiency of online advertising may increase. 3.3 Persuasive e-Commerce Product Information Personalized e-commerce has not been studied widely. It has been found that while personalization of the content substance displayed to each user may provide value, the users have a strong motivation to understand why the system is displaying a particular piece of information for them. Also, users wish to be in control of their user profiles. [see 1, 23] Hence, it seems that users are at least partly suspicious to the system adapting the substance of information to them. However, 250 250 in many cases it may be possible to adapt the form of information in personalized applications in conjunction to content substance variation or even without it. The adaptation of form of information to the user may even be a more transparent way of personalizing user-system- interaction as the user most likely does not question the form of a particular substance. Hence, there are emerging possibilities for personalization and customization in this area. There are at least two different types of advanced e-commerce systems commonly used: i) systems using recommendation engines and other personalization features to present information in a media-like manner and ii) systems using persuasive interface agents, creating a relationship between the user and the agent. The focus here is mostly on presentation of product information, such as information (product properties, comparisons, pricing, functionalities and other information) of a new car, digital camera, computer or garment. Although, in the context of product presentation, users have usually been suggested to prefer a combination site including pictures and text [e.g. 35], individual differences in their preferences are likely to occur. The technological possibilities for persuasive presentation of product information are much like those presented for persuasive advertising seen in Table 2. In other words, different layers of technology may be adapted to the user of an e-commerce system to create various psychological effects when presenting product information. With personalized e-commerce systems for product information presentation one may facilitate positive emotional responses for instance by selecting the modalities of the information to be displayed according to the processing styles and alter visual layouts of the interface according to the personalities of the users. The ease of processing information and the similarity-attraction between visual layouts and personalities may create positive emotional states. As for brand awareness one may indirectly influence memory with the facilitation of positive emotion and increase memory-based performance on the task such as brand recognition and recall. By increasing attention one may increase the likelihood of the user of an e-commerce system to learn product information more efficiently. Positive emotion and mood also has the effect of making the user adapt a less risk-prone approach to making decisions [20]. This may be used to present product information in a familiar manner creating a safe atmosphere around the product to make it more desirable when the user is making purchasing decisions in a positive mood. Psychological Customization may be used for persuasion also with recommendation systems. The system knows the users profile, such as type of personality, and the desired psychological effect is set to positive emotion in as many page-views of the recommendations as possible. The user starts using the system and finds an interesting product that the system recommends to her. The form of the recommendation information is tailored to the users profile and desired psychological effect in real-time when the page uploads to make the realization of positive emotion as probable as possible. The system may select the modality of recommendation from text to audio, or from audio to animation; the system may change the background colors of the page and modify the shape and color of the navigation buttons, for instance. In this case, the system will try to do everything possible to facilitate positive emotion but change the substance of the recommendation itself. Naturally in some cases depending on type of user and the type of recommendation, the available databases of recommendation information and the available means of Psychological Customization of form of recommendation information, the effect to be achieved is more or less likely to occur. However, even effects that provide some percentages or even tens of percents of more targeted positive emotion may make a difference in persuasion and hence attitudes towards the product and buying behavior. This is especially true if the recommendation system website has masses of users and hence even a slight increase in sales effectiveness may add up to significant amounts of revenue. Further, one may discuss interface agents for product information presentation. Often with interface agents an illusion of being in interaction with another human being is created in the user via using for example animated agents that seemingly exhibit various human properties, such as gender, personality, politeness, group membership and other factors. Here one possible application would be to add an agent to float atop of a page with product information to comment or recommend it, to aid the user in navigation and finding interesting information and to act as a feedback channel for the user, such as collecting the users interest profile or other situational relevant information. It is known that both the substance of the interaction (what is being sold, or what information is presented, and what the agent says, or how it acts) and the form of interaction (how information is presented, what is the appearance and personality or other factors of the agent) influence for instance trust, persuasiveness, emotion and liking of the transaction [e.g. 51]. What Psychological Customization may add here may be more systematic and efficient personalization of the way of presenting information together with customizing the selected appearance and other features of the agent in without actually changing the substance of the interaction, i.e. what the agent says or what product information is presented, for instance. CONCLUSION The authors believe that no other comprehensive framework of varying form of information to systematically create emotional and cognitive effects has been presented, specifically in persuasive presentation of online advertising and product information in e-commerce sites. Differences to other approaches to influencing user experience in general are various. Usability studies traditionally address the question of how to make difficult technology easy to use. Usability is at least partly founded on the idea of optimal human-machine performance, i.e. how well a user can manipulate and control a machine. However, there is a growing conviction that, in order to ensure high user satisfaction usability is not sufficient [see 12, 24]. The approach to system design presented in this paper may be beneficial to the fields of e-commerce and online advertising because: i) it provides a possibility to personalize the form of information that may be more transparent and acceptable to the users than adapting the substance of information, ii) it offers a way of more systematically accessing and controlling transient psychological effects of users of e-commerce and advertisement displaying systems, iii) it offers possibilities to more efficiently persuade and consequently influence behavior of individual users and iv) it is compatible with existing and new systems (recommendation engines, click-through-systems, other) as an add-on or a middleware layer in software with many potential application areas. 251 251 The potential drawbacks of the framework include the following: i) it may be costly to build the design-rule databases and actually working real-life systems for creating systematic psychological effects, ii) the rule-databases may have to be adapted also locally and culturally, iii) the method needed to create a rule-database is not easy to use and may be suspect to ecological validity (eye-tracking , behavioral and psychophysiological measures, self-report , field tests are needed to verify laboratory results etc.) and iv) if the system works efficiently it may raise privacy issues, such as the intimacy of a personal psychological user profile (personality, cognitive style, values, other). Also ethical issues related to mind-control or even propaganda may arise. It should be noted that to build a smoothly functioning Psychological Customization system one should do much more research and gain more evidence of the systematic relationships of user profiles, information forms and psychological effects. However, in our research for the past four years we have found many feasible rules for personalization for psychological effects. Regarding future research, content management technologies should be elaborated to provide for the platform that prototypes can be built on. Consequently, we aim to build, evaluate and field-test prototypes of Psychological Customization in various areas, specifically in mobile, urban ad-hoc contexts and situations related to mobile advertising and e-commerce, but also other areas such as mobile gaming communities, mobile content, mobile messaging, knowledge work systems and city navigation. REFERENCES [1] Alpert, S. R.; Karat, J.; Karat, C-M; Brodie, C and Vergo, J.G. (2003) User Attitudes Regarding a User-Adaptive eCommerce Web Site. User Modeling and User-Adapted Interaction 13 (4): 373-396, November 2003. [2] Bandura (1977) Social learning theory. Englewood Cliffs, NJ: Prentice-Hall. [3] Baudisch, P. and Leopold, D. (2000) Attention, indifference, dislike, action: Web advertising involving users. Netnomics Volume 2 , Issue 1 2000, pp. 75-83. [4] Brave S. and Nass, C. (2003). Emotion in human-computer interaction. In Jacko, J.A. and Sears, A. (Ed.), The Human-Computer Interaction Handbook. Fundamentals, Evolving Technologies and Emerging Applications. (pp. 81-96). London : Lawrence Erlbaum Associates. [5] Beniger, J. R. (1987) Personalization of mass media and the growth of pseudo-community. Communication Research, 14(3), pp 352-371. [6] Benkler, Y. (2000) From Consumers to Users: Shifting the Deeper Structures of Regulation. Federal Communications Law Journal 52, 561-63. [7] Billmann, D. (1998) Representations. In Bechtel, W. and Graham, G. (1998) A companion to cognitive science, 649-659 . Blackwell publishers, Malden, MA. [8] Biocca, F. and Levy, M. (1995) Communication in the age of virtual reality. Lawrence Erlbaum, Hillsdale, NJ. [9] Cuperfain, R. and Clarke, T. K. (1985) A new perspective on subliminal perception. Journal of Advertising, 14, 36-41. [10] Clore, G. C. and Gasper, K. (2000). Feeling is believing. Some affective influences on belief. In Frijda, N.H., Manstead, A. S. R. and Bem, S. (Ed.), Emotions and beliefs: How feelings influence thoughts (pp. 10-44). Paris/Cambridge: Editions de la Maison des Sciences de lHomme and Cambridge University Press. [11] Dijkstra, J. J., Liebrand, W.B.G and Timminga, E. (1998) Persuasiveness of expert systems. Behavior and Information Technology, 17(3), pp 155-163. [12] Dillon, A. (2001). Beyond usability: process, outcome and affect in human computer interactions. Online: http://www.ischool.utexas.edu/~adillon/publications/beyond _usability.html. [13] Egan, D. E. (1988). Individual differences in human-computer interaction. In: M. Helander (Ed.), Handbook of Human-Computer Interaction, p. 543 568. Elsevier, New York. [14] Eysenck, M. (1994) Individual Differences: Normal and Abnormal. New Jersey: Erlbaum. [15] Fogg, B.J. (2003) Motivating, influencing and persuading users. In Jacko, J.A. and Sears, A. (Ed.), The Human-Computer Interaction Handbook. Fundamentals, Evolving Technologies and Emerging Applications. (pp. 81-96). London : Lawrence Erlbaum Associates. [16] Fogg, B. J. (2002) Persuasive technology. Using computers to change what we think and do. Morgan Kaufmann Publishers, New York. [17] Ford, M. E. (1992) Motivating humans: goals, emotions, personal agency beliefs. Newbury Park, Ca: Sage. [18] Hampson, S. E. & Colman, A. M. (Eds., 1995) Individual differences and personality. London: Longman. [19] Hristova, N. and O'Hare, G. M. P. (2004) Ad-me: Wireless Advertising Adapted to the User Location, Device and Emotions. Proceedings of the Proceedings of the 37th Annual Hawaii International Conference on System Sciences (HICSS'04) - Track 9 - Volume 9 [20] Isen, A. M. (2000). Positive affect and decision making. In Lewis, M. and Haviland-Jones, J. M. (Ed.), Handbook of emotions (2nd ed.) (pp. 417-435). New York: Guilford Press. [21] Kallinen, K., & Ravaja, N. (in press). Emotion-related effects of speech rate and rising vs. falling background music melody during audio news: The moderating influence of personality. Personality and Individual Differences. [22] Kamins, M.A., Marks, L.J., & Skinner, D. (1991). Television commercial evaluation in the context of program induced mood: Congruency versus consistency effects. Journal of Advertising, 20, 1-14. [23] Karat, M-C., Blom, J. and Karat, J. (in press) Designing Personalized User Experiences in eCommerce. Dordrecht: Kluwer. [24] Karat, J. (2003) Beyond task completion: Evaluation of affective components of use. In Jacko, J.A. and Sears, A. (Ed.), The Human-Computer Interaction Handbook. Fundamentals, Evolving Technologies and Emerging Applications. (pp. 81-96). London : Lawrence Erlbaum Associates. [25] Kihlstrm, J. F., Barnhardt, T. M. and Tataryn, D. J. (1992) Implicit perception. In Bornstein, R. F. and Pittmann, T. S. 252 252 (eds.) Perception without awareness. Cognitive, clinical and social perspectives, 17-54. Guilford, New York. [26] Krohne, H.W., Pieper, M., Knoll, N., & Breimer, N. (2002). The cognitive regulation of emotions: The role of success versus failure experience and coping dispositions. Cognition and Emotion, 16, 217-243. [27] Krosnick, J. A., Betz, A. L., Jussim, J. L. and Lynn, A. R. (1992) Subliminal conditioning of attitudes. Personality and Social Psychology Bulletin, 18, 152-162. [28] Laarni, J. (2003). Effects of color, font type and font style on user preferences. In C. Stephanidis (Ed.) Adjunct Proceedings of HCI International 2003. (Pp. 31-32). Crete University Press, Heraklion. [29] Laarni, J. (2002). Searching for optimal methods of presenting dynamic text on different types of screens. In: O.W. Bertelsen, S. Bdker & K. Kuutti (Eds.), Tradition and Transcendence. Proceedings of The Second Nordic Conference on Human-Computer Interaction, October 19-23, 2002, Arhus, Denmark (Pp. 217 220). [30] Laarni, J. & Kojo, I.(2001). Reading financial news from PDA and laptop displays. In: M. J. Smith & G. Salvendy (Eds.) Systems, Social and Internationalization Design Aspects of Human-Computer Interaction. Vol. 2 of Proceedings of HCI International 2001. Lawrence Erlbaum, Hillsdale, NJ. (Pp. 109 113.) [31] Laarni, J., Kojo, I. & Krkkinen, L. (2002). Reading and searching information on small display screens. In: D. de Waard, K. Brookhuis, J. Moraal, & A. Toffetti (Eds.), Human Factors in Transportation, Communication, Health, and the Workplace. (Pp. 505 516). Shake, Maastricht. (On the occasion of the Human Factors and Ergonomics Society Europe Chapter Annual Meeting in Turin, Italy, November 2001). [32] Lang, A. (1990) Involuntary attention and physiological arousal evoked by structural features and mild emotion in TV commercials. Communication Research, 17 (3), 275-299. [33] Lang, A., Dhillon, P. and Dong, Q. (1995) Arousal, emotion and memory for television messages. Journal of Broadcasting and Electronic Media, 38, 1-15. [34] Lang, A., Newhagen, J. and Reeves. B. (1996) Negative video as structure: Emotion, attention, capacity and memory. Journal of Broadcasting and Electronic Media, 40, 460-477. [35] Lightner, N.J., & Eastman, C. (2002). User preference for product information in remote purchase environments. Journal of Electronic Commerce Research [Online], 3, Available from http://www.csulb.edu/web/journals/ jecr/issues/20023/paper6.pdf . [36] Lombard, M. and Ditton, T. (2000) Measuring presence: A literature-based approach to the development of a standardized paper-and-pencil instrument. Project abstract submitted to Presence 2000: The third international workshop on presence. [37] Lombard, M., Reich, R., Grabe, M. E., Bracken, C. and Ditton, T. (2000) Presence and television: The role of screen size. Human Communication Research, 26(1), 75-98. [38] Meyers-Levy, J. & Malaviya, P. (1998). Consumers' processing of persuasive advertisements. An integrative framework of persuasion theories. Journal of Marketing 63, 4560. [39] McGuire, W. J. (1989) Theoretical foundations of campaigns. In R.E. Rice and C.K. Atkin (eds.) Public communication campaigns (2 nd ed, pp 43-65). Newbury Park, Ca: Sage. [40] Monahan, J.L. (1998). I don't know it but I like you: The influence of nonconscious affect on person perception. Human Communication Research, 24, 480-500. [41] Nowak, G. J., Shamp, S., Hollander, B., Cameron, G. T., Schumann, D. W. and Thorson, E. (1999) Interactive media: A means for more meaningful advertising? Advertising and consumer psychology. Mahwah, NJ: Lawrence Erlbaum. [42] Petty, R.E., & Cacioppo, J.T. (1986). Communication and persuasion: Central and peripheral routes to attitude change. New York: Springer-Verlag. [43] Murphy, S.T., Monahan, J.L., & Zajonc, R.B. (1995). Additivity of nonconscious affect: Combined effects of priming and exposure. Journal of Personality and Social Psychology, 69, 589-602. [44] Pfau, M., Holbert, R.L., Zubric, S.J., Pasha, N.H., & Lin, W.-K. (2000). Role and influence of communication modality in the process of resistance to persuasion. Media Psychology, 2, 1-33. [45] Ravaja, N. (2004). Effects of a small talking facial image on autonomic activity: The moderating influence of dispositional BIS and BAS sensitivities and emotions. Biological Psychology, 65, 163-183. [46] Ravaja, N. (in press). Contributions of psychophysiology to media research: Review and recommendations. Media Psychology. [47] Ravaja, N., & Kallinen, K. (in press). Emotional effects of startling background music during reading news reports: The moderating influence of dispositional BIS and BAS sensitivities. Scandinavian Journal of Psychology. [48] Ravaja, N., Kallinen, K., Saari, T., & Keltikangas-Jrvinen, L. (in press). Suboptimal exposure to facial expressions when viewing video messages from a small screen: Effects on emotion, attention, and memory. Journal of Experimental Psychology: Applied. [49] Ravaja, N., Saari, T., Kallinen, K., & Laarni, J. (2004). The Role of Mood in the Processing of Media Messages from a Small Screen: Effects on Subjective and Physiological Responses. Manuscript submitted for publication. [50] Reardon, K. R. (1991) Persuasion in practice. Newbury Park, Ca: Sage. [51] Reeves, B. and Nass, C. (1996) The media equation. How people treat computers, television and new media like real people and places. Cambridge University Press, CSLI, Stanford. [52] Riding, R. J. and Rayner, S. (1998) Cognitive styles and learning strategies. Understanding style differences in learning and behavior. David Fulton Publishers, London. 253 253 [53] Riecken, D. (2000) Personalized views on personalization. Communications of the ACM, V. 43, 8, 27-28. [54] Rusting, C.L. (1998). Personality, mood, and cognitive processing of emotional information: Three conceptual frameworks. Psychological Bulletin, 124, 165-196. [55] Saari, T. (1998) Knowledge creation and the production of individual autonomy. How news influences subjective reality. Reports from the department of teacher education in Tampere university. A15/1998. [56] Saari, T. (2001) Mind-Based Media and Communications Technologies. How the Form of Information Influences Felt Meaning. Acta Universitatis Tamperensis 834. Tampere University Press, Tampere 2001. [57] Saari, T. (2002) Designing Mind-Based Media and Communications Technologies. Proceedings of Presence 2002 Conference, Porto, Portugal. [58] Saari, T. (2003a) Designing for Psychological Effects. Towards Mind-Based Media and Communications Technologies. In Harris, D., Duffy, V., Smith, M. and Stephanidis, C. (eds.) Human-Centred Computing: Cognitive, Social and Ergonomic Aspects. Volume 3 of the Proceedings of HCI International 2003, pp. 557-561. [59] Saari, T. (2003b) Mind-Based Media and Communications Technologies. A Framework for producing personalized psychological effects. Proceedings of Human Factors and Ergonomics 2003 -conference. 13.-17.10.2003 Denver, Colorado. [60] Saari, T. (in press, a) Using Mind-Based Technologies to facilitate Positive Emotion and Mood with Media Content. Accepted to proceedings of to Positive Emotion, 2 nd European Conference. Italy, July 2004. [61] Saari, T. (in press, b) Facilitating Learning from Online News with Mind-Based Technologies. Accepted to proceedings of EDMedia 2004, Lugano, Switzerland. [62] Saari, T. and Turpeinen, M. (in press, a) Towards Psychological Customization of Information for Individuals and Social Groups. In Karat, J., Blom, J. and Karat. M.-C. (eds.) Personalization of User Experiences for eCommerce, Kluwer. [63] Saari, T. and Turpeinen, M. (in press, b) Psychological customization of information. Applications for personalizing the form of news. Accepted to proceedings of ICA 2004, 27.31 .5. 2004, New Orleans, USA. [64] Schneider, S.L., & Laurion, S.K. (1993). Do we know what we've learned from listening to the news? Memory and Cognition, 21, 198-209. [65] Schwarz, N. (in press). Meta-cognitive experiences in consumer judgment and decision making. Journal of Consumer Psychology. [66] Stretcher, V. J., Kreutzer, M., Den Boer, D.-J., Kobrin, S., Hospers, H. J., and Skinner, C. S. (1994) the effects of computer-tailored smoking cessation messages in family practice settings. Journal of Family Practice, 39(3), 262-270. [67] Stretcher, V. J. (1999) Computer tailored smoking cessation materials: A review and discussion. Special issue: Computer tailored education. Patient Education & Counceling, 36(2), 107-117. [68] Turpeinen, M. (2000) Customizing news content for individuals and communities. Acta Polytechnica Scandinavica. Mathematics and computing series no. 103. Helsinki University of Technology, Espoo. [69] Turpeinen, M. and Saari, T. (2004) System Architechture for Psychological Customization of Information. Proceedings of HICSS-37- conference, 5.-8.1. 2004, Hawaii. [70] Vecchi, T., Phillips, L. H. & Cornoldi, C. (2001). Individual differences in visuo-spatial working memory. In: M. Denis, R. H. Logie, C. Cornoldi, M. de Vega, & J. Engelkamp (Eds.), Imagery, language, and visuo-spatial thinking. Psychology Press, Hove. [71] Zillmann, D. (1971). Excitation transfer in communication-mediated aggressive behavior. Journal of Experimental Social Psychology, 7, 419-434. [72] Zillmann, D. (1978). Attribution and misattribution of excitatory reactions. In J.H. Harvey, W.J. Ickes, & R.F. Kidd (Eds.), New directions in attribution research (Vol. 2, pp. 335-368). Hillsdale, NJ: Lawrence Erlbaum Associates. [73] Zillmann, D., & Bryant, J. (1985). Affect, mood, and emotion as determinants of selective exposure. In D. Zillmann & J. Bryant (Eds.), Selective exposure to communication (pp. 157-190). Hillsdale, NJ: Lawrence Erlbaum. 254 254

personalization emotion;persuasion;advertising;E-commerce

156

Publicly Verifiable Ownership Protection for Relational Databases

Today, watermarking techniques have been extended from the multimedia context to relational databases so as to protect the ownership of data even after the data are published or distributed. However , all existing watermarking schemes for relational databases are secret key based , thus require a secret key to be presented in proof of ownership. This means that the ownership can only be proven once to the public (e.g., to the court). After that, the secret key is known to the public and the embedded watermark can be easily destroyed by malicious users. Moreover, most of the existing techniques introduce distortions to the underlying data in the watermarking process, either by modifying least significant bits or exchanging categorical values. The distortions inevitably reduce the value of the data. In this paper, we propose a watermarking scheme by which the ownership of data can be publicly proven by anyone, as many times as necessary. The proposed scheme is distortion-free , thus suitable for watermarking any type of data without fear of error constraints. The proposed scheme is robust against typical database attacks including tuple/attribute insertion/deletion, ran-dom/selective value modification, data frame-up, and additive attacks

INTRODUCTION Ownership protection of digital products after dissemination has long been a concern due to the high value of these assets and the low cost of copying them (i.e., piracy problem). With the fast development of information technology, an increasing number of digital products are distributed through the internet. The piracy problem has become one of the most devastating threats to networking systems and electronic business. In recent years, realizing that "the law does not now provide sufficient protection to the comprehensive and commercially and publicly useful databases that are at the heart of the information economy" [12], people have joined together to fight against theft and misuse of databases published online (e.g., parametric specifications, surveys, and life sciences data) [32, 4]. To address this concern and to fight against data piracy, watermarking techniques have been introduced, first in the multimedia context and now in relational database literature, so that the ownership of the data can be asserted based on the detection of watermark . The use of watermark should not affect the usefulness of data, and it must be difficult for a pirate to invalidate watermark detection without rendering the data much less useful. Watermarking thus deters illegal copying by providing a means for establishing the original ownership of a redistributed copy [1]. In recent years, researchers have developed a variety of watermarking techniques for protecting the ownership of relational databases [1, 28, 26, 29, 13, 19, 20, 2] (see Section 5 for more on related work). One common feature of these techniques is that they are secret key based, where ownership is proven through the knowledge of a secret key that is used for both watermark insertion and detection . Another common feature is that distortions are introduced to the underlying data in the process of watermarking. Most techniques modify numerical attributes [1, 28, 29, 13, 19, 20], while others swap categorical values [26, 2]. The distortions are made such that the usability of data for certain applications is not affected and that watermark detection can be performed even in the presence of attacks such as value modification and tuple selection. The above two features may severely affect the application of watermarking techniques for relational databases. First, the secret key based approach is not suitable for proving ownership to the public (e.g., in a court). To prove ownership of suspicious data, the owner has to reveal his secret key to the public for watermark detection. After being used one time, the key is no longer secret. With access to the key, a pirate can invalidate watermark detection by either removing watermarks from protected data or adding a false watermark to non-watermarked data. Second, the distortions that are introduced in the process of watermarking may affect the usefulness of data. Even though certain kind of error constraints (e.g., means and variances of watermarked attributes) can be enforced prior to or during the watermarking process, it is difficult or even impossible to quantify all possible constraints, which may include domain constraint, unique-ness constraint, referential integrity constraint, functional dependencies , semantic integrity constraint, association, correlation, car-dinality constraint, the frequencies of attribute values, and statisti cal distributes. In addition, any change to categorical data may be considered to be significant. Another difficulty is that the distortions introduced by watermarking cannot be reduced arbitrarily. A tradeoff has to be made between watermark distortions and the robustness of watermark detection (roughly speaking, the more distortions introduced in the watermarking process, the more likely that a watermark can be detected in the presence of database attacks ). In this paper, we attempt to design a new database watermarking scheme that can be used for publicly verifiable ownership protection and that introduces no distortions. Our research was motivated in part by certain aspects of public key watermarking schemes in the multimedia context, yet it is fundamentally different and particularly customized for relational databases (see also Section 5 for related work). Our scheme has the following unique properties. First, our scheme is publicly verifiable. Watermark detection and ownership proof can be effectively performed publicly by anyone as many times as necessary. Second, our scheme introduces no errors to the underlying data (i.e., it is distortion-free); it can be used for watermarking any type of data including integer numeric, real numeric , character, and Boolean, without fear of any error constraints. Third, our scheme is efficient for incremental updating of data. It is designed to facilitate typical database operations such as tuple insertion, deletion, and value modification. Fourth, our scheme is robust. It is difficult to invalidate watermark detection and ownership proof through typical database attacks and other attacks. With these properties, we believe that our watermarking technique can be applied practically in the real world for the protection of ownership of published or distributed databases. The rest of the paper is organized as follows. Section 2 presents our watermarking scheme, which includes watermark generation and detection. Section 3 studies how to prove ownership publicly using a watermark certificate. It also investigates certificate revocation and incremental update in our scheme. Section 4 analyzes the robustness of our scheme and the tradeoff between its robustness and overhead. Section 5 comments on related work, and section 6 concludes the paper. THE SCHEME Our scheme watermarks a database relation R whose schema is R(P, A 0 , . . . , A -1 ), where P is a primary key attribute (later we discuss extensions for watermarking a relation that does not have a primary key attribute). There is no constraint on the types of attributes used for watermarking; the attributes can be integer numeric , real numeric, character, Boolean, or any other types. Attributes are represented by bit strings in computer systems. Let denote the number of tuples in relation R. For each attribute of a tuple, the most significant bit (MSB) of its standard binary representation may be used in the generation of a watermark. It is assumed that any change to an MSB would introduce intolerable error to the underlying data value. For ease of referencing, Table 1 lists the symbols that will be used in this paper. 2.1 Watermark Generation Let the owner of relation R possess a watermark key K, which will be used in both watermark generation and detection. The watermark key should be capable of publicly proving ownership as many times as necessary. This is contrast to traditional watermarking , where a watermark key is kept secret so that the database owner can prove his ownership by revealing the key for detecting the watermark . However, under that formation, the ownership can be publicly proved only once. In addition, the key should be long enough to thwart brute force guessing attacks to the key. Algorithm 1 genW (R, K, ) // Generating watermark W for DB relation R 1: for each tuple r in R do 2: construct a tuple t in W with the same primary key t.P = r.P 3: for i=0; i &lt; ; i= i+1 do 4: j = G i (K, r.P ) mod (the number of attributes in r) 5: t.W i = MSB of the j-th attribute in r 6: delete the j-th attribute from r 7: end for 8: end for 9: return W In our scheme, the watermark key is public and may take any value (numerical, binary, or categorical) selected by the owner. There is no constraint on the formation of the key. To reduce unnecessary confusion, the watermark key should be unique to the owner with respect to the watermarked relation. We suggest the watermark key be chosen as K = h(ID |DB name|version|...) (1) where ID is the owner's identity, `|' indicates concatenation, and h() is a cryptographic hash function (e.g., SHA-512) [22]. In the case of multiple owners, the public key can be extended to be a combination of all the owners' IDs or generated from them using a threshold scheme. For simplicity, we assume that there is a single owner of DB relation R in the following. Our concept of public watermark key is different from that of a public key in public key infrastructure (PKI) [16]. In the cryptography literature, a public key is paired with a private key such that a message encoded with one key can be decoded with its paired key; the key pair is selected in a specific way such that it is computation-ally infeasible to infer a private key from the corresponding public key. In our watermarking scheme, there is no private key, and the public watermark key can be arbitrarily selected. If the watermark key is derived from the owner's ID as suggested, it is similar to the public key in identity based cryptography [25, 3, 5], though the owner does not need to request a private key from a key distribution center (KDC). The watermark key is used to decide the composition of a public watermark W . The watermark W is a database relation whose scheme is W (P, W 0 , . . . , W -1 ), where W 0 , . . . , W -1 are binary attributes. Compared to DB relation R, the watermark (relation ) W has the same number of tuples and the same primary key attribute P . The number of binary attributes in W is a control parameter that determines the number of bits in W , where = and . In particular, we call the watermark generation parameter . Algorithm 1 gives the procedure genW (R, K, ) for generating the watermark W . In the algorithm, a cryptographic pseudo-random sequence generator (see chapter 16 in [24]) G is seeded with the concatenation of watermark key K and the primary key r.P for each tuple r in R, generating a sequence of numbers {G i (K, r.P ) }. The MSBs of selected values are used for generating the watermark. The whole process does not introduce any distortions to the original data. The use of MSBs is for thwarting potential attacks that modify the data. Since the watermark key K, the watermark W , and the algorithm genW are publicly known, anyone can locate those MSBs in R that are used for generating W . However , an attacker cannot modify those MSBs without introducing intolerable errors to the data. In the construction of watermark W , each tuple in relation R 0 R database relation to be watermarked number of tuples in relation R number of attributes in relation R W database watermark (relation) generated in watermarking (watermark generation parameter) number of binary attributes in watermark W number of bits in W ; = (watermark detection parameter) least fraction of watermark bits required for watermark detection K watermark key Table 1: Notation in watermarking Algorithm 2 detW (R , K, , W, ) // Detecting watermark for DB relation R' 1: match count=0 2: total coutn=0 3: for each tuple r in R do 4: get a tuple t in W with the same primary key t.P = r.P 5: for i=0; i &lt; ; i= i+1 do 6: total count = total count +1 7: j = G i (K, r.P ) mod (the number of attributes in r) 8: if t.W i = MSB of the j-th attribute in r then 9: match count = match count +1 10: end if 11: delete the j-th attribute from r 12: end for 13: end for 14: if match count/total count &gt; then 15: return true 16: else 17: return false 18: end if contributes MSBs from different attributes that are pseudo-randomly selected based on the watermark key and the primary key of the tuple . It is impossible for an attacker to remove all of the watermark bits by deleting some but not all of the tuples and/or attributes from the watermarked data. The larger the watermark generation parameter , the more robust our scheme is against such deletion attacks. 2.2 Watermark Detection Our watermark detection is designed to be performed publicly by anyone as many times as necessary. This is a notable difference compared from previous approaches, which are secret key based. In watermark detection, the public watermark key K and watermark W are needed to check a suspicious database relation R . It is assumed that the primary key attribute has not been changed or else can be recovered. If the primary key cannot be relied on, one can turn to other attributes, as will be discussed in Section 2.4. Algorithm 2 gives the procedure detW (R , K, , W, ) for detecting watermark W from relation R , where is the watermark generation parameter used in watermark generation, and is the watermark detection parameter that is the least fraction of correctly detected watermark bits. Both parameters are used to control the assurance and robustness of watermark detection, as will be ana-lyzed in Section 4. The watermark detection parameter is in the range of [0.5, 1). To increase the robustness of watermark detection , we do not require that all detected MSBs in R match the corresponding bits in W , but that the percentage of the matches is more than (i.e., match count/total count &gt; in algorithm 2). 2.3 Randomized MSBs Most modern computers can represent and process four primitive types of data besides memory addresses: integer numeric, real numeric, character, and Boolean. Regardless of its type, a data item is represented in computer systems as a bit string. The MSB of the bit string is the leftmost digit, which has the greatest weight. In a signed numeric format (integer or real), the MSB can be the sign bit, indicating whether the data item is negative or not 1 . If the sign bit is not chosen (or there is no sign bit), the MSB can be the high order bit (next to the sign bit; in floating point format, it is the leftmost bit of exponent). For character or Boolean data, any bit can be an MSB and we simply choose the leftmost one. We assume that watermark bits generated from selected MSBs are randomly distributed; that is, each MSB has the same probability of 1/2 to be 1 or 0. This randomness is important in our robustness analysis (see Section 4). If this is not the case, then we randomize the MSBs by XOR'ing them with random mask bits. For the MSB of the j-th attribute of tuple r, the corresponding mask bit is the j-th bit of hash value h(K|r.P) if j , where is the bit-length of hash output. In general, if (k - 1) &lt; j k, the mask bit is the (j - (k - 1))-th bit of hash value h k (K |r.P). Since the hash value is computed from the unique primary key, the mask bit is random; thus, the MSB after masking is random. The random-ized MSBs are then used in watermark generation and detection in our scheme. 2.4 Discussion on Relations without Primary Keys Most watermarking schemes (e.g., [1, 20, 26, 2]) for relational databases, including ours, depend critically on the primary key attribute in the watermarking process. In the case that there is no primary key attribute, or that the primary key attribute is destroyed in malicious attacks, one can turn to other attributes and construct a virtual primary key that will be used instead of the primary key in the watermarking process. The virtual primary key is constructed by combining the most significant bits of some selected attributes. The actual attributes that are used to construct the virtual primary key differ from tuple to tuple, and the selection of the attributes is based on a key that could be the watermark key in the context of this paper. The reader is referred to [19] for more details on the construction of a virtual primary key. Since the virtual primary key is constructed from the MSBs of selected attributes, it is difficult to destroy the virtual primary key through value modification or attribute deletion. However, unlike a real primary key, the virtual primary key may not be unique for each tuple; consequently, there could be multiple tuples in both R and W sharing the same value of the primary key. In watermark detection, the exact mapping between pairs of these tuples needs 1 In most commonly used storage formats, the sign bit is 1 for a negative number and 0 for a non-negative number. 0 to be recovered (see line 4 in algorithm 2). This can be done as follows. For each tuple r R with primary key r.P, compute a tuple t the same way as in watermark generation, then choose a tuple t W such that t is the most close (e.g., in terms of Hamming distance) to t among the multiple tuples in W that share the same primary key r.P . The number of tuples sharing the same primary key value (i.e., the severity of the duplicate problem) can be minimized, as shown in the above-mentioned work [19]. PUBLIC OWNERSHIP PROOF We now investigate how to publicly prove ownership as many times as necessary. If the watermark key K is kept secret with the owner, the ownership proof can be done secretly; however, it can be done only once in public since the key has to be revealed to the public during this process. The problem of public ownership proof was originally raised in the multimedia context [15] (see section 5 for details); it has not been studied in the literature of database watermarking. We note that the requirements for watermarking relational data are different from those for watermarking multimedia data. The former must be robust against typical database alterations or attacks such as tuple insertion, deletion, and value modification, while the latter should be robust against multimedia operations such as compression and transformation. An additional requirement for watermarking relational data is that a watermarked relation should be updated easily and incrementally. Public ownership proof in our scheme is achieved by combining watermark detection with a certificate. 3.1 Watermark Certificate D EFINITION 3.1. A watermark certificate C of relation R is a tuple ID,K,h(W),h(R),T,DB-CA,Sig, where ID is the identity of the owner of R, K is the owner's watermark key, W is the public watermark, T is the validity information, DB-CA is the trusted authority who signs the certificate by generating a signature Sig. Similar to the identity certificate [16] in PKI (or attribute certificate [10] in PMI), which strongly binds a public key (a set of attributes) to its owner with a validity period, the watermark certificate strongly binds a watermark key, a watermark, and a DB relation to its owner's ID with validity information. The validity information is a triple T = T origin , T start , T end indicating the original time T origin when the DB relation is first certified, the starting time T start , and the ending time T end of this certificate in the current binding. When the DB relation is certified for the first time, T origin should be the same as T start . Compared with the identity certificate or attribute certificate, the watermark certificate not only has a validity period defined by T start and T end , but also contains the original time T origin . The original time will be useful in thwarting possible attacks that confuse ownership proof. A comparison of the watermark certificate with the traditional identity certificate is illustrated in Figure 1. The two kinds of certificates share a similar structure except that the public key information in the identity certificate is replaced by the watermark key, watermark hash, and database hash in the watermark certificate. In traditional identity certificate, the subject's public key is paired with a private key known only to the subject. In the case of damage or loss of the private key (e.g., due to collision attacks), the identity certificate needs to be revoked before the expiration of the certificate . In the watermark certificate, since there is no private key asso-ciated with the public watermark key, it seems that there is no need Version Serial Number Signature Algorithm Issuer Validity Period Subject Subject Public Key Info Signature Version Serial Number Signature Algorithm DB-CA Validity Info T DB owner ID Watermark Key K Watermark Hash h(W) DB hash h(R) Signature Sig Identity Certificate Watermark Certificate Figure 1: Relation between watermark and identity certificate of certificate revocation. Nonetheless, certificate revocation and recertification may be needed in the case of identity change, ownership change, DB-CA signature compromise, and database update. The role of DB-CA is similar to that of the traditional CA in PKI in terms of authentication of an applicant's identity. The differences are: (i) it binds the applicant's ID to the watermark key, watermark, and watermarked data; and (ii) it confirms the original time when the watermarked data was first certified. The original time is especially useful in the case of recertification so as to thwart false claims of ownership by a pirate. This is addressed in the following subsection. 3.2 Public Verifiability While the watermark detection process can be performed by anyone , voluntarily or in delegation, who has access to the public watermark and watermark key, the ownership is proven by further checking the corresponding watermark certificate. This involves checking (i) if the watermark certificate has been revoked (see the next subsection for details); (ii) if the watermark key and (the hash of) the watermark used in watermark detection are the same as those listed in the watermark certificate; (iii) if the signature is correctly signed by the DB-CA stipulated in the watermark certificate (this is done in traditional PKI and may involve checking the DB-CA's public key certificate, a chain of CA's certificates, and a certificate revocation list); and (iv) the similarity of suspicious data R to the original data R as published by the owner of watermark certificate. If all are proven, the ownership of the suspicious data is publicly claimed to belong to the owner of the watermark certificate for the time period stipulated in the certificate. The original time that the data was certified is also indicated in the certificate. The last requirement is optional, depending on whether data frame-up attack is of concern. In a data frame-up attack, an attacker modifies the watermarked data as much as possible while leaving the watermarked bits (i.e., MSBs of selected values) untouched . Note that in our scheme, an attacker can pinpoint the watermarked bits since the watermark key, watermark, and watermark algorithm are all public. Since the ownership is publicly verifiable, such "frame-up" data may cause confusion and damage to the legitimate ownership. The data frame-up attack has not been discussed before, even though it is also possible in secret key based schemes. For example , in Agrawal and Kiernan's watermarking scheme [1], the watermark information is embedded in one of least significant bits of some selected values. Data frame-up attack is possible if an attacker modifies all significant bits except the last least significant bits in each value. However, this attack is less serious in secret key based schemes because the owner of watermarked data may choose not to claim the ownership for "frame-up" data. In our scheme, this attack is thwarted by requiring that the suspicious data is similar enough to the original data (the authenticity of the original data R can be checked with h(R) in the watermark certificate). The rationale is that when an attacker forges a low quality data R with the MSBs given in the public watermark W , such R will be significantly different from the original R due to its low quality. The similarity between R and R may be measured, for example, by the portion of significant bits that match for each pair of values in R and R whose watermarked MSBs match. The similarity may also be measured in terms of the usefulness of data, such as the difference of individual values, means, and variances. 3.3 Certificate Management Once publicly proven based on a valid watermark certificate, the ownership of watermarked data is established for the owner of the certificate. The current ownership is valid for a time period [T start , T end ] stipulated in the certificate. The original time T origin when the data was first certified is also indicated in the certificate . The use of original time is to thwart additive attack. Additive attack is a common type of attacks to watermarking schemes in which an attacker simply generates another watermark for watermarked data so as to confuse ownership proof. The additional watermark can be generated using a watermark key that is derived from the attacker's ID. It is also possible for the attacker to obtain a valid watermark certificate for this additional watermark. We solve this problem by comparing the original time T origin in the certificate of real owner with the original time T origin in the certificate of the attacker. We assume that the owner of data will not make the data available to potential attackers unless the data is watermarked and a valid watermark certificate is obtained. Therefore , one always has T origin &lt; T origin by which the legitimate ownership can be proven in the case of an ownership dispute. After this, the attacker's valid certificate should be officially revoked. Besides revocation upon losing an ownership dispute, a certificate may be revoked before its expiration based on the following reasons: (1) identity change; (2) ownership change; (3) validity period change; (4) DB-CA compromise; and (5) database update. When the owner of a valid certificate changes his identity, he needs to revoke the certificate and, at the same time, apply for a new certificate to replace the old one. Upon the owner's request, the DB-CA will grant a new validity period [T start , T end ] according to its policy while keeping the original time T origin unchanged in the new certificate. The case of ownership change is handled in a similar manner, except that the DB-CA needs to authenticate the new owner and ensure the ownership change is granted by the old owner. In both cases, a new watermark key and a new watermark may be derived and included in the new certificate. Sometimes the owner wants to prolong or shorten the validity period of his certificate. In this case, the watermark certificate needs to be re-certified with a new validity period. The watermark key or watermark does not need to change in the recertification process. In our scheme, the DB-CA is trusted, similar to the CA in traditional PKI. A traditional PKI certificate would need to be revoked for a variety of reasons, including key compromise and CA compromise . Since a watermark key is not paired with a private key in our scheme, there is no scenario of watermark key compromise. However, there is a possibility of DB-CA compromise if any of the following happens: (i) DB-CA's signature is no longer safe (e.g., due to advanced collision attacks); (ii) DB-CA loses its signature key; (iii) DB-CA ceases its operation or business; or (iv) any CA who certifies the DB-CA's public key is compromised (the public key is used to verify the DB-CA's signature in our scheme). In the case of DB-CA compromise, all related watermark certificates must be revoked and re-examined by a valid DB-CA and recertified with new validity periods but unchanged original times. Due to the similarity between the watermark certificate and the traditional identity certificate, many existing standards and mechanisms regarding certificate management, such as certification path constraints and CRL distribution points, can be borrowed from PKI with appropriate adaptations. For simplicity and convenience, the functionality of a DB-CA may be performed by a CA in traditional PKI. 3.4 Efficient Revocation of Watermark Certificate Micali proposed an efficient public key certificate revocation scheme [23] called CRS (for certificate revocation status). Compared with the CRL-based solution, CRS substantially reduces the cost of management of certificates in traditional PKI. This scheme can easily be adapted to our scheme for efficient revocation of watermark certificates . As pointed out in [23], the costs of running a PKI are staggering and most of the costs are due to CRL transmission. The major reason is that each time a user queries the status of a single certificate , he needs to query a directory, an agent receiving certificate information from a CA and handling user queries about it, and the directory sends him the whole CRL list that has been most recently signed by the CA. Since the CRL list tends to be very long and transmitted very often, the CRL solution is extremely expensive. In CRS, however, the directory responds to a user's query by sending a 100-bit value only, instead of the whole CRL. The 100-bit value is employed by the user to verify whether the relative certificate is valid or has been revoked. In our watermarking scheme, the DB-CA selects a secret 100-bit value Y 0 for a watermark certificate, and recursively applies on it a one-way function F 365 times, assuming that the validity period of the certificate is a normal year. The DB-CA then includes the 100-bit value Y 365 = F 365 (Y 0 ) in the watermark certificate C = ID,K,h(W),h(R),T,DB-CA,Y 365 , Sig . Assume that the current day is the i-th day in the validity period of the certificate. The DB-CA generates a 100-bit value Y 365 -i = F 365 -i (Y 0 ) and gets it published through the directory. It is the DB owner's responsibility to obtain Y 365 -i from the directory and publish it together with the watermark certificate C. Anyone can verify the validity of the certificate by checking whether F i (Y 365 -i ) = Y 365 , where i is the number of days since the start of the validity period (i.e., T start in T ). If this is the case, the certificate is valid; otherwise, it has been revoked before the i-th day, in which case the DB-CA did not get Y 365 -i published. Note that Y 365 -i cannot be computed from previously released Y 365 -j (j &lt; i) due to the one-way property of function F . In this scheme, the DB owner needs to query the directory and update Y 365 -i every day. To make the transition from Y 365 -i to Y 364 -i smooth, one more hour may be granted for the validity period of Y 365 -i (i.e., 25 hours). To avoid high query load at certain hours, the validity period of Y 365 -i should start at a different time each day for a different certificate. A policy stating this may also be included in the watermark certificate. Note that Micali's original scheme requires a CA to (i) sign another 100-bit value besides Y 365 -i to explicitly indicate a certificate being revoked; and (ii) sign a updated list indicating all and only the series numbers of issued and not-yet-expired certificates. The signed value and list are sent to the directory so that any user query can be answered by the directory. In our scheme, it is the DB owner's responsibility (for his own benefit, namely anti-piracy) to query the directory and publish the updated Y 365 -i online together with DB, watermark, and certificate. A user who wants to verify the certificate will obtain the validity information from the owner rather than from the directory. This separation of duty simplifies the scheme and clarifies the responsibility of the DB owner. It is relatively straightforward to analyze the communication cost of our scheme as compared with the CRL based solution. The analysis is very similar to that given in [23] for comparing CRS with CRL (CRS is about 900 times cheaper than CRL in terms of communication cost with the Federal PKI estimates). We omit this analysis due to space limitations. 3.5 Incremental Updatability The proposed scheme is also designed to facilitate incremental database update. In relational database systems, database update has been tailored to tuple operations (tuple insertion, deletion, and modification), where each tuple is uniquely identified by its primary key. In our scheme, both watermark generation and detection are tuple oriented; each tuple is processed independently of other tuples, based on its primary key. The watermark is updated as follows. If a set of new tuples is inserted into the watermarked data, the watermark generation algorithm 1 can be performed on those new tuples only. As a result, a set of corresponding new tuples is generated and inserted into the watermark. If a set of tuples is deleted from the watermarked data, the corresponding tuples with the same primary keys are simply deleted from the watermark. In the case that a set of values is modified , only the related tuples need to be updated in the watermark. This can be done in a similar manner as in the tuple insertion case. Note that if a modified value does not contribute any MSB to the watermark, then no update is needed for that value. The update of the watermark certificate follows the update of the watermark. To update a watermark certificate, the owner of watermarked data needs to authenticate himself to a DB-CA, revoke the old certificate, and get a new certificate for the updated DB and watermark . The new certificate may have an updated validity period, but the original time will not be altered. As this process involves interactions with a DB-CA, it may not be efficient if executed frequently . Fortunately, our scheme is very robust against database update, as will be indicated in the next section. Therefore, the update of the watermark and watermark certificate may lag behind the update of the watermarked data; it can be done periodically after a batch of data updates. The lag-behind watermark and certificate can still be used for checking the ownership of the updated data as long as the updates do not severely degrade the robustness of our scheme. 3.6 Discussion Like traditional PKI, the certificate revocation in our scheme is handled only by the trusted party (i.e., the DB-CA). An alternative solution is to let the DB owner himself handle the certificate revocation. After the DB-CA signs a watermark certificate C = ID,K,h(W),h(R),T,DB - CA,Y 365 , Sig , where Y 365 = F 365 (Y 0 ), it gives Y 0 to the DB owner through a secure channel. The DB owner keeps Y 0 secret. On the i-th day in the validity period of the certificate, the DB owner himself can generate and publish Y 365 -i = F 365 -i (Y 0 ), based on which anyone can verify the validity of the certificate. This solution further simplifies our scheme in the sense that the DB-CA does not need to generate Y-values for all valid certificates each day, and that all DB owners do not need to query a directory to update the Y-values. The communication cost is thus reduced substantially. Whenever the DB owner deems it appropriate (e.g., after database is updated), he can refuse to release new Y-values to the public, thus revoking the certificate in a de facto manner, and apply a new certificate if necessary. This solution works well for database updates because it is to the benefit of the DB owner to maintain the certificate status. However, it may not work well in the case of DB-CA compromise or loss of Y 0 , but this fortunately would not happen very often as compared with database updates. It is possible to develop a hybrid solution that combines the merits of both DB-owner-handled revocation and CA-handled revocation. ROBUSTNESS AND OVERHEAD For a watermarking scheme to be useful, it must be robust against typical attacks and be efficient in practice. In this section, we first present a quantitative model for the robustness of our watermarking scheme. We analyze the robustness of our scheme by the same method (i.e., binomial probability) as was used in [1]. We then investigate the overhead of our watermarking scheme. We also study the tradeoffs between the robustness and overhead in terms of the watermarking generation parameter and watermarking detection parameter . 4.1 Survival Binomial Probability Consider n Bernoulli trials of an event, with probability p of success and q = 1 - p of failure in any trial. Let P p (k; n) be the probability of obtaining exactly k successes out of n Bernoulli trials (i.e., the discrete probability of binomial distribution). Then P p (k; n) = nkp k q n -k (2) nk = n! k!(n - k)! (3) Let C p (k; n) denote the probability of having more than k successes in n Bernoulli trials; that is, C p (k; n) is the survival binomial probability. According to the standard analysis of binomial distribution, we have C p (k; n) = n i=k+1 P p (i; n) (4) In many widely available computation software packages such as Matlab and Mathematica, the survival binomial probability can be computed by C p (k; n) = 1 - binocdf(k,n,p), where binocdf(k, n, p) is the binomial cumulative distribution function with parameters n and p at value k. When n is large, the binomial distribution can be approximated by a normal distribution with mean np, standard deviation npq, at value k + 0.5, where 0.5 is the correction of continuity (for p = 0.5, the normal is a good approximation when n is as low as 10; see chapter 7.6 in [31]). Thus, C p (k; n) = 1 - normcdf(k + 0.5,np,npq), where normcdf is the normal cumulative distribution function. 4.2 Detecting Non-Watermarked Data First consider the robustness of our scheme in terms of false hit, which is the probability of a valid watermark being detected from non-watermarked data. The lower the false hit, the better the robustness . We show that the false hit is under control in our scheme and can be made highly improbable. Recall that in watermark detection, a collection of MSBs are located in suspicious data and compared with the corresponding bits recorded in the public watermark. When the watermark detection is applied to non-watermarked data, each MSB in data has the same probability 1/2 to match or not to match the corresponding bit in the watermark. Assume that the non-watermarked data has the same number of tuples (and the same primary keys) as the original data. Let = be the total number of bits in the watermark , where is the watermark generation parameter. The false hit is the probability that at least portion of bits can be detected from the non-watermarked data by sheer chance, where is the watermark detection parameter. The false hit H can be written as H = C 1/2 ( ,) = C 1/2 ( ,) (5) 1 2 3 4 5 6 7 8 9 10 10 -15 10 -10 10 -5 10 0 False hit H =1000 =0.51 =0.52 =0.53 =0.54 =0.55 Figure 2: False hit as function of 2000 4000 6000 8000 10000 10 -15 10 -10 10 -5 10 0 False hit H =5 =0.51 =0.52 =0.53 =0.54 =0.55 Figure 3: False hit as function of Figure 2 shows the change of the false hit when the watermark insertion parameter increases from 1 to 10 for fixed = 1000 and various values of the watermark detection parameter . The figure illustrates that the false hit is monotonic decreasing with both watermark insertion parameter and detection parameter . On the one hand, the larger the insertion parameter , the more MSBs are included in the watermark and the smaller the false hit. On the other hand, the false hit can be decreased by increasing the detection parameter , which is the least fraction of watermark bits required for ownership assertion. Figure 3 illustrates the trend of false hit when the number of tuples is scaled up from 1000 to 10,000. The trend is that the false hit is monotonic decreasing with . This trend is linear, which is similar to that of increasing , as indicated in figure 2. A conclusion drawn from these two figures is that with reasonably large values of , , and/or , the false hit can be made extremely low. 4.3 Detecting Watermarked Data We now consider the robustness of our scheme in terms of false miss , which is the probability of not detecting a valid watermark from watermarked data that has been modified in typical attacks. The robustness can also be measured in terms of the error introduced by typical attacks. The less the false miss, or the larger the error introduced by typical attacks, the better the robustness. The typical attacks include database update, selective value modification , and suppression. Other typical attacks include the data frame-up attack and the additive attack which have been addressed in a previous section. 4.3.1 Typical Database Update Typical database update includes tuple insertion, tuple deletion, attribute deletion, and value modification. For tuple deletion and attribute deletion, the MSBs in the deleted tuples or attributes will not be detected in watermark detection; however, the MSBs in other tuples or attributes will not be affected. Therefore, all detected MSBs will match their counterparts in the public watermark, and the false miss is zero. Though the deletion of tuples or attributes will not affect the false miss , it will make the false hit worse. The more the tuples or attributes are deleted, the larger the false hit, as indicated in Section 4.2. The effect to the false hit of deleting tuples is equivalent to that of decreasing as shown in Figure 3, while the effect of deleting attributes is equivalent to decreasing proportionally as shown in Figure 2. Since the watermark detection is primary key based, a newly inserted tuple should have a valid primary key value; otherwise, there is no corresponding tuple in the public watermark. We thus consider tuple insertion to be "mix-and-match" [1]; that is, an attacker inserts new tuples to replace watermarked tuples with their primary key values unchanged. For watermark detection to return a false answer, at least - MSBs in those newly added tuples (which consists of MSBs) must not match their counterparts in the public watermark (which consist of bits). Therefore, the false miss M for inserting tuples in mix-and-match can be written as M = C 1/2 ( - - 1,) (6) 80 82 84 86 88 90 92 94 96 98 100 10 -15 10 -10 10 -5 10 0 / (%) False miss M =5, =1000 =0.51 =0.52 =0.53 =0.54 =0.55 Figure 4: False miss (tuple insertion) as function of Figures 4, 5, and 6 show the false miss in the case of tuple insertion . The default parameters in these figures are / = 90% (i.e., bits recorded in the public watermark. When the watermark detection is applied to non-watermarked data, each MSB in data has the same probability 1/2 to match or not to match the corresponding bit in the watermark. Assume that the non-watermarked data has the same number of tuples (and the same primary keys) as the original data. Let = be the total number of bits in the watermark , where is the watermark generation parameter. The false hit is the probability that at least portion of bits can be detected from the non-watermarked data by sheer chance, where is the watermark detection parameter. The false hit H can be written as H = C 1/2 ( ,) = C 1/2 ( ,) (5) 1 2 3 4 5 6 7 8 9 10 10 -15 10 -10 10 -5 10 0 False hit H =1000 =0.51 =0.52 =0.53 =0.54 =0.55 Figure 2: False hit as function of 2000 4000 6000 8000 10000 10 -15 10 -10 10 -5 10 0 False hit H =5 =0.51 =0.52 =0.53 =0.54 =0.55 Figure 3: False hit as function of Figure 2 shows the change of the false hit when the watermark insertion parameter increases from 1 to 10 for fixed = 1000 and various values of the watermark detection parameter . The figure illustrates that the false hit is monotonic decreasing with both watermark insertion parameter and detection parameter . On the one hand, the larger the insertion parameter , the more MSBs are included in the watermark and the smaller the false hit. On the other hand, the false hit can be decreased by increasing the detection parameter , which is the least fraction of watermark bits required for ownership assertion. Figure 3 illustrates the trend of false hit when the number of tuples is scaled up from 1000 to 10,000. The trend is that the false hit is monotonic decreasing with . This trend is linear, which is similar to that of increasing , as indicated in figure 2. A conclusion drawn from these two figures is that with reasonably large values of , , and/or , the false hit can be made extremely low. 4.3 Detecting Watermarked Data We now consider the robustness of our scheme in terms of false miss , which is the probability of not detecting a valid watermark from watermarked data that has been modified in typical attacks. The robustness can also be measured in terms of the error introduced by typical attacks. The less the false miss, or the larger the error introduced by typical attacks, the better the robustness. The typical attacks include database update, selective value modification , and suppression. Other typical attacks include the data frame-up attack and the additive attack which have been addressed in a previous section. 4.3.1 Typical Database Update Typical database update includes tuple insertion, tuple deletion, attribute deletion, and value modification. For tuple deletion and attribute deletion, the MSBs in the deleted tuples or attributes will not be detected in watermark detection; however, the MSBs in other tuples or attributes will not be affected. Therefore, all detected MSBs will match their counterparts in the public watermark, and the false miss is zero. Though the deletion of tuples or attributes will not affect the false miss , it will make the false hit worse. The more the tuples or attributes are deleted, the larger the false hit, as indicated in Section 4.2. The effect to the false hit of deleting tuples is equivalent to that of decreasing as shown in Figure 3, while the effect of deleting attributes is equivalent to decreasing proportionally as shown in Figure 2. Since the watermark detection is primary key based, a newly inserted tuple should have a valid primary key value; otherwise, there is no corresponding tuple in the public watermark. We thus consider tuple insertion to be "mix-and-match" [1]; that is, an attacker inserts new tuples to replace watermarked tuples with their primary key values unchanged. For watermark detection to return a false answer, at least - MSBs in those newly added tuples (which consists of MSBs) must not match their counterparts in the public watermark (which consist of bits). Therefore, the false miss M for inserting tuples in mix-and-match can be written as M = C 1/2 ( - - 1,) (6) 80 82 84 86 88 90 92 94 96 98 100 10 -15 10 -10 10 -5 10 0 / (%) False miss M =5, =1000 =0.51 =0.52 =0.53 =0.54 =0.55 Figure 4: False miss (tuple insertion) as function of Figures 4, 5, and 6 show the false miss in the case of tuple insertion . The default parameters in these figures are / = 90% (i.e., 1 2 3 4 5 6 7 8 9 10 10 -15 10 -10 10 -5 10 0 False miss M =1000, / =90% =0.51 =0.52 =0.53 =0.54 =0.55 Figure 5: False miss (tuple insertion) as function of 2000 4000 6000 8000 10000 10 -15 10 -10 10 -5 10 0 False miss M =5, / =90% =0.51 =0.52 =0.53 =0.54 =0.55 Figure 6: False miss (tuple insertion) as function of 90% of the new tuples are inserted into the data to replace the watermarked tuples), = 5, and = 1000. A general trend shown in these figures is that the false miss is monotonic increasing with watermark detection parameter . This trend is opposite to that of the false hit, which is monotonic decreasing with as indicated in Figures 2 and 3. Therefore, there is a tradeoff between false hit and false miss with respect to . Figure 4 shows that even if 80% of watermarked tuples are replaced with new tuples, the false miss is as low as 10 -15 for all values greater than or equal to 51%. The false miss is close to one only if more than 90% of watermarked tuples are replaced in this figure. Figures 5 and 6 illustrate that the false miss is monotonic decreasing with and , which is similar to the trend of false hit as indicated in Figures 2 and 3. With reasonably large and/or , the false miss can be made extremely low. For value modification, we assume that the modified values are randomly chosen. We leave the selective modification targeted on watermarked values to the next subsection. Recall that there are attributes in the original data in which attributes are watermarked for each tuple. When a random modification happens, it has probability / that a watermarked value is chosen. When a watermarked value is modified, its MSB has probability 1/2 to change (i.e., the value is modified randomly). In watermark detection, a detected MSB has probability /(2) not to match its counterpart in the public watermark. The false miss M for randomly modifying values can be written as M = C /2 ( - - 1,) (7) 80 82 84 86 88 90 92 94 96 98 100 10 -15 10 -10 10 -5 10 0 /() (%) False miss M =10, =5, =1000 =0.51 =0.52 =0.53 =0.54 =0.55 Figure 7: False miss (value modification) as function of 1 2 3 4 5 6 7 8 9 10 10 -15 10 -10 10 -5 10 0 False miss M =10, =1000, /() =90% =0.51 =0.52 =0.53 =0.54 =0.55 Figure 8: False miss (value modification) as function of 2000 4000 6000 8000 10000 10 -15 10 -10 10 -5 10 0 False miss M =10, =5, /() =90% =0.51 =0.52 =0.53 =0.54 =0.55 Figure 9: False miss (value modification) as function of Figures 7, 8, and 9 show the false miss in the case of random value modification. The default parameters in these figures are /() = 90% (i.e., 90% of the values are modified randomly), = 10, = 5, and = 1000. The general trend shown in these figures for value modification is similar to that shown in previous Figures 4, 5, and 6 for tuple insertion. The difference in calculation is due to the use of probability /2 in Equation 7 instead of probability 1/2 in Equation 6. Figure 7 shows that even if 80% of values are modified randomly, which would make the data less useful, the false miss rate in detection is less than 10 -10 in our computation. 4.3.2 Selective Value Modification and Suppression Since both the watermark key and the watermark are public in our scheme, an attacker can pinpoint the MSBs of watermarked values. A simple attack would be to flip some of those MSBs so that the watermark detection will detect no match. Assuming that watermarked MSBs are flipped in selective value modification, the false miss M can be written as M = 1 if-0 otherwise (8) If no less than - watermarked MSBs are flipped, the watermarked data will no longer be detected. The robustness of our scheme can then be measured in terms of the error introduced by this attack. The larger the error introduced for defeating the watermark detection (i.e., achieving M = 1), the better the robustness. Recall that any change to an MSB would introduce intolerable error to the related data value. To defeat the watermark detection, no less than - MSBs have to be flipped; this would introduce intolerable errors to no less than - data values. We thus measure the robustness in terms of failure error rate, which is the least fraction F of total data values that need to be intolerably modified for defeating the watermark detection. This failure error rate can be written as F = (1 - ) (9) A larger failure error rate (or better robustness) can be achieved by increasing (watermark generation parameter) or decreasing (watermark detection parameter). There is a tradeoff between the robustness of our scheme and the size of the public watermark (which has binary attributes). To achieve the best robustness in terms of thwarting the selective modification attacks, one may choose = and 0.5. (However, this would increase the false hit as indicated in Section 4.2.) In this extreme case, approxi-mately 50% of data values have to be intolerably modified so as to defeat the watermark detection. To avoid the intolerable error, an attacker may choose to suppress some watermarked values rather than flipping their MSBs. Since this attack causes no mismatch in watermark detection, the false miss is zero. However, it will increase the false hit because those MSBs will be missed in watermark detection. It is easy to know that the effect of suppressing MSBs to the false hit is the equivalent of decreasing the total number of MSBs by in the computation of false hit. Thus, the false hit formula (see section 4.2) changes from C 1/2 ( ,) to C 1/2 ( ( - ), - ) for selective suppression of watermarked values. Figure 10 shows the influence of selective value suppression to the false hit for fixed = 5, = 1000, and various from 0.51 to 0.55. In the figure, we change the rate /() (the percentage of watermarked bits are suppressed) from 0% to 99%. Even if the rate /() increases up to 50%, the false hit is still below 15.4% for = 0.51, below 2.2% for = 0.52, below 0.13% for = 0.53, below 3 10 -5 for = 0.54, and below 2.6 10 -7 for = 0.55. 4.4 Overhead We now analyze the time and space overhead for both watermark 0 10 20 30 40 50 60 70 80 90 100 10 -15 10 -10 10 -5 10 0 /() (%) False hit H =5, =1000 =0.51 =0.52 =0.53 =0.54 =0.55 Figure 10: False hit (value suppression) as function of generation and watermark detection. Throughout the analysis, we ignore the IO cost (i.e., reading and writing tuples). Table 2 describes the symbols that will be used in this section. Consider watermark generation. For each of tuples to be processed , a random sequence generator G is first seeded, then MSBs are determined based on random numbers generated by G. The MSBs are assigned to the corresponding attributes in the public watermark. For each MSB to be determined, one mod operation is involved and one attribute is deleted from the copy of related tuple. The memory requirement for the process of a tuple is to keep the copy of the tuple, MSBs, and the watermark key in concatenation with the tuple's primary key. Therefore, the time overhead t genW and space overhead m genW for watermark generation are t genW = t seed + (t genS + t mod + t bit + t delA ) = O() (10) m genW = m tuple + + m wkey = O() (11) In watermark detection, the time and space overheads are the same as in watermark generation except for the cost of processing the count information. Let t if denote the cost of the last operation "if match count/total count &gt; ." The time overhead t detW and space overhead m detW for watermark detection can be written as t detW = 2t count + t seed + (t genS + t mod + t bit + t delA + 2t count ) + t if = O() (12) m detW = 2m count + m tuple + + m wkey = O() (13) The generated watermark W will be stored on disk. The disk storage requirement m disk is thus m disk = |W| = m pkey + = O() (14) 4.5 Tradeoffs In our watermark scheme, we have two parameters: watermark generation parameter and watermark detection parameter . The two parameters can be used to balance between the robustness and the overhead of our scheme. Table 3 summarizes the tradeoffs that can be made when choosing the two parameters. The watermark generation parameter is used to balance between robustness and overhead. The larger the , the better the robustness of our scheme and the worse the time and space overhead . While the watermark detection parameter has no effect on figures for value modification is similar to that shown in previous Figures 4, 5, and 6 for tuple insertion. The difference in calculation is due to the use of probability /2 in Equation 7 instead of probability 1/2 in Equation 6. Figure 7 shows that even if 80% of values are modified randomly, which would make the data less useful, the false miss rate in detection is less than 10 -10 in our computation. 4.3.2 Selective Value Modification and Suppression Since both the watermark key and the watermark are public in our scheme, an attacker can pinpoint the MSBs of watermarked values. A simple attack would be to flip some of those MSBs so that the watermark detection will detect no match. Assuming that watermarked MSBs are flipped in selective value modification, the false miss M can be written as M = 1 if-0 otherwise (8) If no less than - watermarked MSBs are flipped, the watermarked data will no longer be detected. The robustness of our scheme can then be measured in terms of the error introduced by this attack. The larger the error introduced for defeating the watermark detection (i.e., achieving M = 1), the better the robustness. Recall that any change to an MSB would introduce intolerable error to the related data value. To defeat the watermark detection, no less than - MSBs have to be flipped; this would introduce intolerable errors to no less than - data values. We thus measure the robustness in terms of failure error rate, which is the least fraction F of total data values that need to be intolerably modified for defeating the watermark detection. This failure error rate can be written as F = (1 - ) (9) A larger failure error rate (or better robustness) can be achieved by increasing (watermark generation parameter) or decreasing (watermark detection parameter). There is a tradeoff between the robustness of our scheme and the size of the public watermark (which has binary attributes). To achieve the best robustness in terms of thwarting the selective modification attacks, one may choose = and 0.5. (However, this would increase the false hit as indicated in Section 4.2.) In this extreme case, approxi-mately 50% of data values have to be intolerably modified so as to defeat the watermark detection. To avoid the intolerable error, an attacker may choose to suppress some watermarked values rather than flipping their MSBs. Since this attack causes no mismatch in watermark detection, the false miss is zero. However, it will increase the false hit because those MSBs will be missed in watermark detection. It is easy to know that the effect of suppressing MSBs to the false hit is the equivalent of decreasing the total number of MSBs by in the computation of false hit. Thus, the false hit formula (see section 4.2) changes from C 1/2 ( ,) to C 1/2 ( ( - ), - ) for selective suppression of watermarked values. Figure 10 shows the influence of selective value suppression to the false hit for fixed = 5, = 1000, and various from 0.51 to 0.55. In the figure, we change the rate /() (the percentage of watermarked bits are suppressed) from 0% to 99%. Even if the rate /() increases up to 50%, the false hit is still below 15.4% for = 0.51, below 2.2% for = 0.52, below 0.13% for = 0.53, below 3 10 -5 for = 0.54, and below 2.6 10 -7 for = 0.55. 4.4 Overhead We now analyze the time and space overhead for both watermark 0 10 20 30 40 50 60 70 80 90 100 10 -15 10 -10 10 -5 10 0 /() (%) False hit H =5, =1000 =0.51 =0.52 =0.53 =0.54 =0.55 Figure 10: False hit (value suppression) as function of generation and watermark detection. Throughout the analysis, we ignore the IO cost (i.e., reading and writing tuples). Table 2 describes the symbols that will be used in this section. Consider watermark generation. For each of tuples to be processed , a random sequence generator G is first seeded, then MSBs are determined based on random numbers generated by G. The MSBs are assigned to the corresponding attributes in the public watermark. For each MSB to be determined, one mod operation is involved and one attribute is deleted from the copy of related tuple. The memory requirement for the process of a tuple is to keep the copy of the tuple, MSBs, and the watermark key in concatenation with the tuple's primary key. Therefore, the time overhead t genW and space overhead m genW for watermark generation are t genW = t seed + (t genS + t mod + t bit + t delA ) = O() (10) m genW = m tuple + + m wkey = O() (11) In watermark detection, the time and space overheads are the same as in watermark generation except for the cost of processing the count information. Let t if denote the cost of the last operation "if match count/total count &gt; ." The time overhead t detW and space overhead m detW for watermark detection can be written as t detW = 2t count + t seed + (t genS + t mod + t bit + t delA + 2t count ) + t if = O() (12) m detW = 2m count + m tuple + + m wkey = O() (13) The generated watermark W will be stored on disk. The disk storage requirement m disk is thus m disk = |W| = m pkey + = O() (14) 4.5 Tradeoffs In our watermark scheme, we have two parameters: watermark generation parameter and watermark detection parameter . The two parameters can be used to balance between the robustness and the overhead of our scheme. Table 3 summarizes the tradeoffs that can be made when choosing the two parameters. The watermark generation parameter is used to balance between robustness and overhead. The larger the , the better the robustness of our scheme and the worse the time and space overhead . While the watermark detection parameter has no effect on Table 2: Symbols used in the analysis of overhead t seed cost of seeding random sequence generator S with public key and a tuple's primary key t genS cost of generating a random number from S t mod cost of mod operation t delA cost of deleting an attribute from a copy of a tuple t bit cost of assigning/comparing a bit value to/with the public watermark t count cost of assigning/updating a count in watermark detection m count number of bits required to store a count in watermark detection m tuple number of bits required to store a copy of a tuple m wkey number of bits to store a watermark key m pkey number of bits to store a primary key value Table 3: Tradeoffs para-false false failure robustness overhead overhead meter hit miss error rate (summary) (time) (space) H M F H M F in terms of H -in terms of M,F the overhead, it is used as a tradeoff between false hit, false miss, and failure error rate. Increasing will make the robustness better in terms of false hit, but worse in terms of false miss and failure error rate. RELATED WORK Watermarking has been extensively studied in the context of multimedia data for the purpose of ownership protection and authentication [7, 17, 18]. Most watermarking schemes proposed so far are secret key based , which require complete disclosure of the watermarking key in watermark verification. These watermarking schemes can be further classified as private (both the secret key and original data are required in watermark verification), blind (only the secret key is needed for watermark bit decoding), and semi-blind (it requires both the secret key and watermark bit sequence in watermark detection). Watermarking schemes can also be classified as being robust (the watermark is hardly destroyed in attacks), or fragile (the watermark is hardly untouched if the watermarked data is modified). The robust watermark may be used for ownership proof while the fragile watermark is suitable for data authentication and integrity check. As database piracy increasingly becomes a serious problem, watermarking techniques have been extended to protect the ownership of published or distributed databases [1, 13, 28, 29, 26, 19, 20, 2]. Agrawal and Kiernan [1] first proposed a robust watermarking scheme for database relations. Their scheme modifies a collection of least significant bits of numerical attributes. The locations of those least significant bits, and the values to which those bits are modified, are all determined by a secret key. With the same secret key, those modified values can be localized in watermark detection, and ownership is claimed if a large portion of the detected values are as expected. As noted by Agrawal and Kiernan [1], database relations differ from multimedia data in significant ways and hence require a different class of watermarking techniques. A major difference is that a database relation is composed of a set of tuples; each tuple represents an independent object which can be added, deleted, and modified frequently in either benign updates or malicious attacks. In contrast, a multimedia object consists of a large number of bits; portions of a multimedia object are bound together in fixed spatial or temporal order that cannot be arbitrarily changed. It is also noted that the frequency domain watermarking being used in the multimedia context is not suitable for watermarking relational data. The reason is that the error introduced in frequency domain will spread over all attribute values (i.e., the whole "image"), which may not be acceptable in certain database applications. There have been other schemes proposed for watermarking relational data. In Sion et al.'s scheme [28], an arbitrary bit is embedded into a selected subset of numeric values by changing the distribution of the values. The selection of the values is based on a secret sorting. In another work, Gross-Amblard [13] designs a query-preserving scheme which guarantees that special queries (called local queries) can be answered up to an acceptable distortion. Recent work also includes watermarking categorical data [26], streaming data [29], XML data [27], and medical databases [2]. The watermarking schemes for categorical data [26, 2] exchange pairs of categorical values so as to embed watermark information. In this case, there is no insignificant change and the error constraint is considered at aggregation level (e.g., k-anonymity). A common feature of this class of work is that a watermark is embedded and detected based on a secret key. Without knowing the key, an attacker is not able to locate exactly where the watermark is embedded, nor does he destroy the embedded watermark unless too many errors are introduced. A drawback of such a solution is that the ownership of watermarked data can be proven only once. After the key is revealed to the public (e.g., to the court) in the proof, anyone knowing the key can easily locate and remove the embedded watermark. Another common feature of these schemes is that the watermarking process introduces errors to the underlying data. This may severely affect database applications unless error constraints are carefully enforced in the watermarking process. In addition, a tradeoff between the watermarking error and the robustness of watermarking schemes has to be made. The concept of public key based watermark (or asymmetric watermark ) was first conceived in the multimedia context. Hachez and Quisquater summarized the work in this area in [14]. As mentioned in [14], one of the first ideas was proposed by Hartung and Girod [15] for watermarking compressed video. The basic idea is to make a part of the embedded watermark public such that a user can check the presence of this part of watermark. However, an attacker is able to remove this part of watermark and thus invalidate a public detector . Another idea is to embed private key information into a host signal and detect a correlation between the signal and a transformation of the signal using a public key [33]. Other correlation-based public watermarking schemes include [9, 30, 11]. However, such watermarks can be removed by certain attacks such as a sensitivity attack [6, 21] or confusing attack [34]. Craver and Katzenbeisser [8] used a zero knowledge protocol to prove the presence of a watermark in a signal "without revealing the exact location and nature of the watermark (specified by a private key)." As in most zero knowledge protocols, the proposed scheme requires many rounds of interactions between prover and verifier, which may not be efficient in practice. It is also not clear how to extend this scheme to watermarking relational databases. Because the original watermark is not certified and because a verifier is allowed to perform the protocol multiple times, this scheme may be subject to oracle attack (an attacker uses a public detector repeatedly to test modified signals so as to remove the watermark), plain-text chosen attack (a special case of oracle attack in which the tested signals are chosen by an attacker), or ambiguity attack (also called invert-ibility attack, in which a fake watermark is discovered from the watermarked signal). In comparison, our scheme requires no interaction between a verifier and the owner of data, thus is immune to both oracle attack and plain-text chosen attack. The watermark is certified in our scheme for thwarting the ambiguity attack (which we call additive attack in this paper). In addition, our scheme is both efficient and robust for typical database operations. CONCLUSION In this paper, we proposed a public watermarking scheme for relational databases. The scheme is unique in that it has the following properties. Public verifiability Given a database relation to be published or distributed, the owner of data uses a public watermark key to generate a public watermark, which is a relation with binary attributes. Anyone can use the watermark key and the watermark to check whether a suspicious copy of data is watermarked, and, if so, prove the ownership of the data by checking a watermark certificate officially signed by a trusted certificate authority, DB-CA. The watermark certificate contains the owner's ID, the watermark key, the hashes of both the watermark and DB relation, the first time the relation was certified, the validity period of the current certificate, and the DB-CA's signature. The watermark certificate may be revoked and re-certified in the case of identity change, ownership change, DB-CA compromise, or data update. Therefore, the revocation status also needs to be checked in ownership proof. To our best knowledge, our scheme is the only one to achieve public ownership proof in database literature. In contrast, all existing schemes are based on secret key, by which ownership cannot be proven more than once in public. Distortion free Different from typical watermarking schemes (e.g., [1]) for database ownership proof that hide watermark information in data by modifying least significant bits (LSBs), our scheme generates a public watermark from a collection of the most significant bits (MSBs). Our scheme does not modify any MSBs; therefore, it is distortion-free. The public watermark is a database relation that has the same primary key attribute as the original data, plus one or more binary attributes to store the MSBs. Even though the MSBs are publicly known, an attacker cannot modify them without introducing intolerable error to the underlying data. In comparison , all previous watermarking schemes for databases introduce some kind of distortion to the watermarked data. They either modify LSB's for numerical data (e.g., [1, 19, 20]), or exchange values among categorical data (e.g., [26, 2]). Those schemes work well for particular types of data only, while our scheme can be applied for any type of data distortion-free. Incremental updatability Following the line of [1], each tuple in a database relation is independently processed in our scheme. Neither watermark generation nor detection depends on any correlation or costly sorting among data items as required in [28, 26, 2]. Therefore, the scheme is particularly efficient for typical database operations, which are mostly tuple oriented. In the case of tuple insertion, deletion, or modification, the watermark can be easily updated by processing those relating tuples only, with simple computation of random sequence numbers and modulus operations. Due to the robustness of our scheme, the update of watermark certificate can be performed periodically after a batch of data updates. Robustness Since the ownership of data is proven after the data is published or distributed, it is crucial that our scheme is robust against various attacks that intend to invalidate watermark detection or ownership proof. The robustness of our scheme is measured in terms of: (i) false hit, the probability of detecting a valid watermark from non-watermarked data; (ii) false miss, the probability of not detecting a valid watermark from watermarked data due to attacks; and (iii) failure error rate, the least portion of data that has to be intolerably modified so as to defeat our watermark detection. Typical database attacks considered in this paper include tuple/attribute insertion, deletion, and random/selecitive value modification/suppression. Both theoretical analysis and experimental study show that our scheme is robust in terms of these measures, which can be adjusted by the watermark generation and detection parameters. We have also studied the tradeoff between the robustness and the overhead of our scheme. Our scheme is robust against the data frame-up attack and additive attack that may be more perilous to public watermarking schemes. The major contribution of this paper is the proposal of a public watermarking scheme that has the above properties. Though our scheme may not necessarily supersede secret key based schemes due to the overhead of using certificate and public watermark, we believe that it can be applied more practically in the real world for database ownership protection. Our future plan includes extending our scheme to other types of data such as XML and streaming data. REFERENCES [1] R. Agrawal and J. Kiernan. Watermarking relational databases. In Proceedings of VLDB, pages 155166, 2002. [2] E. Bertino, B. C. Ooi, Y. Yang, and R. Deng. Privacy and ownership preserving of outsourced medical data. In Proceedings of IEEE International Conference on Data Engineering , pages 521532, 2005. [3] D. Boneh and M. Franklin. Identity-based encryption from the weil pairing. In Proceedings of CRYPTO'2001, LNCS 2139, Springer-Varlag , pages 213229, 2001. [4] Coalition Against Database Piracy (CADP). Piracy is unacceptable in the information age or any other age, July 2, 2005. http://cadp.net/default.asp. [5] C. Cocks. An identity based encryption scheme based on quadratic residues. In Cryptography and Coding - Institute of Mathematics and Its Applications International Conference on Cryp- tography and Coding Proceedings of IMA 2001, LNCS 2260 , pages 360363, 2001. [6] I. J. Cox and J. M. G. Linnartz. Public watermarks and resistance to tampering. In Proceedings of International Conference on Image Processing , pages 36, 1997. [7] I. J. Cox, M. L. Miller, and J. A. Bloom. Digital Watermarking: Principles and Practice . Morgan Kaufmann, 2001. [8] S. Craver and S. Katzenbeisser. Security analysis of public-key watermarking schemes. In SPIE Vol. 4475, Mathematics of Data/Image Coding, Compression, and Encryption IV , pages 172182, 2001. [9] J. J. Eggers, J. K. Su, and B. Girod. Public key watermarking by eigenvectors of linear transforms. In Proceedings of European Signal Processing Conference (EUSIPCO) , 2000. [10] S. Farrell and R. Housley. An internet attribute certificate profile for authorization, internet draft, April, 2002. http://www.ietf.org/rfc/rfc3281.txt. [11] T. Furon, I. Venturini, and P. Duhamel. A unified approach of asymmetric watermarking schemes. In SPIE Vol. 4314, Security and Watermarking of Multimedia Contents III , pages 269279, 2001. [12] B. Gray and J. Gorelick. Database piracy plague. The Washington Times , March 1, 2004. http://www.washingtontimes.com. [13] D. Gross-Amblard. Query-preserving watermarking of relational databases and xml documents. In Proceedings of ACM Symposium on Principles of Database Systems (PODS) , pages 191201, 2003. [14] G. Hachez and J. Quisquater. Which directions for asymmetric watermarking. In Proceedings of XI European Signal Processing Conference (EUSIPCO), Vol. I , pages 283286, 2002. [15] F. Hartung and B. Girod. Fast public-key watermarking of compressed video. In Proceedings of IEEE International Conference on Speech and Signal Processing , 1997. [16] R. Housley, W. Ford, W. Polk, and D. Solo. Internet x.509 public key infrastructure certificate and crl profile, July 2, 2005. http://www.ietf.org/rfc/rfc2459.txt. [17] N. F. Johnson, Z. Duric, and S. Jajodia. Information Hiding: Steganography and WatermarkingAttacks and Countermeasures . Kluwer Publishers, 2000. [18] S. Katzenbeisser and F. A. Petitcolas, editors. Information Hiding Techniques for Steganography and Digital Watermarking . Artech House, 2000. [19] Y. Li, V. Swarup, and S. Jajodia. Constructing a virtual primary key for fingerprinting relational data. In Proceedings of ACM Workshop on Digital Rights Management (DRM) , October 2003. [20] Y. Li, V. Swarup, and S. Jajodia. Fingerprinting relational databases: Schemes and specialties. IEEE Transactions on Dependable and Secure Computing (TDSC) , 2(1):3445, 2005. [21] J. M. G. Linnartz and M. van Dijk. Analysis of the sensitivity attack against electronic watermarks in images. In Proceedings of 2nd Workshop on Information Hiding Workshop , 1998. [22] A. Menezes, P. C. van Oorschot, and S. A. Vanstone. Handbook of Applied Cryptography . CRC Press, 1997. [23] S. Micali. Efficient certificate revocation. In Technical Report: TM-542b. Massachusetts Institute of Technology. Cambridge, MA, USA, 1996. [24] B. Schneier. Applied Cryptography. John Wiley & Sons, Inc., 1996. [25] A. Shamir. Identity-based cryptosystems and signature schemes. In Proceedings of CRYPTO'84, LNCS 196, Springer-Varlag , pages 4753, 1984. [26] R. Sion. Proving ownership over categorical data. In Proceedings of IEEE International Conference on Data Engineering , pages 584596, 2004. [27] R. Sion, M. Atallah, and S. Prabhakar. Resilient information hiding for abstract semi-structures. In Proceedings of the Workshop on Digital Watermarking , 2003. [28] R. Sion, M. Atallah, and S. Prabhakar. Rights protection for relational data. In Proceedings of ACM SIGMOD International Conference on Management of Data , pages 98108, 2003. [29] R. Sion, M. Atallah, and S. Prabhakar. Resilient rights protection for sensor streams. In Proceedings of the Very Large Databases Conference , pages 732743, 2004. [30] J. Smith and C. Dodge. Developments in steganography. In Proceedings of 3rd International Workshop on Information Hiding , pages 7787, 1999. [31] G. W. Snedecor and W. G. Cochran. Statistical Methods. 8th edition, Iowa State Press, 1989. [32] L. Vaas. Putting a stop to database piracy. eWEEK, enterprise news and reviews , September 24, 2003. http://www.eweek.com/print article/0,3048,a=107965,00.asp. [33] R. G. van Schyndel, A. Z. Tirkel, and I. D. Svalbe. Key independent watermark detection. In Proceedings of IEEE International Conference on Multimedia Computing and Systems, Vol. 1 , 1999. [34] Y. Wu, F. Bao, and C. Xu. On the security of two public key watermarking schemes. In Proceedings of 4th IEEE Pacific-Rim Conference on Multimedia , 2003.

public verifiability;certificate;Relational database;watermark;ownership protection

157

Putting Integrated Information in Context: Superimposing Conceptual Models with SPARCE

A person working with diverse information sources--with possibly different formats and information models--may recognize and wish to express conceptual structures that are not explicitly present in those sources. Rather than replicate the portions of interest and recast them into a single, combined data source, we leave base information where it is and superimpose a conceptual model that is appropriate to the task at hand. This superimposed model can be distinct from the model(s) employed by the sources in the base layer. An application that superimposes a new conceptual model over diverse sources, with varying capabilities, needs to accommodate the various types of information and differing access protocols for the base information sources. The Superimposed Pluggable Architecture for Contexts and Excerpts (SPARCE) defines a collection of architectural abstractions, placed between superimposed and base applications, to demarcate and revisit information elements inside base sources and provide access to content and context for elements inside these sources. SPARCE accommodates new base information types without altering existing superimposed applications. In this paper, we briefly introduce several superimposed applications that we have built, and describe the conceptual model each superimposes. We then focus on the use of context in superimposed applications. We describe how SPARCE supports context and excerpts. We demonstrate how SPARCE facilitates building superimposed applications by describing its use in building our two, quite diverse applications.

Introduction When a physician prepares for rounds in a hospital intensive care unit, she often creates a quick synopsis of important problems, with relevant lab tests or observations, for each patient, as shown in Figure 1. The information is largely copied from elsewhere, e.g., from the patient medical record, or the laboratory system. Although the underlying data sources use various information structures, including dictated free text, tabular results and formatted reports, the physician may organize the selected information items into the simple cells or groups as shown in Figure 1 (without concern for the format or information model of the base sources). Each row contains information about a single patient, with the four columns containing patient identifying information, (a subset of) the patient's current problems, (a subset of) recent lab results or other reports, and notes (including a "To Do" list for the patient). While the information elements selected for this synopsis will generally suffice for the task at hand (patient rounds), the physician may need to view an element (such as a problem or a lab result) in the original source [Gorman 2000, Ash 2001]. However, this paper artefact obviously provides no means of automatically returning to the original context of an information element. In an ICU, we have observed a clinician actively working with a potentially diverse set of underlying information sources as she prepares to visit a patient, selecting bits of information from the various information sources, organizing them to suit the current purpose, possibly elaborating them with highlighting or annotation, or mixing them with new additional information, including new relationships among bits of information [Gorman 2000]. In our work [Delcambre 2001], we have put forth the notion of superimposed information for use in such scenarios . The superimposed layer contains marks, which are encapsulated addresses, to the information elements of interest in the base layer. More than that, the superimposed layer may contain additional information (beyond marks) and may be structured according to an appropriate conceptual model. We are particularly interested in viewing and manipulating base information using tools appropriate for the information source (e.g., Microsoft Word for .doc files, Adobe Acrobat for .PDF files, and an electronic medical record system for patient data). We have built several superimposed applications that use conceptual models that are quite different from those of any of the underlying base information sources. In past work we have implemented superimposed applications and models that rely solely on the ability of a base application to create a mark and to return to the marked region. In this paper, we explore the use of excerpts and context for marks in superimposed applica-71 tions. An excerpt consists of the extracted content for a mark and the context contains additional descriptive information (such as section heading and font characteristics) about the marked information. In Section 2 we present two superimposed applications that superimpose a new conceptual model over the base information (which is largely text documents), and makes use of excerpt and mark capabilities. In Section 3 we describe the notion of excerpts and contexts in more detail and provide the rationale for using middleware to access them. The main contribution of this paper is our architecture for building superimposed applications called the Superimposed Pluggable Architecture for Contexts and Excerpts (SPARCE), presented in Section 4. This architecture makes it easy for a developer to build superimposed applications, including those that superimpose a conceptual model that is different from any of the base conceptual models. The paper concludes with a discussion of how to structure and access context, a summary of related work, and conclusions and plans for future work, in Sections 5, 6, and 7, respectively. Sample Applications We present two superimposed applications built using SPARCE to demonstrate the ability to superimpose different conceptual models, over the same corpus of base information. These applications are designed for use in the Appeals Decision Process in the Forest Services of the US Department of Agriculture (USFS). USFS routinely makes decisions to solve (or prevent) problems concerning forests. The public may appeal any USFS decision after it is announced. The appeal process begins with a set period of time during which an appellant can send in an appeal letter that raises one or more issue with a USFS decision or the decision-making process. A USFS editor processes all appeal letters pertaining to a decision and prepares an appeal packet for a reviewing officer. An appeal packet contains all documents a reviewing officer might need to consult while formulating a recommended decision about the complete set of issues raised in the appeals. This set of documents is called the Records, Information, and Documentation (RID) section of the appeal packet. This section contains a RID letter that lists the issues raised and a summary response for each issue. An Editor synthesizes a RID letter using documents in the RID such as the Decision Notice, the Environmental Assessment, the Finding of No Significant Impact (FONSI), and specialists' reports. In the RID letter, the editor presents information from other documents in a variety of forms such as excerpts, summaries, and commentaries. In addition, the editor documents the location and identity of the information sources referenced in the RID letter. 2.1 RIDPad Composing a RID letter requires an editor to maintain a large working set of information. Since it is not unusual for an editor to be charged with preparing appeal packets for several decisions simultaneously, the editor may need to maintain several threads of organization. Though using documents in electronic form can be helpful, such use does not necessarily alleviate all problems. For example, the editor still needs to document the identity and location of information. In using electronic documents, the editor may have to cope with more than a dozen documents simultaneously. RIDPad is a superimposed application for the USFS appeal process. A USFS editor can use this application to collect and organize information needed to prepare a RID letter. A RIDPad instance is a collection of items and groups. An item is a superimposed information element associated with a mark. It has a name and a description. The name is user-defined and the description is the text excerpt from the associated mark. A group is a convenient collection of items and other groups. Figure 2 shows a RIDPad instance with information concerning the "Road 18 Caves" decision (made in the Pacific Northwest Region of USFS). The instance shown has eight items (labeled Summary, Details, Comparison of Issues, Alternative A, Alternative B, Statement, Details, and FONSI) in four groups (labeled Environmental Assessment, Proposed Action, Other Alternatives, and Decision). The group labeled "Environmental Assessment" contains two other groups. Figure 1: (Hand-drawn) Information summary as prepared by a resident prior to conducting rounds in a hospital intensive care unit (used with permission) 72 The information in the instance shown comes from three distinct base documents in two different base applications . (The item labeled "Comparison of Issues" contains an MS Excel mark; all other items contain MS Word marks.) All items were created using base-layer support included in the current implementation of SPARCE. Figure 2: A RIDPad Instance RIDPad affords many operations on items and groups. A user can create new items and groups, and move items between groups. The user can also rename, resize, and change visual characteristics such as colour and font for items and groups. With the mark associated with an item, the user can navigate to the base layer if necessary, or browse the mark's context from within RIDPad via the Context Browser (as shown in Figure 3). Briefly, the Context Browser is a superimposed application window with information related to a mark. Figure 3 shows the Context Browser for the item labelled "FONSI". From the context elements listed on the left we see that this item has both content and presentation kinds of context elements. The browser displays the value of the selected context element to the right. The formatted text content is currently selected and displayed in the browser. Figure 3: Context of a RIDPad Item RIDPad superimposes a simple conceptual model over the selected base information with Group and Item as the only model constructors. A group contains a name, size, location, and an ID. An item contains a name, description, size, location, and an ID. Items can occur within a Group and Groups can be nested within a Group. Figure 4 shows the model as a UML Class Diagram. The class RIDPadDoc represents the RIDPad instance which includes information that will likely be used to prepare the RIDPad document. Figure 4: RIDPad Information Model (Simplified) 2.2 Schematics Browser Appeal letters from different appellants in the USFS appeal process tend to share features. They all contain appellant names and addresses, refer to a Decision Notice, and raise issues. Such similarities suggest a schema for appeal letters. A superimposed schematic is an E-R schema superimposed over base information [Bowers 2002]. The Schematics Browser (see Figure 5) is a superimposed application that demonstrates the use of superimposed schematics. It is meant to allow USFS personnel to consider a set of appeal decisions to look for important issues or trends. The Schematics Browser might be used to support strategic planning activities. Figure 5: Schematics Browser Name RIDPadDoc ID Name Size Location Group ID Name Description Size Location Item Belongs to 0..1 * 0..1 * 0..1 Contains * 0..1 * ID Address Mark 73 Figure 5 shows an instance of a USFS appeal decision schematic opened in the Schematics Browser. The upper left frame lists instances of the appeal decision schematic. The user can select one of these instances, and then use the large middle frame to browse through information associated with the decision. The "1997 Ranch House Timber Sale" appeal decision is selected in Figure 5. This schematic allows the user to easily browse from a particular issue to the appeal letter(s) where the issue was raised to the appellant who raised the issue, for example. Marks into any number of base sources can be associated with entities, relationships, and attributes (but only one mark per entity and attribute). When an entity, relationship, or an attribute has an associated mark, a user can either visit the base layer or choose to view the excerpt from within the browser. Figure 6 shows a simplified version of the information model the Schematics Browser uses in superimposing the E-R model over base information. The browser stores all superimposed information in a relational database. This structure is a simple generic model that accommodates arbitrary Entity-Relationship style schematics. Name Schematic ID Name Description Entity ID Name Value Attribute 1 1..* 1 * 1 * ID Address Mark ID Name SchematicInst ID Address Mark Figure 6: Schematics Browser's Information Model Figure 7 uses the Schematic Browser's meta model to show a partial superimposed schematic instance. It shows an instance of the "1997 Ranch House Timber Sale" appeal decision schematic (also shown in Figure 5) and an Issue entity. It also shows the two attribute instances, desc and number , of the Issue entity. The desc attribute is associated with a mark instance (ID 41). In this simple implementation, the schematic instance data has its corresponding type information stored in the Name field. 2.3 Impact of Superimposed Information on Conceptual Model(s) Superimposed information introduces one significant modeling construct the mark. The mark spans between information at the superimposed layer and information in the various base layer sources. The mark thus serves as a bridge between the conceptual model used in the superimposed layer and the conceptual model used in a base information source. Name = Appeal Decision : Schematic ID = 2 Name = 1997 Ranch House Timber Sale : SchematicInst ID = 1 Name = Issue Description = Failed to meet Treaty and trust obligations : Entity ID = 1 Name = desc Value = The Forest Service i... : Attribute ID = 2 Name = number Value = 1 : Attribute ID = 41 Address = Win1997.pdf|1|79|115 : Mark Figure 7: Partial Superimposed Schematic Instance In the RIDPad application, the superimposed model consists of groups and items, where groups can be nested. This model is somewhat like a simplified XML model where groups are analogous to elements. But one important difference is that items contain marks, as opposed to PCDATA or other content. In a similar manner, the Schematics Browser uses a superimposed model that is similar to an entity-relationship model, but marks may appear as attribute values. In addition, each entity and relationship instance may be anchored, i.e., may be in one-to-one correspondence with a mark. Any superimposed application, by definition, includes marks in the superimposed layer. Thus, the conceptual model used in the superimposed layer must, necessarily, be extended to include marks in some manner. The use of marks has no impact on the conceptual model of the base layer. In fact, the use of marks, in general, requires no change to the base information or the base application. Marks encapsulate an address to an information element in the base source. Thus, the use of marks requires an addressing scheme for each base source that participates in a superimposed application. The addressing scheme may exploit the data model of the base information source. As an example, we could use XPath expressions to address information elements in an XML document. It is also possible to use addressing schemes that are independent of the data model used in the base information source. For example, a MS Word document could be converted to a PDF document and a user could create a mark using a bounding box where the interior of the box contains parts of individual characters. Regardless of the addressing scheme used in a mark, the superimposed layer is shielded from the details of the addressing scheme as well as the details of the conceptual model used in the base information source. Excerpts and Contexts Superimposed applications may want to incorporate contents of base-layer elements in the superimposed layer. For example, an application might use the extracted base-layer content as the label of a superimposed element. We call the contents of a base-layer element an excerpt. An excerpt can be of various types. For example it may be plain text, formatted text, or an image. An excerpt of one 74 type could also be transformed into other types. For example , formatted text in a word processor could also be seen as plain text, or as a graphical image. In addition to excerpts, applications may use other information related to base-layer elements. For example, an application may group superimposed information by the section in which the base-layer elements reside. To do so, the application needs to retrieve the section heading (assuming one exists) of each base-layer element. We call information concerning a base-layer element, retrieved from the base layer, its context. Presentation information such as font name and location information such as line number might be included in the context of a mark. The context of a base-layer element may contain more than one piece of information related to the base-layer element . Each such piece of information is a context element (and context is a collection of context elements). Figure 8: A Base-Layer Selection Figure 8 shows a fragment of an HTML page as displayed by a web browser. The highlighted region of the fragment is the marked region. Table 1 shows an excerpt and a few context elements of this marked region. The column on the left lists names of context elements whereas the column on the right shows values of those context elements. Name Value Excerpt Cheatgrass, Bromus tectorum, grows near many caves in this project area. HTML Cheatgrass,&nbsp; &lt;i&gt;Bromus tectorum &lt;/i&gt;, &nbsp; grows near many caves in this project area. Font name (Inherited) Times New Roman Font size (Inherited) 12 Table 1: Sample Context Elements of an HTML Mark Note that superimposed applications may access context information that a user might not explicitly access (or even be aware of). For example, consider the marked region shown in Figure 8. The HTML markup for this region (shown in Table 1) does not contain font information. If a superimposed application needs to display the mark's excerpt exactly as it is in the base layer, the application needs to examine the markup of the enclosing element, possibly traversing to the beginning of the document (because font characteristics can be inherited in HTML). The superimposed application may also need to examine the configuration of the Web browser to retrieve some or all of the format specification. Several kinds of context are possible for a mark. The following is a representative list of context kinds along with example context elements for each kind. Content: Text, graphics. Presentation: Font name, color. Placement: Line number, section. Sub-structure: Rows, sentences. Topology: Next sentence, next paragraph. Container: Containing paragraph, document. Application: Options, preferences. Contexts can vary across base-layer types. For example, the context of a mark to a region in a graphics-format base layer might include background colour and foreground colour, but not font name. However, the context of a mark to a selection in a web page might include all three elements. Contexts can also vary between marks of the same base-layer type. For example, an MS Word mark to text situated inside a table may have a "column heading" context element, but a mark to text not situated in a table does not include that context element. Lastly, the context of a mark itself may change with time. For example, the context of a mark to a figure inside a document includes a "caption" context element only as long as a caption is attached to that figure. Supporting excerpts and contexts for marks are a natural extension of our original notion of mark as an encapsulated address. Because we use the same mechanism to support both contexts and excerpts, we will often use the term "context" broadly to refer to both kinds of information about a base-layer element. Accessing information inside diverse base-layer types requires superimposed applications to work with a variety of base information models, addressing mechanisms, and access protocols. In addition, base applications may have different capabilities. For example, base applications may vary in their support for navigation or querying, but users of superimposed applications may want to navigate through selected base information elements seamlessly and uniformly, e.g., using the Schematics Browser. We use middleware to ease communication between the two layers and make up for deficiencies of base applications. And we want the middleware to allow independent evolution of components in these layers. By providing a uniform interface to base information and its context, the middleware reduces the complexity of superimposed applications and allows superimposed application developers to focus on the needs of their applications such as the intricacies of the conceptual model they aim to superimpose. SPARCE The Superimposed Pluggable Architecture for Contexts and Excerpts (SPARCE) is a middleware for mark and context management [Murthy 2003]. It is designed to be extensible in terms of supporting new base-layer types 75 and contexts, without adversely affecting existing superimposed applications. Figure 9: SPARCE Reference Model Figure 9 shows a reference model for SPARCE. The Mark Management module implements operations such as mark creation. It also maintains a repository of marks. The Context Management module is responsible for retrieving context of base information. This module depends on the Mark Management module to locate information inside base layers. The Clipboard module is modelled after the Clipboard object in operating systems such as Macintosh and MS Windows. The Superimposed Information Management module provides storage service to superimposed applications. We have developed a generic representation for information, called the Uni-Level Description [Bowers 2003], that can represent information (including superimposed information) structured according to various data models or representation schemes, such as XML, RDF or database models, in a uniform way. In this architecture, superimposed applications can choose whether they use this module for storage, or another storage manager. 4.1 Key Abstractions Table 2 provides a brief description of the classes and interfaces SPARCE uses for mark and context management . SPARCE supports context for three classes of objects: marks, containers, and applications (using the classes Mark, Container, and Application respectively). A Container is an abstraction for a base document (or a portion of that document). An Application is an abstraction for a base application. SPARCE also defines the interface Context-Aware Object to any base-layer element that supports context. The classes Mark, Container, and Application implement this interface. Superimposed applications use the class SPARCE Manager to create new marks and to retrieve existing marks. The SPARCE Manager maintains a repository of marks. SPARCE treats context as a property set (a collection of name-value pairs). Context is the entire set of properties of a base-layer element and a context element is any one property. For example, the text excerpt and font name of a mark are context elements. Modelling context as a property set makes it possible to support a variety of contexts, both across and within base layers, without affecting existing superimposed applications. This model also provides a uniform interface to context of any base-layer element, for any base-layer type. SPARCE uses the interface Context Agent to achieve its extensibility goal. A class that implements this interface takes a context-aware object and returns its context. That is, SPARCE does not access base-layer elements or their contexts directly. It uses external agents to do so on its behalf. However, SPARCE is responsible for associating a context-aware object with an appropriate context agent. The SPARCE Manager obtains the name of the class that will be the context agent for a mark from the description of the marks. The SPARCE Manager instantiates the context agent class by name whenever a superimposed application accesses the context of a context-aware object . Typically, there is one implementation of the context agent interface per base-layer type. For example, a PDF Agent is an implementation of this interface for use with PDF documents. A context agent implementation determines the constitution of context for its context-aware objects. SPARCE does not require an implementation to support particular context elements (nor does it prevent an implementation from defining any context element). However, we expect implementations to support kinds of context elements commonly expected (such as those listed in Section 3), and use meaningful names for context kinds and elements. Class/Interface Description Mark A mark to base-layer information. Container The base document (or a portion of it) in which a mark is made. Application The base application in which a mark is made. Context-Aware Object (interface) Interface to any base-layer element able to provide context. Classes Mark, Container, and Application implement this interface. Context Context of a context-aware object. It is a collection of context elements. Context Element A single piece of context information about a context-aware object. Context Agent (interface) Interface to any base-layer. An implementation will retrieve context from a context-aware object. SPARCE Manager Creates, stores, and retrieves marks; associates context-aware objects with appropriate context agents. Table 2: SPARCE Classes and Interfaces 4.2 Creating Marks A user initiates mark creation after selecting some information in a base application. The mark creation process consists of two steps: (1) generating the address of the selected base information, perhaps with other auxiliary information (collectively called mark fodder) and (2) creating a mark object in the mark repository. The address contained in mark fodder uses the addressing mechanism appropriate for the base information source. For example, the address to information inside a PDF document contains the page number and the starting and ending word indexes; the address to a selection in a spreadsheet contains the row and column numbers for the first and last cell in the selection. (Other addressing schemes are possible for these base types.) Superimposed Application Superimposed Information Management Mark Management Context Management Clipboard Base Application 76 Figure 10 depicts two possible mark-creation scenarios as a UML Use Case Diagram. (The boxes in this figure denote system boundaries; the broken arrows denote object flows.) In both scenarios, a user starts mark creation in a base application and completes it in a superimposed application. In the first scenario, labelled "Copy", the user is able to use the normal copy operation, e.g., of a word processor, to create the mark fodder. In the "Mark" use case, the user invokes a newly introduced function (such as the Mark menu item shown in Figure 8). The superimposed application retrieves the mark fodder from the Clipboard, and passes it to the SPARCE Manager. The SPARCE Manager creates a mark object (from the fodder), assigns it a unique ID, stores it in the mark repository, and returns the new object to the superimposed application. Copy User Mark Base Application Clipboard Operating System Complete Superimposed Application Figure 10: Two Mark-creation Scenarios The first scenario allows a user to select base information in a preferred base application and copy it to the Clipboard without having to learn any new application, tool, or process to create marks. However, supporting this scenario requires cooperative base applications such as Microsoft Word and Excel. Some base applications do not directly support Clipboard operations, but they provide mechanisms (such as plug-ins or add-ins) to extend their environments. A special mark creation tool or menu option can be inserted in to the user interface of such applications. The Mark use case in Figure 10 demonstrates this scenario. Early versions of Adobe Acrobat and Netscape Navigator are examples of base applications in this category. Figure 11 shows the internal representation of a mark. This mark corresponds to the selection in the HTML page shown in Figure 8. Superimposed applications do not have visibility of a mark's internal representation. They simply use the mark's interface to access its details. 4.3 Accessing Marks and Context A superimposed application sends a mark ID to the SPARCE Manager to retrieve the corresponding mark object from the marks repository. The SPARCE Manager instantiates an implementation of the context agent interface that is appropriate for the mark. The superimposed application can work with the mark object directly (for example, to navigate to the base layer) or can interact with the mark's context agent object (for example, to retrieve mark context). With a context object in hand, a superimposed application can find out what context elements are available. It can also retrieve values for context elements of interest. The superimposed application may use a context-element's value in various ways. For example, it may use the text content of the mark as a label, or it may apply the font characteristics of the marked region to some superimposed information. &lt;Mark ID=&quot;HTML2003Apr22065837YZXsmurthy&quot;&gt; &lt;Agent&gt;HTMLAgents.IEAgent&lt;/Agent&gt; &lt;Class&gt;HTMLMark&lt;/Class&gt; &lt;Address&gt;4398|4423&lt;/Address&gt; &lt;Description/&gt;Noxius Weeds in ea1.html &lt;/Description&gt; &lt;Excerpt&gt;Cheatgrass, Bromus tectorum, grows near many caves in this project area.&lt;/Excerpt&gt; &lt;Who&gt;smurthy&lt;/Who&gt; &lt;Where&gt;YZX&lt;/Where&gt; &lt;When&gt;2003-04-22 06:58:37&lt;/When&gt; &lt;ContainerID&gt;cdocsea1html&lt;/ContainerID&gt; &lt;/Mark&gt; Figure 11: Internal Representation of a Mark For ease of use, our design also allows the application to retrieve the value of a context element from the context-aware object or even from the context-agent object. An application developer may choose the access path that is most convenient to his or her particular situation. 4.4 Implementation We have implemented SPARCE for Microsoft-Windows operating systems using ActiveX technology [COM]. The current implementation includes support for the following base applications: MS Word, MS Excel, Adobe Acrobat (PDF files), and MS Windows Media Player (a variety of audio/video file types). The agents for these base applications support the following kinds of context: content, presentation, containment, placement, sub-structure, topology, document, and application. (Some possible context elements of these kinds are listed in Section 3.) We have implemented reusable view facilities such as the Context Browser to display the complete context of a context-aware object, and tabbed property pages to display properties of context-aware objects. We have also implemented a few testing aids. For example, we have implemented a generic context agent with limited functionality (that can be used with any base-layer type) to test integration with new base-layer types. The Context Browser is also a good testing tool when support for a new a base type is added or when definition of context is altered for a base type. 4.5 Extensibility Supporting new context elements is straightforward in SPARCE: The new context element name is just added to the property set. Superimposed applications may ignore the new context elements if they are not capable of handling them. Supporting new base-layer types is more involved. It requires a developer to understand the base layer and its addressing mechanisms. The developer must implement the context agent interface for the base-layer type. And the developer must implement a means to allow users to 77 select regions within this type of base information and copy mark fodder to the Clipboard. As we mention in Section 4.2, the developer might be able to exploit extensibility mechanisms of base applications for creating mark fodder. We have used the extensibility mechanism to add support for MS Word, MS Excel, Adobe Acrobat, and MS Windows Media Player. It took us about 7-12 hours to support each of these base types. The SPARCE implementation and the superimposed applications were not changed or recompiled when new base types were added. 4.6 Evaluation Our observations show that developing superimposed applications with SPARCE is relatively easy. Although the effort required to develop a superimposed application depends on the specifics of that application, using abstractions such as marks and contexts alleviate the need to model those entities in each application. For example, RIDPad is a complex application due to its graphical nature and the variety of operations it supports. However, we were able to develop that application in approximately 30 hours. As we added support for new base types using the extensibility mechanism of SPARCE, RIDPad was able to automatically work with the new base types. The original Schematics Browser application worked only with PDF files. The application was responsible for managing marks and interacting with Adobe Acrobat. The application had no context-management capabilities. We altered this application to use SPARCE and it instantaneously had access to all base-layer types SPARCE supported (and those it will support in future). In addition, it also had access to context of base information. In less than 7 hours, we were able to alter the Schematics Browser to use SPARCE. There are many ways to deploy the components of SPARCE and its applications (based on application and user needs). For example, RIDPad is expected to be a single-user application. Thus, all components of RIDPad and SPARCE may run on a single computer. In contrast, the Schematics Browser is likely to be used by many USFS personnel to browse schematic instances of past appeal decisions. That is, shared repositories of superimposed information and marks can be useful. Based on such analyses, we are currently in the process of evaluating different deployment configurations of SPARCE and its applications. In addition to studying performance of these configurations, we intend to explore the benefits of caching context information. Issues in Context Representation One of the areas of SPARCE design we are still exploring is the representation of context. We have considered defining contexts via data types (say, a context type for each base type), but feel that approach would be too restrictive. The set of context elements available for a mark might vary across a document. For example, a mark in a Word document might have a "column name" context element if it is in a table, but not otherwise. It is even possible that the context elements available for a single mark may change over time. For instance, the "image" context element might only be available while the invocation of the base application in which the mark was originally created is still running. A context type could define all possible context elements, where a particular mark produces null values on elements undefined for it, but that approach complicates the application programming interface (especially for context elements of scalar types such as numbers and strings). Another issue with types is making it possible to write a superimposed application without specifying in advance all the base sources it will be used with (and their context types). We have demonstrated with our current approach the ability of a superimposed application to work with new context agents without modifying the application. The superimposed application can make use of any context elements it knows about (from the elements the new agent supplies). While inheritance schemes can support some polymorphism in types, they do not seem adequate to support the arbitrary kinds of overlap we have seen among context elements across base types. Another issue is the internal structure of a context. Cur-rently a context is a property set of context elements, where each element is a name-value pair. Context elements also have kinds (such as presentation and substructure), which allows grouping context elements in user interfaces. We are considering giving contexts an explicit hierarchical structure. There are several alternatives for such an approach: Make a context a compound object capable of holding sub-contexts, use of qualified names (for example, format.font.fontsize), or employ a hierarchical namespace as in a directory structure. We do not see great differences in these three alternatives. The advantage of some kind of hierarchical structure, however , versus the current flat structure might come in the interface between superimposed applications and the context agent. Rather than the application asking for context elements individually (or for all context elements ), it could ask for a particular subgroup of elements of interest. A methodological issue related to context structure is how to coordinate the naming of context elements across multiple base types and multiple superimposed applications. There is no requirement currently that the "same" context element be named the same thing for different base types (or, in fact, in alternative context agents for the same base type). Even if the same name is used, the types of the associated values could be different. With an individual or small group writing context agents and superimposed applications, informal methods will work for consistency in naming. However, a more structured process will be needed at the point that context agents and superimposed applications are being produced by different organizations Related Work Memex and Evolutionary List File were visionary proposals for organizing information from diverse sources [Bush 1945, Nelson 1965]. Hypertext and compound document models are two classes of systems that attempt to realize these visions. Hypertext systems are helpful in 78 preparing information for non-linear media. Although designed to help organize information, they tend to be limited in the types of source, granularity of information, and location of information that can be organized. For example, NoteCards and Dexter both require information consulted to be stored in a proprietary database [Halasz 1987, 1994]. Intermedia can address base information only at sub-document granularity [Yankelovich 1988]. Hypertext systems typically do not support retrieval of contextual information from sources. Compound document systems are helpful in preparing information for linear media (such as paper). They can address base information at both document and sub-document granularity, but they tend to constrain display models of tools developed. For example, OLE 2 requires rectangular display areas [COM]. Like SPARCE (and unlike hypertext systems), compound document systems provide architectural support for building applications. Compound document systems support only retrieval of contents. Information sources decide the content, its format , and geometry. Table 3 provides a brief comparison of SPARCE with hypertext and compound document systems. NoteCards, Intermedia, and Dexter are hypertext systems. OpenDoc [Apple 1994] and OLE 2 are compound document systems . No teC a r d s In te rm e d ia De x t er Op e n D o c OLE 2 SPAR C E Base types 2 3 Any Any Any Any Base location Custom Files Custom Any Any Any Base granularity Whole Part Both Both Both Both Context kinds None None None Content Content Many Table 3: SPARCE Compared with Related Systems Multivalent documents [Phelps 2000b] allow multiple behaviours to be superimposed on to a single base document using an abstraction similar to the context-agent interface in SPARCE. The system uses contents of a region of interest (and its surrounding), but only to address that region [Phelps 2000a]. In the area related to dynamism and representation of context, OLE Automation [Microsoft 1996] provides an interesting comparison to our approach. An OLE automation object exposes an interface to the type information object (ITypeInfo) that corresponds to itself. The type information object is resident in a type library (that contains type information for possibly many automation object types). Changing type information (deleting members or adding new members) requires creation of a new type-information object and a new type library. Although the framework allows each instance of an object type to return a different type-information object, the requirement to create new type information and a type library makes it impractical to do so. Consequently, type information of an OLE automation object tends to include all possible elements, without regard to whether those members are relevant in a given situation. For example, the type information for a Range object of a MS Word document contains over 30 members [Microsoft]. The value of a member of scalar type that is not applicable for a given Range object will be equivalent of NULL (and the inapplicable collection-type members will be empty). In SPARCE, the context of a mark contains only those elements that apply to the mark. It might seem that links in OLE 2 compound documents provide similar functionality to marks. An OLE 2 compound document supports only retrieval of contents from links. It does not provide a mechanism from within a compound document to obtain the OLE automation object that corresponds to a link (even when the link source defines an automation object corresponding to the region the link represents). As a consequence, context-like information about the linked region cannot be accessed via the link directly. For example, linking a selection S from the main body of MS Word document D1 into Word document D2 makes D2 a compound document. However , the Range object for S (which is available in D1) is not accessible through the link in D2. A user needing more information about S must navigate to the source document D1. SPARCE not only provides the ability to link information via marks, it also provides access to context of the mark through the mark itself. Discussion and Future Work One way to view our work is that we have extended the standard modelling building blocks (integers, floats, dates, strings, etc.) with a new primitive--mark--that encapsulates an information element from an external source. A conceptual model (extended to be a superimposed model) can permit the use of marks in any of its structuring constructs (tuples, relationships, attributes, entities, etc.), without regard to the complexities of the underlying element. Support for context allows superimposed applications to extract information from that element and its surrounding elements or the information source in a controlled manner, to augment what is explicitly stored in the superimposed model. As a means to provide "new models for old data," our approach is quite different from data integration approaches such as mediators and data warehouses. Such approaches seek to provide an integrated view through a global schema describing base information that faithfully reflects the structure of the base source. In our work, we are exploring the use of selected base information elements (using marks). Note that the selection of marks is often performed manually, by a domain expert (e.g., a clinician or a USFS scientist), for a specific purpose (e.g., to treat a patient or prepare a RID). We have no requirement to represent the structure or relationships present within the base layer. Rather, we rely on the original application to provide interpretation for a mark and, if appropriate, to describe any relationships among marks. Standard integration approaches describe information from various sources and expect the mediator to be responsible for its interpretation. 79 The superimposed layer, by definition, allows the user to mix marks with additional information that may not exist in any of the base information sources. Such information may augment the base layer, e.g., by making implicit information explicit (e.g., "this issue relates only to Alternative A") or by providing commentary. Another use of superimposed information is to link related information from multiple sources, e.g., by placing marks in the same group or by explicitly linking between information elements in two sources. Finally, the superimposed approach permits reinterpretations that are much less structured than the original. For example, base information elements can be grouped or linked without having to observe any type constraints imposed in the original source. Exploring different representations of context and ways to reconcile context definition from different context agents is one area of our future work. Understanding the needs of new superimposed conceptual models (other than those we have described), and exploiting contexts to superimpose richer conceptual models is another area of our interest. A natural application of superimposed conceptual models would be to create means of querying jointly over superimposed and base information. We are also interested in superimposed applications that facilitate "schema later" organization of diverse information. That is, a user can start accumulating and arranging information items of interest, and--as he or she starts forming a mental conceptual model--incrementally define a superimposed model that reflects it. Acknowledgements This work has been supported by US NSF grants IIS 9817492 and IIS 0086002. We thank John Davis for helping us understand the USFS appeal process. We also thank the anonymous reviewers for their comments. References Acrobat SDK: Acrobat Software Development Kit, Adobe Systems Incorporated. Apple (1994): The OpenDoc Technical Summary. Apple World Wide Developers Conference Technologies CD, San Jose; CA. Ash, J., Gorman P., Lavelle, M., Lyman J., Delcambre, L., Maier, D., Bowers, S. and Weaver, M. (2001): Bundles: Meeting Clinical Information Needs. Bulletin of the Medical Library Association 89(3):294-296. Bowers, S., Delcambre, L. and Maier, D. (2002): Superimposed Schematics: Introducing E-R Structure for In-Situ Information Selections. Proc. ER 2002, pp 90 104, Springer LNCS 2503. Bowers, S. and Delcambre, L. (2003): The Uni-Level Description: A Uniform Framework for Representing Information in Multiple Data Models. Proc. of the 22 nd International Conference on Conceptual Modeling (ER 2003), Chicago, IL, October 2003. Bush, V. (1945): As We May Think. The Atlantic Monthly; 1945; July. Delcambre, L., Maier, D., Bowers, S., Weaver, M., Deng, L., Gorman, P., Ash, J., Lavelle, M. and Lyman, J. (2001): Bundles in Captivity: An Application of Superimposed Information. Proc. ICDE 2001, Heidelberg , Germany, pp 111-120. Gorman, P., Ash, J., Lavelle, M., Lyman, J., Delcambre, L. and Maier, D. (2000): Bundles in the wild: Managing information to solve problems and maintain situation awareness. Lib. Trends 2000 49(2):266-289. Halasz, F.G., Moran, T.P. and Trigg, R.H. (1987): NoteCards in a Nutshell. Proc. ACM CHI+GI Conference , New York, NY, pp 45-52, ACM Press. Halasz, F.G. and Schwartz, F. (1994): The Dexter Hypertext Reference Model. Communications of the ACM, 37(2):30-39, ACM Press. Maier, D. and Delcambre, L. (1999): Superimposed Information for the Internet. Proc. WebDB 1999 (informal ), Philadelphia, PA, pp 1-9. COM: The Component Object Model Specification, Microsoft Corporation. Microsoft Corporation. (1996): OLE Automation Pro-

superimposed information;SPARCE .;excerpts;Conceptual modelling;software architecture;context

158

Quality-of-Service in IP Services over Bluetooth Ad-Hoc Networks

Along with the development of multimedia and wireless networking technologies, mobile multimedia applications are playing more important roles in information access. Quality of Service (QoS) is a critical issue in providing guaranteed service in a low bandwidth wireless environment. To provide Bluetooth-IP services with differentiated quality requirements, a QoS-centric cascading mechanism is proposed in this paper. This innovative mechanism, composed of intra-piconet resource allocation, inter-piconet handoff and Bluetooth-IP access modules, is based on the Bluetooth Network Encapsulation Protocol (BNEP) operation scenario. From our simulations the handoff connection time for a Bluetooth device is up to 11.84 s and the maximum average transmission delay is up to 4e-05 s when seven devices join a piconet simultaneously. Increasing the queue length for the Bluetooth-IP access system will decrease the traffic loss rate by 0.02 per 1000 IP packets at the expense of a small delay performance.

Introduction Wireless communications have evolved rapidly over the past few years. Much attention has been given to research and development in wireless networking and personal mobile computing [10,17]. The number of computing and telecommunications devices is increasing and consequently, portable computing and communications devices like cellular phones, personal digital assistants, tablet PCs and home appliances are used widely. Wireless communication technologies will offer the subscriber greater flexibility and capability than ever before [14]. In February 1998, mobile telephony and computing leaders Ericsson, Nokia, IBM, Toshiba, and Intel formed a Special Interest Group (SIG) to create a standard radio interface named Bluetooth [13]. The main aim of Bluetooth has been the development of a wireless replacement for cables between electronic devices via a universal radio link in the globally available and unlicensed 2.4 GHz Industrial Scientific and Medical (ISM) frequency band [9]. Bluetooth technologies have the potential to ensure that the best services , system resources and quality are delivered and used efficiently . However, global services will embrace all types of networks. Therefore, bluetooth-based service networks will interconnect with IPv4/v6 existing networks to provide wide area network connectivity and Internet access to and between, individuals and devices [7]. In Reference [2], the BLUEPAC (BLUEtooth Public ACcess) concepts presented ideas for enabling mobile Bluetooth devices to access local area networks in public areas, such as airports, train stations and supermarkets . The Bluetooth specification defined how Bluetooth-enabled devices (BT) can access IP network services using the IETF Point-to-Point Protocol (PPP) and the Bluetooth Network Encapsulation Protocol (BNEP) [12,19,20]. By mapping IP addresses on the corresponding BT addresses (BD_ADDR), common access across networks is enabled [3]. This means that devices from different networks are allowed to discover and use one another's services without the need for service translation or user interaction. To support communications between all Bluetooth-based home appliances and the existing IP world, IPv6 over Bluetooth (6overBT) technology was proposed [8]. The 6overBT technology suggested that no additional link layer or encapsulation protocol headers be used. However, the development of 6overBT technology is still in progress. The BNEP protocol was referred to as the key technology in our research. What with the development of applications and wireless networking technologies, mobile multimedia applications are playing more important roles in information access [21]. To provide responsive multimedia services (high QoS) in a Bluetooth-IP mobile environment, a pre-requisite for our discussion is seamless data transmission. A cascading system with fair resource allocation scheme, seamless handoff strategy , and transparent bridging system that provides a way of integrating IP networks and Bluetooth-based service networks to relay multimedia applications within residual and enterprise is thus inevitable [6]. The rest of this paper is organized as follows. The following section describes Bluetooth background information, including the piconet, scatternet, IP over Bluetooth service architecture . Section 3 presents the proposed QoS-centric cascading mechanism, including the intra-piconet resource allocation , inter-piconet handoff, and Bluetooth-IP access system . The simulation model and performance analysis of the 700 W.-C. CHAN ET AL. queue length, loss rate, average delay, are introduced in section 4. Section 5 presents our concluding remarks. Bluetooth-IP services Figure 1 is a Bluetooth-IP service network environment. When a BT user wants to receive a networked video stored on a remote video server, the BT (maybe a slave in a picocell ) will initiate a connection procedure with the picocell master. The master initiates a connection procedure with the video server through the Bluetooth-IP access system. During the connection state, the video stream will be fed through the access system, master to BT user (the dotted line of figure 1). 2.1. Piconet When two BTs establish a connection, one BT will act as master and the other as the slave. More BTs can join the piconet . A single piconet can accommodate up to seven active slaves. If a slave does not need to participate in the channel , it should still be frequency-hop synchronized and switch to a low-power Park mode. The master can also request that the slave enter the Park mode so the master can communicate more than seven slave BTs. The master determines the frequency-hop sequence, the timing and the scheduling of all packets in the piconet. Within a piconet, the master initiates the connection procedure , although the application may necessitate that the master and slave roles be exchanged. For instance, such an exchange is necessary when a BT receives network services through Bluetooth-IP access systems. In this circumstance, the access system provides an IP service that may be used by many other BTs. This situation requires that the access system be the master of the piconet and the other BTs to act as slaves. However, when the device is initially activated and looks for an access system, it may be the device initiating the connection . This will make the device the master and the access system the slave [1]. 2.2. Scatternet Several piconets can be established and linked together to form an ad-hoc network. This is called a scatternet. A BT can participate in two or more overlaying piconets by applying time multiplexing. To participate on the proper channel, the BT should use the associated master device address and proper clock offset to obtain the correct phase. A BT can act as a slave in several piconets, but only as a master in a sin-Figure 1. Bluetooth-IP service network environment. QUALITY-OF-SERVICE IN IP SERVICES 701 Figure 2. An inter-piconet node in the scatternet. gle piconet: because two piconets with the same master are synchronous and use the same hopping sequence. This syn-chronization makes them one and the same piconet. Time multiplexing must be used to switch between piconets . In figure 2, an inter-piconet node is capable of time-sharing between multiple piconets. This allows the traffic to flow within and between the piconets [18]. In the case of data links only, a BT can request to enter the Hold or Park mode in the current piconet during which time it may join another piconet by just changing the channel parameters. BTs in the Sniff mode may have sufficient time to visit another piconet in between the Sniff slots. If audio links are established, other piconets can only be visited in the non-reserved slots. 2.3. IP networking The LAN Access profile defines IP service access using PPP over RFCOMM. TCP/IP runs on the PPP protocol. BTs can receive IP services using the PPP protocol [20]. When a BT wants to receive IP service, it will find a LAN Access Point (LAP) within radio range through inquiry and paging. After the data links have been setup, the LMP (Link_Manager Protocol) will process the master/slave switch and the L2CAP/RFCOMM/PPP connection will then be established. A suitable IP address is negotiated between devices in the PPP layer. The BT can forward IP packets through the LAP. Encapsulating an Ethernet packet inside a PPP packet is not an efficient solution. Moreover PPP is not sufficient for ad-hoc networks that contain multiple wireless hops. The best way of providing networking would be to Ethernet over the L2CAP layer. The Bluetooth Network Encapsulation Protocol (BNEP) was pursued by the Bluetooth SIG PAN working group to provide an Ethernet-like interface to IP services [19], as depicted in figure 3. QoS-centric cascading mechanism From figure 1, a BT that accesses multimedia services on the Internet may do so through a piconet or scatternet. QoS is critical in transmitting of different network segments. To provide Bluetooth-IP services with differentiated quality requirements , a QoS-centric cascading mechanism is proposed Figure 3. BNEP protocol stack. Figure 4. The proposed QoS-centric cascading modules. in our research to tunnel the guaranteed applications. The operational modules are illustrated in figure 4. The innovative mechanism consists of three modules: intra-piconet resource allocation, inter-piconet handoff and Bluetooth-IP access system . These modules were developed based on the BNEP operation scenarios. 3.1. Intra-piconet resource allocation Two service types: Synchronous Connection Oriented (SCO) and Asynchronous Connectionless Link (ACL) were defined in the Bluetooth service environment. Through the QoS setup, the ACL link can be configured to provide QoS requirements. The ACL link can be configured with the Flush Timeout setting , which prevents re-transmission when it is no longer useful. Acknowledgement can be received within 1.25 ms. This makes the delay small and it is possible to perform re-transmission for real-time applications. The ACL link also supports variable and asymmetrical bandwidth requirement applications. Currently, the IP QoS architecture is based purely on IP-layer decision making, packet buffering and scheduling through a single link-layer service access point. The mechanism in the link layer has better understanding of the communications medium status. However, simplicity has been important design objective for the Bluetooth interface and the number of IP-based protocols is becoming increasingly more complex. As depicted in figure 5, QoS architecture at the network layer such as differentiated and integrated services provides different services to applications. These services at 702 W.-C. CHAN ET AL. Figure 5. Bluetooth IP QoS mechanism. Figure 6. General QoS framework. the high layer are sufficient depending on the particular scenario . With shared bandwidth and a re-transmission scheme, the Host Controller (HC) buffering will invoke delays. The HC buffer size can be decreased to reduce the buffer delay. In bluetooth-based service layer, the L2CAP layer informs the remote side of the non-default parameters and sends a Configuration Request. The remote L2CAP-side responds to the Configuration Response that agrees or disagrees with these requirements. If the Configuration Response disagrees, the local side sends other parameters to re-negotiate or stop the connection. The Link Manager uses the poll interval and repetitions for broadcast packets to support QoS. The poll interval , which is defined as the maximum time between subsequent transmissions from the master to a particular slave on the ACL link, is used to support bandwidth allocation and latency control. Figure 6 depicts the general framework, which defines the basic functions required to support QoS. In figure 6, the traffic and QoS requirements for the QoS flow from the high layer sends the request to the Resource Requester (RR). Based on this request, the RR generates a resource request to the Resource Manager (RM). When the RM accepts the request the RR configures parameters to the local Resource Allocation (RA) entity. The RA actually reserves resources to satisfy the QoS requirements. The QoS is satisfied by the application that receives the appropriate resource. In our scheme, bluetooth-based operation identifies the following functions and procedures to determine the amount of resources assigned to a traffic flow. A polling algorithm determines which slave is polled next and bandwidth is assigned to that slave. The slave uses the air-interface scheduler to determine which data to send when it is polled. An inter-piconet scheduling algorithm is used by the inter-piconet node to efficiently control the traffic flow between two piconets . Bluetooth also uses the Flush Timeout setting to determine the maximum delay involved with re-transmissions. The Link Manager module in the master selects the baseband packet type for transmission in the single, three and five time slots [4]. 3.2. Inter-piconet handoff The inter-piconet environment suffers many challenges. First, the formation of Bluetooth networks is spontaneous and the problem of scatternet formation has not been defined in the Bluetooth specification [16]. Some researches have addressed these issues in the formation of an efficient scatternet [5,22]. The data sent forward between nodes must been sent via the master. Sometimes this data will traffic through multiple hops in the scatternet. Efficient routing protocols are needed for Bluetooth. The inter-piconet node as the bridge over which a piconet controls communications between piconets. Smarter traffic control is needed to coordinate with these masters. Different data rates exist in each link in different scenarios. The inter-piconet node becomes the bottleneck for the scatternet. In a piconet, the master controls all of the slaves in the piconet . The inter-piconet node joins more than two piconets, but it is only active in one piconet at a time. To efficiently move traffic between different piconets scheduling is needed to coordinate the inter-piconet and the master. In Reference [11], inter-piconet (IPS) and intra-piconet scheduling (IPRS) were presented to interact with one another to provide an efficient scatternet scheduler, as illustrated in figure 2. The IPS and IPRS must coordinate with one another to schedule when the inter-piconet node belongs to which piconet and how and when to transfer data packets between the different masters. Bluetooth connection progress includes two steps: the inquiry progress and the paging progress. This causes the bottleneck in the handoff. Two situations were discussed in BLUEPAC. When the Access Point is the master, the mobile node joins the piconet as a slave. The Access Point can efficiently control the traffic to the Internet. However, the disadvantage is that the Access Point must periodically enter the inquiry stage to find the newly arrived mobile node. This will interrupt the Access Point transfer a packet to the Internet. In a situation in which the BT is the master, the Access Point involves itself in multiple piconets. The BT can actively inquire the Access Point when it wants to connect to the Access Point. However, the traffic control becomes difficult and complex when the Access Point must switch between various piconets. The scheduler is still not supported for seamless handoffs in real time applications. To solve this problem, reference [15] proposes the Next Hop Handoff Protocol (NHHP) to support fast handoffs. The major focus is on reducing the connection time in the strategy. A scheme that passes the inquiry information to the next Access Point was used to avoid wasting time in the inquiry stage. The disadvantage is that the Access Points are divided into two categories: Entry Points QUALITY-OF-SERVICE IN IP SERVICES 703 which constantly make inquiries to the new BT and pass information to the Access Points in the internal regions have the responsibility in the handoff. If a newly arrived BT is initiated in an internal region, the scheme doesn't work. To resolve the above problem, a fast handoff scheme was proposed. This scheme assumes Bluetooth service environment is a micro-cellular network architecture. However, it also adapts Bluetooth as a macro-cellular network. Based on the fast handoff mission, the major focus is reducing the connection procedure that causes significant delay. We assumed the following conditions; the Access Point and mobile BT periodically scan for page attempts and inquiry requests. To obtain seamless and efficient handoff support, the Access Point RF range should cover the nearby Access Point. The neighborhood set records all Access Point locations. 3.2.1. Connection procedure As depicted in figure 7, when the mobile BT accesses the Internet it initiates an inquiry to the Access Point and makes a connection. The Access Point passes the BT addresses and clock information to the nearest Access Point according to the neighborhood set. The nearest Access Point uses this information to page the BT and form the piconets. These piconets form the scatternet and the BT becomes the inter-piconet node between them. The BT depends on the received signal strength indicator (RSSI) to determine which Access Point to use to access the Internet. The BT is a dedicated Access Point only in the connection state. The remaining piconets are all in the Hold state. 3.2.2. Handoff The BT periodically monitors the RSSI and bit error rate. When the RSSI decreases to the lower threshold a handoff may take place. To know where the mobile BT moves, the BT detects which RSSI becomes stronger. It then informs the Access Points and Foreign Agent that a handoff is imminent (figure 8(A)). The BT leaves the Hold state to the connection state in the piconet that contains the coming Access Point (AP0 in figure 8(B)). The routing path also changes to a new path. Additional caches may be needed in the Access Point to avoid packet losses. The new nearest Access Point (APa in figure 8(C)) in accordance with the neighbor set is notified that the mobile BT is within range to receive BT information . It begins to page the BT and enter the Hold state with the BT. The original connection state also turns into the Hold state in the piconet that contains the original access (AP1 in figure 8(D)). When the BT does not seek access service from the Access Point it should inform the Access Point that it no longer wants a connection. A connection could break down without prior warning. In the Bluetooth specification, both the master and slave use the link supervision time to supervise the loss. The supervision timeout period is chosen so that the value is longer than the hold periods. For simplicity of discussion we assumed that the Bluetooth AP is the application sender and divided the architecture into two parts, wired parts: the correspondent node (CN) to the Bluetooth AP and wireless part: the Bluetooth AP to the mobile BT device. The wired part is the same as the current general mechanism. We will only discuss the mechanism that combines the wireless part and our handoff mechanism in the Bluetooth environment. After making a connection and switching roles as mentioned earlier, the Bluetooth AP becomes a master. As illustrated in figure 9, the Bluetooth AP1 sends an active PATH message to the BT in figure 9(1). After the BT receives the PATH message, if the BT needs RSVP Figure 7. Connection procedures. 704 W.-C. CHAN ET AL. Figure 8. Handoff procedures. Figure 9. RSVP messages for bluetooth resource reservation. support, it sends a resource reservation request with a RESV message to the AP1 in figure 9(2). When the traffic specification contains a RSVP message, the traffic and QoS requirements for the QoS flow from the high layer sends a request to the Resource Requester (RR). The Bluetooth low levels will thus coordinate with one another. Once the Bluetooth AP1 accepts the request, the reservation along the flow between the AP1 and BT is made. After this point, the Bluetooth AP1 must have reservations in the neighboring APs. The resource reservation request is the same as an active reservation. The current AP1 sends a Passive PATH message to the neighboring AP2. The AP2 responds with a Passive RESV message and reserves the resources that the BT may need. Because Bluetooth can have only seven active slaves in a piconet at the same time, the resources must be used efficiently. One way is adding more Bluetooth devices in an AP. This can be easily achieved by modifying the application layer. As discussed before, to support seamless handoffs, information about the BT is sent to the AP2 after the RSVP process is performed. The Hold state is maintained between the BT and AP2. However, if the BT does not need a QoS guarantee, Figure 10. The protocol stack for Bluetooth-IP interworking. the QoS mechanism is not added. After supporting the end-to -end QoS using the RSVP protocol, the packet classification and scheduler can be used to control the traffic. 3.3. Bluetooth-IP access system Th difference between existing Bluetooth piconets and IP LANs is the communication protocol stack illustrated in figure 10. From figure 10, these differences are shown in the lower OSI seven layers. The lower layers are responsible for connection and addressing. We therefore focused on the connection management and address resolution issues in our research . The access function allows connections to be established without requiring any particular knowledge or interaction. The Bluetooth-IP access system plays the role of bridg-ing/routing multimedia traffic between the various LANs and piconets. Their operational scenario is illustrated in figure 11. When different piconet devices are connected directly to the access system, the access system function is referred to as a bridge (piconet Master and Slave role). If a LAN host (piconet BT) communicates with a piconet BT (LAN host) through the access system, because of the different protocol stacks between the piconet and LAN networks, two issues, addressing and connection must be resolved in the design. QUALITY-OF-SERVICE IN IP SERVICES 705 Figure 11. Bluetooth-IP interworking operational scenario. 3.3.1. Address resolution To achieve the interconnection function in a Bluetooth-IP environment , the access system must refer to both the piconet and LAN networks as members. Thus, two different protocol stacks must be combined to form a new communication protocol stack suitable as a routing server. The combined protocol stack is shown in figure 10. Using the protocol stack specifications , a scenario for addressing is identified as follows: Step 1: Each LAN host assigns an IP address. Each host thus possesses two addresses; an IP address and a MAC address. Step 2: Each piconet BT acquires two addresses, a BT address (BD_ADDR) and an IP address. Step 3: A routing table must be built for interconnection in the access system. A lookup for destination addresses is needed to find the corresponding outbound BD_ADDR or MAC address to which the packet must be forwarded. 3.3.2. Connection management Because the existing LAN is a packet switching service and BT connections are made on an ad hoc basis, interconnection is very difficult. To solve this problem, a mechanism based on IP services over Bluetooth was proposed in the Bluetooth specification. The system is as follows. Step 1: Bluetooth adapters and an Ethernet card are embed-ded into a desktop computer. These adapters and card are referred to as network attachment units (wired or Figure 12. The Bluetooth-IP access system. radio). Each port in the interface is directly attached to different networks (see figure 12). Step 2: The API of the Bluetooth adapter and the Ethernet packet driver are used to design an access system. The operational procedures for this system follow the scenario in figure 11. Performance analysis The Bluehoc toolkit was used to simulate the various scenarios in the Bluetooth-IP service environment. As depicted in figure 13, the data is dumped from the L2CA_DataWrite 706 W.-C. CHAN ET AL. Figure 13. L2CAP packet flow. and L2CA_DataRead into the connection state. The L2CA_DataWrite and L2CA_DataRead events are the upper-Layer to the L2CAP events. The L2CA_DataRead is the event that requests transfer for received data from the L2CAP entity to the upper layer. The L2CA_DataWrite is the event that requests data transfer from the upper layer to the L2CAP entity for transmission over an open channel. In the Bluehoc toolkit the connection procedures such as inquiry and paging are simulated according to the Bluetooth specifications. The master sends the QoS parameters, which depend on the application. QoS parameters are then passed to the Deficit Round Robin (DRR) based scheduler that determines if the connection can be accepted by the LMP. The DRR-based scheduler finds the appropriate ACL link baseband packet type (DM1, DM3, DM5, DH1, DH3 and DH5) depending upon the application level MTU and loss sensitivity . The simulation applications include packetized voice, Telnet and FTP. 4.2. Simulation results In figure 14 the average delay for various numbers of slaves using packetized voice in a piconet is shown. The voice application is real time and would be sensitive to a loss of several consecutive small packets. Figure 15 shows the average delay in the mixed traffic source. The short-packet delay, such as telnet applications, are significantly increased by the long-packet in the DRR-based scheduler. An efficient and simple scheme that does not add to the Bluetooth load is important. 4.2.1. Queue length analysis The queue length analysis was based on the access system queue size. We observed the variations in queue length using specified traffic levels. In figure 16, the queue length of the traffic from 100 M LAN to 1 M piconet increases very fast. It reaches 50000 packets within 10 s. The queue length for the traffic from 10 M LAN to 1 M piconet increases more smoothly and the queue length of the traffic from 1 M piconet to 100 M or 10 M LANs is almost zero. Figure 14. Average delay with voice services. Figure 15. Average delay in the mixed traffic. Figure 16. Queue length analysis. 4.2.2. Loss rate analysis In the loss rate analysis the queue length was changed to observe the variations in the loss rate. In figure 17, when the queue length is smaller than 1000 packets, the loss rate is close to 0.9. When the queue length increases to 5000 packets , the loss rate decreases to 0.8. Therefore, increasing the queue length will decrease the traffic loss rate. The loss rate from 100M LAN to 1 M piconet is a little more than that for 10 M LAN to 1 M piconet because of the difference in the LAN transport speed. QUALITY-OF-SERVICE IN IP SERVICES 707 Figure 17. Loss rate analysis. Figure 18. Throughput analysis. 4.2.3. Throughput analysis From figure 18, increasing the queue length has no effect on improving the throughput. The throughput is smooth in both the LAN to piconet and piconet to LAN traffic. 4.2.4. Delay analysis From figure 19, when the queue length is 500 packets, the delay is about 0.0005 seconds per bit. If the queue length increases to 1000 packets, the delay becomes almost double. When the queue length reaches 5000 packets, the delay is close to 0.003 seconds per bit. The transfer delay has no obvious change when the queue length increases to 10000 packets. Conclusions In a wireless environment the QoS guarantee provision becomes more important. The frequent mobility of a host increases the service disruption in real-time applications. Even though efficient RSVP enhances the resource reservation ability and allows for requesting a specific QoS from the network, these mechanisms do not have enough QoS guarantee to prevent service disruption during handoff. In this paper a cascading mechanism for QoS guarantee in a Bluetooth-IP environment was proposed. A fast and efficient handoff scheme that supports BT roaming handoffs between different Access Points was proposed. Concepts for the mobile RSVP issues in Bluetooth networks were presented with our fast handoff mechanism. The Bluetooth-IP access system can be implemented using available technology such as Network Addressing Translation (NAT) and Linux Bluetooth Stack to connect Bluetooth piconets and LAN. In our simulations Bluetooth required a long time to process the inquiry and paging procedures . These results show that the connection time is up to 11.84 sec when seven slaves join a piconet at the same time. Moreover, the access system queue length increases by about 10000 packets per second in a 100 Mbps LAN and about 1000 packets per second in a 10 Mbps LAN when the traffic load is 80%. In the loss rate analysis, the loss rate was close to 90% when the queue length was less than 1000 packets. However , when the queue length increased to 10000 packets the 708 W.-C. CHAN ET AL. Figure 19. Delay analysis. loss rate decreased to 70%. In the delay analysis, the delay was about 0.000045 seconds per bit when the queue length was 500 packets. The delay doubles when the queue length doubles. However, when the queue length is more than 5000 packets the delay has no obvious variance. Acknowledgement This paper is a partial result of project no. NSC-90-2218-E-259 -006 conducted by National Dong Hwa University under the sponsorship of National Science Council, ROC. References [1] M. Albrecht, M. Frank, P. Martini, M. Schetelig, A. Vilavaara and A. Wenzel, IP service over bluetooth: Leading the way to a new mobility, in: Proceedings of the International Conference on Local Computer Networks (1999) pp. 143154. [2] S. Baatz, M. Frank, R. Gopffarth, D. Kassatkine, P. Martini, M. Schetelig and A. Vilavaara, Handoff support for mobility with IP over bluetooth, in: Proceedings of the 25th Annual IEEE Conference on Local Computer Networks, USA (2000) pp. 143154. [3] J.L. Chen and K.C. Yen, Transparent bridging support for bluetooth-IP service interworking, International Journal of Network Management 12 (2002) 379386. [4] J.L. Chen, A.P. Shu, H.W. Tzeng and P.T. Lin, Fair scheduling for guaranteed services on personal area networks, in: Proceedings of 2002 International Conference on Communications, Circuits and Systems, China (2002) pp. 440444. [5] L. Ching and K.Y. Siu, A bluetooth scatternet formation algorithm, in: Proceedings of IEEE Global Telecommunications Conference, Vol. 5 (2001) pp. 28642869. [6] A. Das, A. Ghose, A. Razdan, H. Saran and R. Shorey, Enhancing performance of asynchronous data traffic over the bluetooth wireless ad-hoc network, in: Proceedings of the IEEE INFOCOM (2001) pp. 591 600. [7] D. Famolari and P. Agrawal, Architecture and performance of an em-bedded IP bluetooth personal area network, in: Proceedings of the International Conference on Personal Wireless Communications, India (2000) pp. 7579. [8] M. Frank, R. Goepffarth, W. Hansmann and U. Mueller, Transmission of native IPv6 over bluetooth, http://www.ispras.ru/~ipv6/ docs/draft-hansmann-6overbt-00.txt [9] C. Haartsen and S. Mattisson, Bluetooth a new low-power radio interface providing short-range connectivity, Proceedings of the IEEE 88(10) (2000) 16511661. [10] G. Ivano, D. Paolo and F. Paolo, The role of Internet technology in future mobile data systems, IEEE Communications Magazine 38(11) (2000) 6873. [11] P. Johansson, R. Kapoor, M. Kazantzidis and M. Gerla, Rendezvous scheduling in bluetooth scatternets, in: Proceedings of IEEE International Conference on Communications, Vol. 1 (2002) pp. 318324. [12] D.J.Y. Lee and W.C.Y. Lee, Ricocheting bluetooth, in: Proceedings of the 2nd International Conference on Microwave and Millimeter Wave Technology (2000) pp. 432435. [13] D.G. Leeper, A long-term view of short-range wireless, IEEE Computer 34(6) (2001) 3944. [14] Y.B. Lin and I. Chlamtac, Wireless and Mobile Network Architectures (Wiley, 2000). [15] I. Mahadevan and K.M. Sivalingam, An architecture and experimental results for quality if service in mobile networks using RSVP and CBQ, ACM/Baltzer Wireless Networks Journal 6(3) (2000) 221234. [16] B. Raman, P. Bhagwat and S. Seshan, Arguments for cross-layer opti-mizations in Bluetooth scatternets, in: Proceedings of 2001 Symposium on Applications and the Internet (2001) pp. 176184. [17] T.S. Rappaport, Wireless Communications Principles and Practice, 2nd ed. (2002). [18] T. Salonidis, P. Bhagwat, L. Tassiulas and R. LaMaire, Distributed topology construction of bluetooth personal area networks, in: Proceedings of the IEEE INFOCOM (2001) pp. 15771586. [19] The Bluetooth Special Interest Group, Bluetooth Network Encapsulation Protocol Specification (2001). [20] The Bluetooth Special Interest Group, Documentation available at http://www.bluetooth.com/techn/index.asp [21] The Bluetooth Special Interest Group, Quality of service in bluetooth networking, http://ing.ctit.utwente.nl/WU4/ Documents/Bluetooth_QoS_ING_A_part_I.pdf [22] V. Zaruba, S. Basagni and I. Chlamtac, Bluetrees-scatternet formation to enable bluetooth-based ad hoc networks, in: Proceedings of IEEE International Conference on Communications, Vol. 1 (2001) pp. 273 277. QUALITY-OF-SERVICE IN IP SERVICES 709 Wah-Chun Chan received the Ph.D. degree from University of British Columbia in 1965. He is cur-rently a Visiting Professor in the Department of Computer Science at National Chiao Tung University . Dr. Chan's research interest has been in the areas of queueing theory and telecommunication networks . Research on telecommunication networks has been in the development of models for the performance analysis of computer communication networks . Jiann-Liang Chen received the Ph.D. degree in electrical engineering from National Taiwan University , Taipei, Taiwan in 1989. Since August 1997, he has been with the Department of Computer Science and Information Engineering of National Dong Hwa University, where he is a Professor now. His current research interests are directed at cellular mobility management and personal communication systems

handoff;quality of service;Bluetooth-IP access system;BNEP protocol;resource allocation

159

Web Question Answering: Is More Always Better?

This paper describes a question answering system that is designed to capitalize on the tremendous amount of data that is now available online. Most question answering systems use a wide variety of linguistic resources. We focus instead on the redundancy available in large corpora as an important resource. We use this redundancy to simplify the query rewrites that we need to use, and to support answer mining from returned snippets. Our system performs quite well given the simplicity of the techniques being utilized. Experimental results show that question answering accuracy can be greatly improved by analyzing more and more matching passages. Simple passage ranking and n-gram extraction techniques work well in our system making it efficient to use with many backend retrieval engines.

INTRODUCTION Question answering has recently received attention from the information retrieval, information extraction, machine learning, and natural language processing communities [1][3][19][20] The goal of a question answering system is to retrieve `answers' to questions rather than full documents or even best-matching passages as most information retrieval systems currently do. The TREC Question Answering Track which has motivated much of the recent work in the field focuses on fact-based, short-answer questions such as "Who killed Abraham Lincoln?" or "How tall is Mount Everest?". In this paper we focus on this kind of question answering task, although the techniques we propose are more broadly applicable. The design of our question answering system is motivated by recent observations in natural language processing that, for many applications, significant improvements in accuracy can be attained simply by increasing the amount of data used for learning. Following the same guiding principle we take advantage of the tremendous data resource that the Web provides as the backbone of our question answering system. Many groups working on question answering have used a variety of linguistic resources part-of-speech tagging, syntactic parsing, semantic relations, named entity extraction, dictionaries, WordNet, etc. (e.g., [2][8][11][12][13][15][16]).We chose instead to focus on the Web as gigantic data repository with tremendous redundancy that can be exploited for question answering. The Web, which is home to billions of pages of electronic text, is orders of magnitude larger than the TREC QA document collection, which consists of fewer than 1 million documents. This is a resource that can be usefully exploited for question answering. We view our approach as complimentary to more linguistic approaches, but have chosen to see how far we can get initially by focusing on data per se as a key resource available to drive our system design. Automatic QA from a single, small information source is extremely challenging, since there is likely to be only one answer in the source to any user's question. Given a source, such as the TREC corpus, that contains only a relatively small number of formulations of answers to a query, we may be faced with the difficult task of mapping questions to answers by way of uncovering complex lexical, syntactic, or semantic relationships between question string and answer string. The need for anaphor resolution and synonymy, the presence of alternate syntactic formulations, and indirect answers all make answer finding a potentially challenging task. However, the greater the answer redundancy in the source data collection, the more likely it is that we can find an answer that occurs in a simple relation to the question. Therefore, the less likely it is that we will need to resort to solving the aforementioned difficulties facing natural language processing systems. EXPLOITING REDUNDANCY FOR QA We take advantage of the redundancy (multiple, differently phrased, answer occurrences) available when considering massive amounts of data in two key ways in our system. Enables Simple Query Rewrites. The greater the number of information sources we can draw from, the easier the task of rewriting the question becomes, since the answer is more likely to be expressed in different manners. For example, consider the difficulty of gleaning an answer to the question "Who killed Abraham Lincoln?" from a source which contains only the text "John Wilkes Booth altered history with a bullet. He will forever be known as the man who ended Abraham Lincoln's life," Question Rewrite Query &lt;Search Engine&gt; Collect Summaries, Mine N-grams Filter N-Grams Tile N-Grams N-Best Answers Where is the Louvre Museum located? "+the Louvre Museum +is located" "+the Louvre Museum +is +in" "+the Louvre Museum +is near" "+the Louvre Museum +is" Louvre AND Museum AND near in Paris France 59% museums 12% hostels 10% Figure 1. System Architecture Question Rewrite Query &lt;Search Engine&gt; Collect Summaries, Mine N-grams Filter N-Grams Tile N-Grams N-Best Answers Where is the Louvre Museum located? "+the Louvre Museum +is located" "+the Louvre Museum +is +in" "+the Louvre Museum +is near" "+the Louvre Museum +is" Louvre AND Museum AND near in Paris France 59% museums 12% hostels 10% Figure 1. System Architecture versus a source that also contains the transparent answer string, "John Wilkes Booth killed Abraham Lincoln." Facilitates Answer Mining. Even when no obvious answer strings can be found in the text, redundancy can improve the efficacy of question answering. For instance, consider the question: "How many times did Bjorn Borg win Wimbledon?" Assume the system is unable to find any obvious answer strings, but does find the following sentences containing "Bjorn Borg" and "Wimbledon", as well as a number: (1) Bjorn Borg blah blah Wimbledon blah blah 5 blah (2) Wimbledon blah blah blah Bjorn Borg blah 37 blah. (3) blah Bjorn Borg blah blah 5 blah blah Wimbledon (4) 5 blah blah Wimbledon blah blah Bjorn Borg. By virtue of the fact that the most frequent number in these sentences is 5, we can posit that as the most likely answer. RELATED WORK Other researchers have recently looked to the web as a resource for question answering. The Mulder system described by Kwok et al. [14] is similar to our approach in several respects. For each question, Mulder submits multiple queries to a web search engine and analyzes the results. Mulder does sophisticated parsing of the query and the full-text of retrieved pages, which is far more complex and compute-intensive than our analysis. They also require global idf term weights for answer extraction and selection, which requires local storage of a database of term weights. They have done some interesting user studies of the Mulder interface, but they have not evaluated it with TREC queries nor have they looked at the importance of various system components. Clarke et al. [9][10] investigated the importance of redundancy in their question answering system. In [9] they found that the best weighting of passages for question answering involves using both passage frequency (what they call redundancy) and a global idf term weight. They also found that analyzing more top-ranked passages was helpful in some cases and not in others. Their system builds a full-content index of a document collection, in this case TREC. In [10] they use web data to reinforce the scores of promising candidate answers by providing additional redundancy, with good success. Their implementation requires an auxiliary web corpus be available for full-text analysis and global term weighting. In our work, the web is the primary source of redundancy and we operate without a full-text index of documents or a database of global term weights. Buchholz's Shapaqa NLP system [7] has been evaluated on both TREC and Web collections. Question answering accuracy was higher with the Web collection (although both runs were poor in absolute terms), but few details about the nature of the differences are provided. These systems typically perform complex parsing and entity extraction for both queries and best matching web pages ([7][14]), which limits the number of web pages that they can analyze in detail. Other systems require term weighting for selecting or ranking the best-matching passages ([10][14]) and this requires auxiliary data structures. Our approach is distinguished from these in its simplicity (simple rewrites and string matching) and efficiency in the use of web resources (use of only summaries and simple ranking). We now describe how our system uses redundancy in detail and evaluate this systematically. SYSTEM OVERVIEW A flow diagram of our system is shown in Figure 1. The system consists of four main components. Rewrite Query . Given a question, the system generates a number of rewrite strings, which are likely substrings of declarative answers to the question. To give a simple example, from the question "When was Abraham Lincoln born?" we know that a likely answer formulation takes the form "Abraham Lincoln was born on &lt;DATE&gt;". Therefore, we can look through the collection of documents, searching for such a pattern. We first classify the question into one of seven categories, each of which is mapped to a particular set of rewrite rules. Rewrite rule sets range in size from one to five rewrite types. The output of the rewrite module is a set of 3-tuples of the form [ string, L/R/-, weight ], where " string " is the reformulated 292 search query, " L/R/-" indicates the position in the text where we expect to find the answer with respect to the query string (to the left, right or anywhere) and " weight " reflects how much we prefer answers found with this particular query. The idea behind using a weight is that answers found using a high precision query (e.g., "Abraham Lincoln was born on") are more likely to be correct than those found using a lower precision query (e.g., "Abraham" AND "Lincoln" AND "born"). We do not use a parser or part-of-speech tagger for query reformulation, but do use a lexicon in order to determine the possible parts-of-speech of a word as well as its morphological variants. We created the rewrite rules and associated weights manually for the current system, although it may be possible to learn query-to-answer reformulations and weights (e.g. see Agichtein et al. [4]; Radev et al. [17]). The rewrites generated by our system are simple string-based manipulations. For instance, some question types involve query rewrites with possible verb movement; the verb "is" in the question "Where is the Louvre Museum located?" should be moved in formulating the desired rewrite to "The Louvre Museum is located in". While we might be able to determine where to move a verb by analyzing the sentence syntactically, we took a much simpler approach. Given a query such as "Where is w 1 w 2 ... w n ", where each of the w i is a word, we generate a rewrite for each possible position the verb could be moved to (e.g. "w 1 is w 2 ... w n ", "w 1 w 2 is ... w n ", etc). While such an approach results in many nonsensical rewrites (e.g. "The Louvre is Museum located in"), these very rarely result in the retrieval of bad pages, and the proper movement position is guaranteed to be found via exhaustive search. If we instead relied on a parser, we would require fewer query rewrites, but a misparse would result in the proper rewrite not being found. For each query we also generate a final rewrite which is a backoff to a simple ANDing of the non-stop words in the query. We could backoff even further to ranking using a best-match retrieval system which doesn't require the presence of all terms and uses differential term weights, but we did not find that this was necessary when using the Web as a source of data. There are an average of 6.7 rewrites for the 500 TREC-9 queries used in the experiments described below. As an example, the rewrites for the query "Who created the character of Scrooge?" are: LEFT_5_"created +the character +of Scrooge" RIGHT_5_"+the character +of Scrooge +was created +by" AND_2_"created" AND "+the character" AND "+of Scrooge" AND_1_"created" AND "character" AND "Scrooge" To date we have used only simple string matching techniques. Soubbotin and Soubbotin [18] have used much richer regular expression matching to provide hints about likely answers, with very good success in TREC 2001, and we could certainly incorporate some of these ideas in our rewrites. Note that many of our rewrites require the matching of stop words like "in" and "the", in the above example. In our system stop words are important indicators of likely answers, and we do not ignore them as most ranked retrieval systems do, except in the final backoff AND rewrite. The query rewrites are then formulated as search engine queries and sent to a search engine from which page summaries are collected and analyzed. Mine N-Grams . From the page summaries returned by the search engine, n-grams are mined. For reasons of efficiency, we use only the returned summaries and not the full-text of the corresponding web page. The returned summaries contain the query terms, usually with a few words of surrounding context. In some cases, this surrounding context has truncated the answer string, which may negatively impact results. The summary text is then processed to retrieve only strings to the left or right of the query string, as specified in the rewrite triple. 1-, 2-, and 3-grams are extracted from the summaries. An N-gram is scored according the weight of the query rewrite that retrieved it. These scores are summed across the summaries that contain the n-gram (which is the opposite of the usual inverse document frequency component of document/passage ranking schemes). We do not count frequency of occurrence within a summary (the usual tf component in ranking schemes). Thus, the final score for an n-gram is based on the rewrite rules that generated it and the number of unique summaries in which it occurred. When searching for candidate answers, we enforce the constraint that at most one stopword is permitted to appear in any potential n-gram answers. The top-ranked n-grams for the Scrooge query are: Dickens 117 Christmas Carol 78 Charles Dickens 75 Disney 72 Carl Banks 54 A Christmas 41 uncle 31 Filter/Reweight N-Grams. Next, the n-grams are filtered and reweighted according to how well each candidate matches the expected answer-type, as specified by a handful of handwritten filters. The system uses filtering in the following manner. First, the query is analyzed and assigned one of seven question types, such as who-question, what-question, or how-many-question. Based on the query type that has been assigned, the system determines what collection of filters to apply to the set of potential answers found during n-gram harvesting. The answers are analyzed for features relevant to the filters, and then rescored according to the presence of such information. A collection of approximately 15 filters were developed based on human knowledge about question types and the domain from which their answers can be drawn. These filters used surface string features, such as capitalization or the presence of digits, and consisted of handcrafted regular expression patterns. After the system has determined which filters to apply to a pool of candidate answers, the selected filters are applied to each candidate string and used to adjust the score of the string. In most cases, filters are used to boost the score of a potential answer when it has been determined to possess the features relevant to the query type. In other cases, filters are used to remove strings from the candidate list altogether. This type of exclusion was only performed when the set of correct answers was determined to be a 293 closed set (e.g. "Which continent....?") or definable by a set of closed properties (e.g. "How many...?"). Tile N-Grams. Finally, we applied an answer tiling algorithm, which both merges similar answers and assembles longer answers out of answer fragments. Tiling constructs longer n-grams from sequences of overlapping shorter n-grams. For example, &quot;A B C&quot; and &quot;B C D&quot; is tiled into &quot;A B C D.&quot; The algorithm proceeds greedily from the top-scoring candidate - all subsequent candidates (up to a certain cutoff) are checked to see if they can be tiled with the current candidate answer. If so, the higher scoring candidate is replaced with the longer tiled n-gram, and the lower scoring candidate is removed. The algorithm stops only when no n-grams can be further tiled. The top-ranked n-grams after tiling for the Scrooge query are: Charles Dickens 117 A Christmas Carol 78 Walt Disney's uncle 72 Carl Banks 54 uncle 31 Our system works most efficiently and naturally with a backend retrieval system that returns best-matching passages or query-relevant document summaries. We can, of course, post-process the full text of matching documents to extract summaries for n-gram mining, but this is inefficient especially in Web applications where the full text of documents would have to be downloaded over the network at query time. EXPERIMENTS For our experimental evaluations we used the first 500 TREC-9 queries (201-700) [19]. For simplicity we ignored queries which are syntactic rewrites of earlier queries (701-893), although including them does not change the results in any substantive way. We used the patterns provided by NIST for automatic scoring. A few patterns were slightly modified to accommodate the fact that some of the answer strings returned using the Web were not available for judging in TREC-9. We did this in a very conservative manner allowing for more specific correct answers (e.g., Edward J. Smith vs. Edward Smith) but not more general ones (e.g., Smith vs. Edward Smith), and simple substitutions (e.g., 9 months vs. nine months). These changes influence the absolute scores somewhat but do not change relative performance, which is our focus here. Many of the TREC queries are time sensitive that is, the correct answer depends on when the question is asked. The TREC database covers a period of time more than 10 years ago; the Web is much more current. Because of this mismatch, many correct answers returned from the Web will be scored as incorrect using the TREC answer patterns. 10-20% of the TREC queries have temporal dependencies. E.g., Who is the president of Bolivia? What is the exchange rate between England and the U. S.? We did not modify the answer key to accommodate these time differences, because this is a subjective job and would make comparison with earlier TREC results impossible. For the main Web retrieval experiments we used Google as a backend because it provides query-relevant summaries that make our n-gram mining techniques more efficient. Thus we have access to more than 2 billion web pages. For some experiments in TREC retrieval we use the standard QA collection consisting of news documents from Disks 1-5. The TREC collection has just under 1 million documents [19]. All runs are completely automatic, starting with queries and generating a ranked list of 5 candidate answers. Candidate answers are a maximum of 50 bytes long, and typically much shorter than that. We report the Mean Reciprocal Rank (MRR) of the first correct answer, the Number of Questions Correctly Answered (NumCorrect), and the Proportion of Questions Correctly Answered (PropCorrect). Correct answers at any rank are included in the number and proportion correct measures. Most correct answers are at the top of the list -- 70% of the correct answers occur in the first position and 90% in the first or second positions. Using our system with default settings for query rewrite weights, number of summaries returned, etc. we obtain a MRR of 0.507 and answer 61% of the queries. The average answer length was 12 bytes, so the system is really returning short answers not passages. This is very good performance and would place us near the top of 50-byte runs for TREC-9. However, since we did not take part in TREC-9 it is impossible to compare our results precisely with those systems (e.g., we used TREC-9 for tuning our TREC-10 system increasing our score somewhat, but we return several correct answers that were not found in TREC-9 thus decreasing our score somewhat). Redundancy is used in two key ways in our data-driven approach. First, the occurrence of multiple linguistic formulations of the same answers increases the chances of being able to find an answer that occurs within the context of a simple pattern match with the query. Second, answer redundancy facilitates the process of answer extraction by giving higher weight to answers that occur more often (i.e., in more different document summaries). We now evaluate the contributions of these experimentally. 5.1 Number of Snippets We begin by examining the importance of redundancy in answer extraction. To do this we vary the number of summaries (snippets) that we get back from the search engine and use as input to the n-gram mining process. Our standard system uses 100 snippets. We varied the number of snippets from 1 to 1000. The results are shown in Figure 2. Performance improves sharply as the number of snippets increases from 1 to 50 (0.243 MRR for 1 snippet, 0.370 MRR for 5, 0.423 MRR for 10, and 0.501 for 50), somewhat more slowly after that 0 0.1 0.2 0.3 0.4 0.5 0.6 1 10 100 1000 Num ber of Snippets MR R Figure 2. MRR as a function of number of snippets returned. TREC-9, queries 201-700. 294 (peaking 0.514 MRR with 200 snippets), and then falling off somewhat after that as more snippets are included for n-gram analysis. Thus, over quite a wide range, the more snippets we consider in selecting and ranking n-grams the better. We believe that the slight drop at the high end is due to the increasing influence that the weaker rewrites have when many snippets are returned. The most restrictive rewrites return only a few matching documents. Increasing the number of snippets increases the number of the least restrictive matches (the AND matches), thus swamping the restrictive matches. In addition, frequent n-grams begin to dominate our rankings at this point. An example of failures resulting from too many AND matches is Query 594: What is the longest word in the English language? For this query, there are 40 snippets matching the rewrite "is the longest word in the English language" with weight 5, 40 more snippets matching the rewrite "the longest word in the English language is" with the weight 5, and more than 100 snippets matching the backoff AND query ("longest" AND "word" AND "English" AND "language") with a weight of 1. When 100 snippets are used, the precise rewrites contribute almost as many snippets as the AND rewrite. In this case we find the correct answer, "pneumonoultramicroscopicsilicovolcanokoniosis", in the second rank. The first answer, "1909 letters long", which is incorrect, also matches many precise rewrites such as "the longest word in English is ## letters long", and we pick up on this. When 1000 snippets are used, the weaker AND rewrites dominate the matches. In this case, the correct answer falls to seventh on the list after "letters long", "one syllable", "is screeched", "facts", "stewardesses" and "dictionary", all of which occur commonly in results from the least restrictive AND rewrite. A very common AND match contains the phrase "the longest one-syllable word in the English language is screeched", and this accounts for two of our incorrect answers. Using differential term weighting of answer terms, as many retrieval systems do, should help overcome this problem to some extent but we would like to avoid maintaining a database of global term weights. Alternatively we could refine our weight accumulation scheme to dampen the effects of many weakly restrictive matches by sub-linear accumulation, and we are currently exploring several alternatives for doing this. Our main results on snippet redundancy are consistent with those reported by Clarke et al. [9], although they worked with the much smaller TREC collection. They examined a subset of the TREC-9 queries requiring a person's name as the answer. They varied the number of passages retrieved (which they call depth) from 25 to 100, and observed some improvements in MRR. When the passages they retrieved were small (250 or 500 bytes) they found improvement, but when the passages were larger (1000 or 2000 bytes) no improvements were observed. The snippets we used are shorter than 250 bytes, so the results are consistent. Clarke et al. [9] also explored a different notion of redundancy (which they refer to as c i ). c i is the number of different passages that match an answer. Their best performance is achieved when both c i and term weighting are used to rank passages. We too use the number of snippets that an n-gram occurs in. We do not, however, use global term weights, but have tried other techniques for weighting snippets as described below. 5.2 TREC vs. Web Databases Another way to explore the importance of redundancy is to run our system directly on the TREC documents. As noted earlier, there are three orders of magnitude more documents on the Web than in the TREC QA collection. Consequently, there will be fewer alternative ways of saying the same thing and fewer matching documents available for mining the candidate n-grams. We suspect that this lack of redundancy will limit the success of our approach when applied directly on TREC documents. While corpus size is an obvious and important difference between the TREC and Web collections there are other differences as well. For example, text analysis, ranking, and snippet extraction techniques will all vary somewhat in ways that we can not control. To better isolate the size factor, we also ran our system against another Web search engine. For these experiments we used only the AND rewrites and looked at the first 100 snippets. We had to restrict ourselves to AND rewrites because some of the search engines we used do not support the inclusion of stop words in phrases, e.g., "created +the character +of Scrooge". 5.2.1 TREC Database The TREC QA collection consists of just under 1 million documents. We expect much less redundancy here compared to the Web, and suspect that this will limit the success of our approach. An analysis of the TREC-9 query set (201-700) shows that no queries have 100 judged relevant documents. Only 10 of the 500 questions have 50 or more relevant documents, which the results in Figure 2 suggest are required for the good system performance. And a very large number of queries, 325, have fewer than 10 relevant documents. We used an Okapi backend retrieval engine for the TREC collection. Since we used only Boolean AND rewrites, we did not take advantage of Okapi's best match ranking algorithm. However, most queries return fewer than 100 documents, so we wind up examining most of the matches anyway. We developed two snippet extraction techniques to generate query-relevant summaries for use in n-gram mining. A Contiguous technique returned the smallest window containing all the query terms along with 10 words of context on either side. Windows that were greater than 500 words were ignored. This approach is similar to passage retrieval techniques albeit without differential term weighting. A Non-Contiguous technique returned the union of two word matches along with 10 words of context on either side. Single words not previously covered are included as well. The search engine we used for the initial Web experiments returns both contiguous and non-contiguous snippets. Figure 3 shows the results of this experiment. MRR NumCorrect PropCorrect Web1 0.450 281 0.562 TREC, Contiguous Snippet 0.186 117 0.234 TREC, Non-Contiguous Snippet 0.187 128 0.256 AND Rewrites Only, Top 100 Figure 3: Web vs. TREC as data source 295 Our baseline system using all rewrites and retrieving 100 snippets achieves 0.507 MRR (Figure 2). Using only the AND query rewrites results in worse performance for our baseline system with 0.450 MRR (Figure 3). More noticeable than this difference is the drop in performance of our system using TREC as a data source compared to using the much larger Web as a data source. MRR drops from 0.450 to 0.186 for contiguous snippets and 0.187 for non-contiguous snippets, and the proportion of questions answered correctly drops from 56% to 23% for contiguous snippets and 26% for non-contiguous snippets. It is worth noting that the TREC MRR scores would still place this system in the top half of the systems for the TREC-9 50-byte task, even though we tuned our system to work on much larger collections. However, we can do much better simply by using more data. The lack of redundancy in the TREC collection accounts for a large part of this drop off in performance. Clarke et al. [10] have also reported better performance using the Web directly for TREC 2001 questions. We expect that the performance difference between TREC and the Web would increase further if all the query rewrites were used. This is because there are so few exact phrase matches in TREC relative to the Web, and the precise matches improve performance by 13% (0.507 vs. 0.450). We believe that database size per se (and the associated redundancy) is the most important difference between the TREC and Web collections. As noted above, however, there are other differences between the systems such as text analysis, ranking, and snippet extraction techniques. While we can not control the text analysis and ranking components of Web search engines, we can use the same snippet extraction techniques. We can also use a different Web search engine to mitigate the effects of a specific text analysis and ranking algorithms. 5.2.2 Another Web Search Engine For these experiments we used the MSNSearch search engine. At the time of our experiments, the summaries returned were independent of the query. So we retrieved the full text of the top 100 web pages and applied the two snippet extraction techniques described above to generate query-relevant summaries. As before, all runs are completely automatic, starting with queries, retrieving web pages, extracting snippets, and generating a ranked list of 5 candidate answers. The results of these experiments are shown in Figure 4. The original results are referred to as Web1 and the new results as Web2. MRR NumCorrect PropCorrect Web1 0.450 281 0.562 TREC, Contiguous Snippet 0.186 117 0.234 TREC, Non-Contiguous Snippet 0.187 128 0.256 Web2, Contiguous Snippet 0.355 227 0.454 Web2, Non-Contiguous Snippet 0.383 243 0.486 AND Rewrites Only, Top 100 Figure 4: Web vs. TREC as data source The Web2 results are somewhat worse than the Web1 results, but this is expected given that we developed our system using the Web1 backend, and did not do any tuning of our snippet extraction algorithms. In addition, we believe that the Web2 collection indexed somewhat less content than Web1 at the time of our experiments, which should decrease performance in and of itself. More importantly, the Web2 results are much better than the TREC results for both snippet extraction techniques, almost doubling MRR in both cases. Thus, we have shown that QA is more successful using another large Web collection compared to the small TREC collection. The consistency of this result across Web collections points to size and redundancy as the key factors. 5.2.3 Combining TREC and Web Given that the system benefits from having a large text collection from which to search for potential answers, then we would expect that combining both the Web and TREC corpus would result in even greater accuracy. We ran two experiments to test this. Because there was no easy way to merge the two corpora, we instead combined the output of QA system built on each corpus. For these experiments we used the original Web1 system and our TREC system. We used only the AND query rewrites, looked at the Top1000 search results for each rewrite, and used a slightly different snippet extraction technique. For these parameter settings, the base TREC-based system had a MRR of .262, the Web-based system had a MRR of .416. First, we ran an oracle experiment to assess the potential gain that could be attained by combining the output of the Web-based and TREC-based QA systems. We implemented a "switching oracle", which decides for each question whether to use the output from the Web-based QA system or the TREC-based QA system, based upon which system output had a higher ranking correct answer. The switching oracle had a MRR of .468, a 12.5% improvement over the Web-based system. Note that this oracle does not precisely give us an upper bound, as combining algorithms (such as that described below) could re-order the rankings of outputs. Next, we implemented a combining algorithm that merged the outputs from the TREC-based and Web-based systems, by having both systems vote on answers, where the vote is the score assigned to a particular answer by the system. For voting, we defined string equivalence such that if a string X is a substring of Y, then a vote for X is also a vote for Y. The combined system achieved a MRR of .433 (a 4.1% improvement over the Web-based system) and answered 283 questions correctly. 5.3 Snippet Weighting Until now, we have focused on the quantity of information available and less on its quality. Snippet weights are used in ranking n-grams. An n-gram weight is the sum of the weights for all snippets in which that n-gram appears. Each of our query rewrites has a weight associated with it reflecting how much we prefer answers found with this particular query. The idea behind using a weight is that answers found using a high precision query (e.g., "Abraham Lincoln was born on") are more likely to be correct than those found using a lower precision query (e.g., "Abraham" AND "Lincoln" AND "born"). Our current system has 5 weights. These rewrite weights are the only source of snippet weighting in our system. We explored how important these weight are and considered several other factors that could be used as additional sources of information for snippet weighting. Although we specify Boolean queries, the retrieval engine can provide a ranking, based on factors like link analyses, proximity of terms, 296 location of terms in the document, etc. So, different weights can be assigned to matches at different positions in the ranked list. We also looked at the number of matching terms in the best fixed width window, and the widow size of the smallest matching passage as indicators of passage quality. Rewrite Wts uses our heuristically determined rewrite weights as a measure the quality of a snippet. This is the current system default. Equal Wts gives equal weight to all snippets regardless of the rewrite rule that generated them. To the extent that more precise rewrites retrieve better answers, we will see a drop in performance when we make all weights equal. Rank Wts uses the rank of the snippet as a measure of its quality, SnippetWt = (100-rank )/100. NMatch Wts uses the number of matching terms in a fixed-width window as the measure of snippet quality. Length Wts uses a measure of the length of the snippet needed to encompass all query terms as the measure of snippet quality. We also look at combinations of these factors. For example, Rewrite+Rank Wts uses both rewrite weight and rank according to the following formula, SnippetWt = RewriteScore + (100-rank )/100. All of these measures are available from query-relevant summaries returned by the search engine and do not require analyzing the full text of the document. The results of these experiments are presented in Figure 4. Weighting MRR NumCorrect PropCorrect Equal Wts 0.489 298 0.596 Rewrite Wts (Default) 0.507 307 0.614 Rank Wts 0.483 292 0.584 Rewrite + Rank Wts 0.508 302 0.604 NMatch Wts 0.506 301 0.602 Length Wts 0.506 302 0.604 Figure 5: Snippet Weighting Our current default 5-level weighting scheme which reflects the specificity of the query rewrites does quite well. Equal weighting is somewhat worse, as we expected. Interestingly search engine rank is no better for weighting candidate n-grams than equal weighting. None of the other techniques we looked at surpasses the default weights in both MRR and PropCorrect. Our heuristic rewrite weights provide a simple and effective technique for snippet weighting, that can be used with any backend retrieval engine. Most question answering systems use IR-based measures of passage quality, and do not typically evaluate the best measure of similarity for purposes of extracting answers. Clarke et al. [9] noted above is an exception. Soubbotin and Soubbotin [18] mention different weights for different regular expression matches, but they did not describe the mechanism in detail nor did they evaluate how useful it is. Harabagiu et al. [11] have a kind of backoff strategy for matching which is similar to weighting, but again we do not know of parametric evaluations of its importance in their overall system performance. The question of what kinds of passages can best support answer mining for question answering as opposed to document retrieval is an interesting one that we are pursuing. DISCUSSION AND FUTURE DIRECTIONS The design of our question answering system was motivated by the goal of exploiting the large amounts of text data that is available on the Web and elsewhere as a useful resource. While many question answering systems take advantage of linguistic resources, fewer depend primarily on data. Vast amounts of data provide several sources of redundancy that our system capitalizes on. Answer redundancy (i.e., multiple, differently phrased, answer occurrences) enables us to use only simple query rewrites for matching, and facilitates the extraction of candidate answers. We evaluated the importance of redundancy in our system parametrically. First, we explored the relationship between the number of document snippets examined and question answering accuracy. Accuracy improves sharply as the number of snippets included for n-gram analysis increases from 1 to 50, somewhat more slowly after that peaking at 200 snippets, and then falls off somewhat after that. More is better up to a limit. We believe that we can increase this limit by improving our weight accumulation algorithm so that matches from the least precise rewrites do not dominate. Second, in smaller collections (like TREC), the accuracy of our system drops sharply, although it is still quite reasonable in absolute terms. Finally, snippet quality is less important to system performance than snippet quantity. We have a simple 5-level snippet weighting scheme based on the specificity of the query rewrite, and this appears to be sufficient. More complex weighting schemes that we explored were no more useful. The performance of our system shows promise for approaches to question answering which makes use of very large text databases even with minimal natural language processing. Our system does not need to maintain its own index nor does it require global term weights, so it can work in conjunction with any backend retrieval engine. Finally, since our system does only simple query transformations and n-gram analysis, it is efficient and scalable. One might think that our system has limited applicability, because it works best with large amounts of data. But, this isn't necessarily so. First, we actually perform reasonably well in the smaller TREC collection, and could perhaps tune system parameters to work even better in that environment. More interestingly, Brill et al. [6] described a projection technique that can be used to combine the wealth of data available on the Web with the reliability of data in smaller sources like TREC or an encyclopedia. The basic idea is to find candidate answers in a large and possibly noisy source, and then expand the query to include likely answers. The expanded queries can then be used on smaller but perhaps more reliable collections either directly to find support for the answer in the smaller corpus, or indirectly as a new query which is issued and mined as we currently do. This approach appears to be quite promising. Our approach seems least applicable in applications that involve a small amount of proprietary data. In these cases, one might need to do much more sophisticated analyses to map user queries to the exact lexical form that occur in the text collection rather than depend on primarily on redundancy as we have done. Although we have pushed the data-driven perspective, more sophisticated language analysis might help as well by providing more effective query rewrites or less noisy data for mining. 297 Most question answering systems contain aspects of both we use some linguistic knowledge in our small query typology and answer filtering, and more sophisticated systems often use simple pattern matching for things like dates, zip codes, etc. There are a number of open questions that we hope to explore. In the short term, we would like to look systematically at the contributions of other system components. Brill et al. [5] have started to explore individual components in more detail, with interesting results. In addition, it is likely that we have made several sub-optimal decisions in our initial implementation (e.g., omitting most stop words from answers, simple linear accumulation of scores over matching snippets) that we would like to improve. Most retrieval engines have been developed with the goal of finding topically relevant documents. Finding accurate answers may require somewhat different matching infrastructure. We are beginning to explore how best to generate snippets for use in answer mining. Finally, time is an interesting issue. We noted earlier how the correct answer to some queries changes over time. Time also has interesting implications for using redundancy. For example, it would take a while for a news or Web collection to correctly answer a question like "Who is the U. S. President?" just after an election. An important goal of our work is to get system designers to treat data as a first class resource that is widely available and exploitable. We have made good initial progress, and there are several interesting issues remaining to explore. REFERENCES [1] AAAI Spring Symposium Series (2002). Mining Answers from Text and Knowledge Bases. [2] S. Abney, M. Collins and A. Singhal (2000). Answer extraction. In Proceedings of the 6 th Applied Natural Language Processing Conference (ANLP 2000), 296-301. [3] ACL-EACL (2002). Workshop on Open-domain Question Answering. [4] E. Agichtein, S. Lawrence and L. Gravano (2001). Learning search engine specific query transformations for question answering. In Proceedings of the 10 th World Wide Web Conference (WWW10), 169-178. [5] E. Brill, S. Dumais and M. Banko (2002). An analysis of the AskMSR question-answering system. In Proceedings of 2002 Conference on Empirical Methods in Natural Language Processing (EMNLP 2002). [6] E. Brill, J. Lin, M. Banko, S. Dumais and A. Ng (2002). Data-intensive question answering. In Proceedings of the Tenth Text REtrieval Conference (TREC 2001). [7] S. Buchholz (2002). Using grammatical relations, answer frequencies and the World Wide Web for TREC question answering. In Proceedings of the Tenth Text REtrieval Conference (TREC 2001). [8] J. Chen, A. R. Diekema, M. D. Taffet, N. McCracken, N. E. Ozgencil, O. Yilmazel, E. D. Liddy (2002). Question answering: CNLP at the TREC-10 question answering track. In Proceedings of the Tenth Text REtrieval Conference (TREC 2001). [9] C. Clarke, G. Cormack and T. Lyman (2001). Exploiting redundancy in question answering. In Proceedings of the 24 th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR'2001), 358-365. [10] C. Clarke, G. Cormack and T. Lynam (2002). Web reinforced question answering. In Proceedings of the Tenth Text REtrieval Conference (TREC 2001). [11] S. Harabagiu, D. Moldovan, M. Pasca, R. Mihalcea, M. Surdeanu, R. Bunescu, R. Girju, V. Rus and P. Morarescu (2001). FALCON: Boosting knowledge for question answering. In Proceedings of the Ninth Text REtrieval Conference (TREC-9), 479-488. [12] E. Hovy, L. Gerber, U. Hermjakob, M. Junk and C. Lin (2001). Question answering in Webclopedia. In Proceedings of the Ninth Text REtrieval Conference (TREC-9 ), 655-664. [13] E. Hovy, U. Hermjakob and C. Lin (2002). The use of external knowledge in factoid QA. In Proceedings of the Tenth Text REtrieval Conference (TREC 2001). [14] C. Kwok, O. Etzioni and D. Weld (2001). Scaling question answering to the Web. In Proceedings of the 10 th World Wide Web Conference (WWW'10), 150-161. [15] M. A. Pasca and S. M. Harabagiu (2001). High performance question/answering. In Proceedings of the 24 th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR'2001), 366-374 . [16] J. Prager, E. Brown, A. Coden and D. Radev (2000). Question answering by predictive annotation. In Proceedings of the 23 rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR'2000), 184-191. [17] D. R. Radev, H. Qi, Z. Zheng, S. Blair-Goldensohn, Z. Zhang, W. Fan and J. Prager (2001). Mining the web for answers to natural language questions. In Proceeding of the 2001 ACM CIKM: Tenth International Conference on Information and Knowledge Management, 143-150 [18] M. M. Soubbotin and S. M. Soubbotin (2002). Patterns and potential answer expressions as clues to the right answers. In Proceedings of the Tenth Text REtrieval Conference (TREC 2001). [19] E. Voorhees and D. Harman, Eds. (2001). Proceedings of the Ninth Text REtrieval Conference (TREC-9). NIST Special Publication 500-249. [20] E. Voorhees and D. Harman, Eds. (2002). Proceedings of the Tenth Text REtrieval Conference (TREC 2001). ). NIST Special Publication 500-250. 298

rewrite query;n-gram extraction techniques;automatic QA;Experimentation;information extraction;Algorithms;question answering system;redundancy in large corpora;facilitates answer mining;natural language processing;information retrieval;machine learning;TREC QA;simple passage ranking

16

A Programming Languages Course for Freshmen

Programming languages are a part of the core of computer science. Courses on programming languages are typically offered to junior or senior students, and textbooks are based on this assumption. However, our computer science curriculum offers the programming languages course in the first year. This unusual situation led us to design it from an untypical approach. In this paper, we first analyze and classify proposals for the programming languages course into different pure and hybrid approaches. Then, we describe a course for freshmen based on four pure approaches, and justify the main choices made. Finally, we identify the software used for laboratories and outline our experience after teaching it for seven years.

INTRODUCTION The topic of programming languages is a part of the core of computer science. It played a relevant role in all the curricula recommendations delivered by the ACM or the IEEE-CS since the first Curriculum'68 [2]. Recent joint curricular recommendations of the ACM and the IEEE-CS identified several "areas" which structure the body of knowledge of the discipline. The list of areas has grown since the first proposal made by the Denning Report [4] up to 14 in the latest version, Computing Curricula 2001 [12]. Programming languages has always been one of these areas. Internationally reputed curricular recommendations are a valuable tool for the design of particular curricula. However, each country has specific features that constrain the way of organizing their studies. In Spain, the curriculum of a discipline offered by a university is the result of a trade-off. On the one hand, the university must at least offer a number of credits of the core subject matters established by the Government. On the other hand, the university may offer supplementary credits of the core as well as mandatory and optional courses defined according to the profile of the University and the faculty. Any proposal of a new curriculum follows a well-established process: (1) the curriculum is designed by a Center after consulting the departments involved; (2) it must be approved by the University Council; (3) the Universities Council of the Nation must deliver a (positive) report; and (4) the curriculum is published in the Official Bulletin of the Nation. This scheme has a number of advantages, e.g. a minimum degree of coherence among all the universities is guaranteed. However, it also has a number of disadvantages, e.g. the process to change a curriculum is very rigid. The Universidad Rey Juan Carlos is a young university, now seven years old. It offered studies of computer science since the very first year. The curriculum was designed by an external committee, so the teachers of computer science thereafter hired by the university did not have the opportunity to elaborate on it. The curriculum had a few weak points that would recommend a light reform, but the priorities of the new university postponed it. The curriculum establishes the features of the "Foundations of programming languages" course. The course is scheduled to last for fifteen weeks, with three lecture hours per week and two supervised laboratory hours per week. However, some flexibility is allowed, so that some weeks may be released from the lab component. This course is both a strong and a weak feature of the curriculum. It is a strong feature because programming languages are marginal in the official core. Consequently, our curriculum is closer to international recommendations than most Spanish universities. However, it is a weak feature, because the course is offered in the second semester of the first year! Notice that the programming languages course is more typically offered as an intermediate or advanced course in the third or fourth year. Our problem was how to teach the programming languages course to freshmen. The paper presents our design of the course and our experience. In the second section we first analyze and classify proposals for the programming languages course into different pure and hybrid approaches. In section 3, we describe a course for freshmen based on four pure approaches, and justify the choices made with respect to the factors that most influenced its design. Finally, we identify the software used for laboratories and outline our experience after teaching it for seven years. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. ITiCSE'05, June 2729, 2005, Monte de Caparica, Portugal. Copyright 2005 ACM 1-59593-024-8/05/0006...$5.00. 271 APPROACHES TO TEACHING PROGRAMMING LANGUAGES Since Curriculum'68, several issues on programming languages have received attention in the different curriculum recommendations: particular programming languages, language implementation, etc. It is formative to study (or to browse, at least) such recommendations, even though the large number of topics can be discouraging for the teacher. In this section, we try to organize different contributions. Firstly, we identify "pure" approaches to the programming languages course. Secondly, we describe their implementation, usually as hybrid approaches. Finally, we briefly discuss the issue of which programming languages and paradigms to use for the course. 2.1 Pure Approaches Probably, the best study on approaches to the programming languages course was given by King [7]. He made a study of 15 textbooks and found 10 different goals. Furthermore, he identified 3 approaches on which these textbooks were based and discussed a number of issues. We have extended his classification up to 5 approaches. Although most books and courses follow a hybrid approach, it is clarifying to distinguish the following pure ones: Catalogue approach. It provides a survey of several programming languages. This approach has several advantages: the student acquires a good education in several languages, it allows studying the interaction among different constructs of a language, and the languages may be studied in chronological order. However, it also exhibits disadvantages: there is much redundancy in studying similar features in different languages, and there is no guarantee that the student acquires a solid education. Descriptive (or anatomic [5]) approach. Programming languages have many common elements which can be grouped into categories and studied independently. Typical examples are: data types, hiding mechanisms, execution control, etc. The advantages and disadvantages of this approach are roughly the opposite of the previous one. Paradigm approach. Although each language has different characteristics and constructs, it is based on a basic model of computation called programming paradigm. The paradigm approach is an evolution of the descriptive approach described above since it generalizes language constructs and groups them consistently. Typical examples of paradigms are functional, logic and imperative programming. Formal approach. It studies the foundations of programming languages, mainly their syntax and semantics. The main advantage of this approach is that the student acquires a solid conceptual background. However, it has the risk of being too formal and therefore keeping far from the study of the programming languages themselves. Implementation approach. It comprises language processing topics. This approach is usually adopted jointly with the descriptive one, so that the run-time mechanisms that support each language construct are also described. This allows estimating the computational cost of each construct. However, the student may associate each concept with a particular implementation; this approach may also be in contradiction with the idea that a high level language should be understandable independently from its implementation. 2.2 Implementation of Pure Approaches The formal and implementation approaches are the basis of two well known and established courses: computation theory and language processors. They are not studied in this paper as standalone courses, but we consider their integration into the programming languages course. Despite these "pure" approaches, it is more common to adopt a hybrid one, formed by a combination of several approaches. For instance, we have explained that a descriptive course may address the implementation of language constructs. The descriptive and paradigm approaches also are commonly complemented by a small catalogue of selected languages. It is also common to find a descriptive part based on the imperative paradigm, followed by a second part based on the catalogue or the paradigm approaches. Finally, the formal approach can complement the descriptive or implementation ones. From a historical point of view, the catalogue approach was the most common in the first years of computing curricula. However, it has almost universally been abandoned, with some interesting exceptions, such as the experience by Maurer [9] on a subfamily of four C-like languages. The trend has been towards giving more importance to the foundations of programming languages, mainly elements and paradigms. After this evolution, it seems that the two most common organizations are: Descriptive approach, complemented with some languages or paradigms. It is the most common organization, according to currently available textbooks. Descriptive approach, illustrated by means of interpreters of some selected languages. Interpreters and paradigms can be combined in two symmetrical ways: either implementing an interpreter in each paradigm [14], or implementing an interpreter for one language of each paradigm; the latter approach can be adopted with an imperative language [6] or, more probably, with a functional language [1]. There are few proposals for a holistic approach. One exception is Wick and Stevenson's proposal [17], which combines the descriptive, formal, paradigm and implementation approaches into a "reductionistic" one. 2.3 Choice of Languages and Paradigms Courses on programming languages do not simply consist in the study of one or several languages, but their study usually is a part of the course. The selection of these languages rises some questions, that we cannot discuss here. A related issue is the selection of programming paradigms. Not all the paradigms can be equally useful for a course on programming languages for freshmen. Firstly, some paradigms are richer to illustrate language elements than others. Secondly, some paradigms can be more adequate to freshmen than others. There is not a catalogue of paradigms classified by suitable academic year, but it is important to use a more objective criterion than just the personal opinion of faculty. 272 We have used the Computing Curricula 2001 [12] as an objective basis to identify feasible paradigms. They identify the paradigms that have succeeded in CS1: procedural, object-oriented, functional, algorithmic notations, and low-level languages. The last two choices are useful for CS1 but not for a programming languages course, thus the remaining choices are procedural, object-oriented, and functional. An additional choice, not cited by Computing Curricula 2001, consists in the use of tiny languages [8][10]. These languages can be ad hoc designed for a specific domain or embedded into operating systems or applications. OUR PROPOSAL We discarded the catalogue approach because it would only contribute to a memorization effort by freshmen. Consequently, our course is based on the remaining four approaches: Formal approach. Formal grammars and syntax notations are given in depth. Language semantics is simply introduced. Implementation approach. Only basic concepts are given. Descriptive approach. Basic language elements are reviewed. Paradigm approach. A programming paradigm is given in depth (functional programming) but others are just sketched. Table 1 contains the contents of the course we offer. There are several major factors to consider for the design of a programming languages course. Firstly, the course when it is offered to students determines to a large extent the knowledge and maturity of students. Secondly, the existence of related courses, such as courses on automata theory or programming methodology, may recommend removing some overlapping topics. Thirdly, the specialization profile of faculty can foster the choice of a given topic instead of another one. Fourthly, a course with so many different topics must guarantee coherence among them. Finally, time constraints typically limit the number of topics to consider. In the following subsections we justify the adequacy of the choices made with respect to these factors. Table 1. Syllabus of our programming languages course PART I. INTRODUCTION Chapter 1. General issues Computer and programming languages. Elements, properties and history of programming languages. Classifications. PART II. SYNTACTIC FOUNDATIONS OF PROGRAMMING LANGUAGES Chapter 2. Grammars and formal languages Alphabets, symbols and chains. Languages. Grammars. Derivation of sentences. Recursive grammars. Classification of grammars: Chomsky's hierarchy. Abstract machines. Chapter 3. Regular grammars Definition. Uses and limitations. Regular expressions and regular languages. Finite-state automata. Chapter 4. Context-free grammars Definition. Uses and limitations. Parsing trees. The ambiguity problem and its removal. PART III. DESCRIPTION AND PROCESSING OF PROGRAMMING LANGUAGES Chapter 5. Language processors Abstract (or virtual) machines. Classes of processors. Stages in language processing. Concrete vs. abstract syntax. Chapter 6. Lexical and syntactic notations Lexical and syntactic elements. Regular definitions. Syntax notations: BNF, EBNF, syntax charts. Chapter 7. Semantics Semantics. Classes of semantics. Static semantics. Binding. PART IV. THE FUNCTIONAL PARADIGM Chapter 8. Basic elements Function definition and application. Programs and expressions. Overview of functional languages. Built-in types, values and operations. The conditional expression. Chapter 9. Advanced elements Operational semantics: term rewriting. Recursive functions. Local definitions. Chapter 10. Functional data types Constructors. Equations and patterns. Pattern matching. PART V. ELEMENTS OF PROGRAMMING LANGUAGES Chapter 11. Lexical and syntactical elements Identifiers. Numbers. Characters. Comments. Delimeters. Notations for expressions. Program structure and blocks. Chapter 12. Data types Recursive types. Parametric types: polymorphism. Polymorphic functions. Type systems. Type checking and inference. Type equivalence. Type conversion. Overloading. PART VI. PROGRAMMING PARADIGMS Chapter 13. Other paradigms and languages Imperative paradigms. Logic paradigm. Motivation of concurrency. Other computer languages: mark-up languages. 3.1 Maturity of Students One major concern was the fact that the course is offered to freshmen. A single approach could not be used because of the freshmen's lack of knowledge of programming languages. A variety of contents from the different approaches must be selected in order to give them a comprehensible and rich view of programming languages. The lack of maturity and capability of students to understand certain topics was a bottleneck for course organization. The different topics can be given with varying degrees of depth, but always making sure that freshmen can master them. We found that some topics are especially difficult to understand, even formulated in the simplest way. This mainly applies to: Semantics of programming languages. Implementation of programming languages. Some programming paradigms, such as concurrency. Consequently, these topics were included in a summarized way, so that students could achieve a global view of them and 273 understand the main issues involved. The rest of the topics could potentially be taught more deeply, but without forgetting that they were offered to freshmen. In terms of Bloom's taxonomy [2], the three topics above can be mastered at the knowledge level, or even comprehension level. However, for the rest of topics, we can expect students to achieve the application and analysis levels, at least. 3.2 Overlapping with Other Courses Some topics are also offered in other courses, either in the same or in a subsequent year. Consequently, these topics can be removed or dealt with more shallowly. The most probable conflicts are: Imperative programming, either procedural or object-oriented. Grammars and formal languages. Language processors. In our case, there is an annual course on programming methodology based on the imperative paradigm, but the other two topics are not included in the curriculum. The programming methodology course is offered in the first academic year. Consequently, we removed the imperative paradigm, except for its use in some illustrating examples, mainly in part V. However, we kept chapters on grammars and formal languages, and on language processors. 3.3 Preferences and Specialization of Faculty This factor is important in order to choose among equally eligible options, or to give broader coverage of some topics. In particular, faculty can be more familiar with some paradigms than with others. This was over-riding for our choice of the programming paradigm. In subsection 2.3 we discussed suitable paradigms for freshmen, and we concluded that Computing Curricula 2001 fosters the selection of the procedural, object-oriented, and functional paradigms. We discarded the procedural paradigm as it is concurrently taught in the programming methodology courses. Finally, our specialization gave priority to the functional paradigm over object-orientation. The reader can find many experiences in the literature, but we recommend a monograph on functional programming in education [13]. Functional programming is a paradigm with several advantages, such as short definition of languages, simple and concise programs, high level of abstraction, etc. However, its main advantage for us is richness of elements. This allows us to deal with many aspects of programming languages (e.g. data types, recursion, polymorphism, etc.) in a natural and easy way. The use of tiny languages is another attractive choice in a course for freshmen. However, we also discarded them in favor of the functional paradigm because they have fewer language elements. 3.4 Coherence and Unifying Themes A key issue in a course based on several approaches is to provide contents coherence. A network of relationships among the different parts makes possible their coherent integration. Part III (description and processing of languages) is the pragmatic continuation of part II (formal grammars). Thus, EBNF and syntactic charts are introduced in part III as more adequate notations for language description than pure grammar definitions. Language processing is given at a conceptual level, but the role of regular and syntax-free grammars in the architecture of language processors is highlighted. Parts IV and V are both based on a functional language, which is described with the tools given in part III, mainly EBNF and type constraints. Parts IV and V are also related because they are based on the same language. In order to provide more homogeneity, language elements studied in part V are introduced in a universal way, but they are mainly illustrated with the functional paradigm. Last, but not least, recursion is adopted as a recurring theme during the course. In effect, it is found in grammars, functions and data types. The recurrent presentation of this topic fosters deeper understanding by students. The "pure" definition of recursion is given early in the course, but its three instantiations enumerated above are studied later. For each instantiation, the mechanisms that accompany a recursive definition are clearly identified, in particular representation of information and operational semantics [16]. For instance, recursive grammars represent sentences as strings of terminal symbols, and its operating semantics is defined in terms of derivation of sentences. However, recursive functions represent information as expressions, and its operating semantics is defined in terms of term rewriting. 3.5 Time Constraints As a final factor, time constraints limit the depth of study of those topics that could be studied longer. A global view of the course schedule is given in Table 2. Table 2 Schedule of the course Part Theory #hours Lab #hours Part I 5 Part II 14 8 Part III 10 Part IV 12 6 Part V 8 6 Part VI 4 2 In part II (formal grammars), regular and context-free grammars are the only ones studied in depth because of their importance for language description and processing. Moreover, only one paradigm can be studied in depth. Even so, the lack of time limits the presentation of functional programming (parts IV and V) to the core elements of the paradigm. Other elements, important for the functional programmer, can not be addressed (e.g. higher-order, lazy evaluation, or currification). However, this is not a serious drawback since the aim of including functional programming in the course is teaching the essentials of a new paradigm as well as illustrating language elements. 274 LABORATORY COURSEWARE A course on programming languages must have a laboratory component. The laboratory schedule includes sessions for those parts of the course where problem solving can be faced, mainly formal grammars and functional programming. Laboratory tools were selected carefully so that they are adequate for freshmen to exercise non-trivial concepts; simple user interaction and visualization facilities are of great help here. There are a number of tools that fulfill these requirements. For formal grammars, we require simulators that allow at least manipulating regular expressions, finite automata, context-free grammars and derivation trees. Our final selection was JFLAP [11]. For functional programming, we require a programming environment that shows term rewriting as the operational semantics. Our final selection was WinHIPE [15]. EXPERIENCE We have been teaching this course for seven years. Although the basic structure has roughly been constant, it was refined according to our experience. In particular, the emphasis on recursion was introduced after several years as we noticed student problems with this concept. We consider that we have succeeded, at least in eliminating the magical connotation of recursion. A major change was the relative order of chapters on formal grammars and the functional paradigm. During the first year, they were given in reverse order. However, students had problems in understanding the syntax of functional declarations that led us to teach in the first place formal grammars (and therefore syntax notations such as EBNF). Thus, a foundation to declare syntax was laid and then used to introduce functional programming. The literature classifies the main difficulties for teaching functional programming into syntactical, conceptual and "psychological" problems [13]. In our approach, the two former kinds of problems are avoided, but the latter remains. As freshmen learn concurrently the functional and one imperative language, they get the idea that functional is an exotic, useless paradigm. CONCLUSION We have described a course on programming languages for freshmen. It comprises elements from four different approaches. We have described the contents of the course, and we have explained the factors that led us to its current design. The experience has been very positive both for teachers and for students. As the Denning report sought for CS1, we consider that our course illustrates that it is feasible to offer some traditionally intermediate or advanced matters in introductory courses. ACKNOWLEDGMENTS This work is supported by the research project TIN2004-07568 of the Spanish Ministry of Education and Science. REFERENCES [1] Abelson, H., and Sussman, G.J. Structure and Interpretation of Computer Programs. MIT Press, 2 ed., 1996. [2] Bloom, B., Furst, E., Hill, W., and Krathwohl, D.R. Taxonomy of Educational Objectives: Handbook I, The Cognitive Domain. Addison-Wesley, 1959. [3] Curriculum Committee on Computer Science. Curriculum '68: Recommendations for academic programs in computer science. Comm. ACM, 11, 3 (March 1968), 151-197. [4] Denning, P. et al. Computing as a Discipline. ACM Press, New York, 1988. [5] Fischer, A.E., and Grodzinsky, F.S. The Anatomy of Programming Languages. Prentice-Hall, 1993. [6] Kamin, S.N. Programming Languages: An Interpreter-Based Approach. Addison-Wesley, 1990. [7] King, K.N. The evolution of the programming languages course. In 23 rd SIGCSE Technical Symposium on Computer Science Education (SIGCSE'92). ACM Press, New York, 1992, 213-219. [8] Kolesar, M.V., and Allan, V.H. Teaching computer science concepts and problem solving with a spreadsheet. In 26 th SIGCSE Technical Symposium on Computer Science Education (SIGCSE'95). ACM Press, New York, 1995, 10-13 . [9] Maurer, W.D. The comparative programming languages course: A new chain of development. In 33 rd SIGCSE Technical Symposium on Computer Science Education (SIGCSE 2002). ACM Press, New York, 2002, 336-340. [10] Popyack, J.L., and Herrmann, N. Why everyone should know how to program a computer. In IFIP World Conference on Computers in Education VI (WCCE'95). Chapman & Hall, 1995, 603-612. [11] Hung, T., and Rodger, S.H. Increasing visualization and interaction in the automata theory course. In 31 st SIGCSE Technical Symposium on Computer Science Education (SIGCSE 2000). ACM Press, New York, 2000, 6-10. [12] The Joint Task Force on Computing Curricula IEEE-CS/ACM : Computing Curricula 2001 Computer Science, http://www.computer.org/education/cc2001/final, 2001. [13] Thomson, S., and Wadler, P. (eds.) Functional programming in education. Journal of Functional Programming, 3, 1 (1993). [14] Tucker, A.B., and Noonan, R.E. Integrating formal models into the programming languages course. In 33 rd SIGCSE Technical Symposium on Computer Science Education (SIGCSE 2002). ACM Press, New York, 2002, 346-350. [15] Velzquez-Iturbide, J.. Improving functional programming environments for education. In M. D. Brouwer-Hanse y T. Harrington (eds.), Man-Machine Communication for Educational Systems Design. Springer-Verlag, NATO ASI Series F 124, 1994, 325-332. [16] Velzquez-Iturbide, J.. Recursion in gradual steps (is recursion really that difficult?). In 31 st SIGCSE Technical Symposium on Computer Science Education (SIGCSE 2000). ACM Press, New York, 2000, 310-314. [17] Wick, M.R., and Stevenson, D.E. A reductionistic approach to a course on programming languages. In 32 nd SIGCSE Technical Symposium on Computer Science Education (SIGCSE 2001). ACM Press, New York, 2001, 253-257. 275

programming language course;language description;formal grammars;laboratory component;functional programming;computer science;programming methodology;Programming languages;recursion;curriculum;freshmen;topics;programming paradigms

160

Query Result Ranking over E-commerce Web Databases

To deal with the problem of too many results returned from an E-commerce Web database in response to a user query, this paper proposes a novel approach to rank the query results. Based on the user query, we speculate how much the user cares about each attribute and assign a corresponding weight to it. Then, for each tuple in the query result, each attribute value is assigned a score according to its "desirableness" to the user. These attribute value scores are combined according to the attribute weights to get a final ranking score for each tuple. Tuples with the top ranking scores are presented to the user first. Our ranking method is domain independent and requires no user feedback. Experimental results demonstrate that this ranking method can effectively capture a user's preferences.

INTRODUCTION With the rapid expansion of the World Wide Web, more and more Web databases are available. At the same time, the size of existing Web databases is growing rapidly. One common problem faced by Web users is that there is usually too many query results returned for a submitted query. For example, when a user submits a query to autos.yahoo.com to search for a used car within 50 miles of New York with a price between $5,000 and $10,000, 10,483 records are returned. In order to find "the best deal", the user has to go through this long list and compare the cars to each other, which is a tedious and time-consuming task. Most Web databases rank their query results in ascending or descending order according to a single attribute (e.g., sorted by date, sorted by price, etc.). However, many users probably consider multiple attributes simultaneously when judging the relevance or desirableness of a result. While some extensions to SQL allow the user to specify attribute weights according to their importance to him/her [21], [26], this approach is cumbersome and most likely hard to do for most users since they have no clear idea how to set appropriate weights for different attributes. Furthermore, the user-setting -weight approach is not applicable for categorical attributes. In this paper, we tackle the many-query-result problem for Web databases by proposing an automatic ranking method, QRRE (Query Result Ranking for E-commerce), which can rank the query results from an E-commerce Web database without any user feedback. We focus specifically on E-commerce Web databases because they comprise a large part of today's online databases. In addition, most E-commerce customers are ordinary users who may not know how to precisely express their interests by formulating database queries. The carDB Web database is used in the following examples to illustrate the intuitions on which QRRE is based. Example 1: Consider a used Car-selling Web database D with a single table carDB in which the car instances are stored as tuples with attributes: Make, Model, Year, Price, Mileage and Location. Each tuple t i in D represents a used car for sale. Given a tuple t i in the query result T q for a query q that is submitted by a buyer, we assign a ranking score to t i , based on its attribute values, which indicates its desirableness, from an E-commerce viewpoint, to the buyer. For instance, it is obvious that a luxury, new and cheap car is globally popular and desired in the used car market. However, sometimes the desired attribute values conflict with each other. For example, a new luxury car with low mileage is unlikely to be cheap. Hence, we need to decide which attributes are more important for a buyer. Some buyer may care more about the model of a car, while some other buyer may be more concerned about its price. For each attribute, we use a weight to denote its importance to the user. In this work, we assume that the attributes about which a user cares most are present in the query he/she submits, from which the attribute importance can be inferred. We define specified attributes to be attributes that are specified in a query and unspecified attributes to be attributes that are not specified in a query. Furthermore, we also consider that a subset of the unspecified attributes, namely, those attributes that are closely correlated to the query, is also important. Example 2: Given a query with condition "Year &gt; 2005", the query condition suggests that the user wants a relatively new car. Intuitively, besides the Year attribute, the user is more concerned about the Mileage than he/she is concerned about the Make and Location, considering that a relatively new car usually has low mileage. Given an unspecified attribute A i , the correlation between A i and the user query q is evaluated by the difference between the distribution of A i 's values over the query results and their distribution over the whole database D. The bigger the difference, the more A i correlates to the specified attribute value(s). Consequently, we assign a bigger attribute weight to A i . Example 3 explains our intuition. Example 3: Suppose a used car database D contains car instances for which the Year has values 1975 and onwards and D returns a subset d containing the tuples that satisfy the query with condition "Year &gt; 2005". Intuitively, Mileage values of the tuples in d distribute in a small and dense range with a relatively low average, while the Mileage values of tuples in D distribute in a large range with a relatively high average. The distribution difference shows a close correlation between the unspecified attribute, namely, Mileage, and the query "Year &gt; 2005". Besides the attribute weight, we also assign a preference score to each attribute value, including the values of both specified and unspecified attributes. In the E-commerce context, we first assume that an expensive product is less preferred than a cheap product if other attribute values are equal. Hence, we assign a small preference score for a high Price value and a large preference score for a low Price value. We further assume that a non-Price attribute value with high desirableness, such as a luxury car or a new car, correlates positively with a high Price value. Thus, a luxury car is more expensive than a standard car and a new car is usually more expensive than an old car. Based on this assumption, we convert a value a i of a non-Price attribute A i to a Price value p' I where p' I is the average price of the product for A i = a i in the database. Consequently, the preference score for a i will be large if p' I is large because a large Price value denotes a high desirableness for the user. Finally, the attribute weight and the value preference score are combined to get the final ranking score for each tuple. The tuples with the largest ranking scores are presented to the user first. The contributions of this paper include the following: 1. We present a novel approach to rank the tuples in the query results returned by E-commerce Web databases. 2. We propose a new attribute importance learning approach, which is domain independent and query adaptive. 3. We also propose a new attribute-value preference score assignment approach for E-commerce Web databases. In the entire ranking process, no user feedback is required. The rest of the paper is organized as follows. Section 2 reviews some related work. Section 3 gives a formal definition of the many-query -result problem and presents an overview of QRRE. Section 4 proposes our attribute importance learning approach while Section 5 presents our attribute preference score assignment approach. Experimental results are reported in Section 6. The paper is concluded in Section 7. RELATED WORK Query result ranking has been investigated in information retrieval for a long time. Cosine Similarity with TF-IDF weighting of the vector space model [2] and [26], the probabilistic ranking model [30] and [31] and the statistical language model [12] have been successfully used for ranking purposes. In addition, [10], [11], [14] and [15] explore the integration of database and information retrieval techniques to rank tuples with text attributes. [1], [5] and [17] propose some keyword-query based retrieval techniques for databases. However, most of these techniques focus on text attributes and it is very difficult to apply these techniques to rank tuples with categorical or numerical attributes. Some recent research addresses the problem of relational query result ranking. In [9], [26], [28] and [33], user relevance feedback is employed to learn the similarity between a result record and the query, which is used to rank the query results in relational multimedia databases. In [21] and [26], the SQL query language is extended to allow the user to specify the ranking function according to their preference for the attributes. In [18] and [19], users are required to build profiles so that the query result is ranked according to their profile. Compared with the above work, our approach is fully automatic and does not require user feedback or other human involvement. In [1] and [12], two ranking methods have been proposed that take advantage of the links (i.e., associations) among records, such as the citation information between papers. Unfortunately, linking information among records does not exist for most domains. The work that is most similar to ours is the probabilistic information retrieval (PIR) model in [8], which addresses the many-query-result problem in a probabilistic framework. In PIR, the ranking score is composed of two factors: global score, which captures the global importance of unspecified values, and conditional score, which captures the strength of the dependencies between specified and unspecified attribute values. The two scores are combined using a probabilistic approach. Our approach differs from that in [8] in the following aspects: 1. PIR only focuses on point queries, such as "A i = a i ". Hence, both a query with condition "Mileage &lt; 5000" and a query with condition "Mileage &lt; 2500" may have to be converted to a query with condition "Mileage = small" to be a valid query in PIR, which is not reasonable for many cases. In contrast, QRRE can handle both point and range queries. 2. PIR focuses on the unspecified attributes during query result ranking while QRRE deals with both specified and unspecified attributes. For example, suppose a car with price less than $10,000 is labeled as a "cheap" car. For a query "Price &lt; 10000", PIR will only consider the value difference for non-Price attributes among tuples and ignore the price difference, which is usually important for a poor buyer. On the contrary, QRRE will consider the value difference for all attributes. 3. A workload containing past user queries is required by PIR in order to learn the dependency between the specified and unspecified attribute values, which is unavailable for new online databases, while QRRE does not require such a workload. The experimental results in Section 6 show that QRRE produces a better quality ranking than does PIR. The attribute-importance learning problem was studied in [23] and [24], in which attribute importance is learned according to the attribute dependencies. In [23], a Bayesian network is built to discover the dependencies among attributes. The root attribute is the most important while the leaf attributes are less important. 576 In [24], an attribute dependency graph is built to discover the attribute dependencies. Both of these methods learn the attribute importance based on some pre-extracted data and their result is invariant to the user queries. Furthermore, both methods can only determine the attribute importance sequence. They are incapable of giving a specific value to show how important each attribute is. In contrast, the attribute-importance learning method presented in this paper can be adapted to the user's query and thus can be tailored to take into account the desire of different users, since each attribute is assigned a weight that denotes its importance for the user. To our knowledge, this is the first work that generates attribute weights that are adaptive to the query the user submitted. QUERY RESULT RANKING In this section, we first define the many-query-result problem and then present an overview of QRRE. 3.1 Problem Formulation Consider an autonomous Web database D with attributes A={A 1 , A 2 , ..., A m } and a selection query q over D with a conjunctive selection condition that may include point queries, such as "A i = a i ", or range queries, such as "a i1 &lt; A i &lt; a i2 ". Let T={t 1 , t 2 , ..., t n } be the set of result tuples returned by D for the query q. In many cases, if q is not a selective query, it will produce a large number of query results (i.e., a large T). The goal is to develop a ranking function to rank the tuples in T that captures the user's preference, is domain-independent and does not require any user feedback. 3.2 QRRE Initially, we focus on E-commerce Web databases because E-commerce Web databases comprise a large proportion of the databases on the Web. We further assume that each E-commerce Web database has a Price attribute, which we always assume to be A 1 . The Price attribute A 1 plays an intermediate role for all attributes during the attribute preference score assignment. Example 4: Consider the tuples in Table 1 that represent an example query result set T. It can be seen that most tuples have their own advantages when compared with other tuples. For example, t 1 is a relatively new car while t 2 is a luxury car and t 3 is the cheapest among all cars. Hence, depending on a user's preferences, different rankings may be needed for different users. Assuming that a user would prefer to pay the smallest amount for a car and that all other attribute values are equal, then the only certainty is that t 4 should always be ranked after t 3 because its mileage is higher than t 3 while it is more expensive than t 3 . Table 1. Examples of used car tuples. Year Make Model Mileage Price Location t 1 2005 Toyota Corolla 16995 26700 Seattle t 2 2002 Mercedes-Benz G500 47900 39825 Seattle t 3 2002 Nissan 350Z 26850 17448 Seattle t 4 2002 Nissan 350Z 26985 18128 Seattle According to Example 4, two problems need to be solved when we assign a ranking score for a tuple t i ={t i1 , t i2 , ..., t im } in the query result T: 1. How can we surmise how much a user cares about an attribute A j and how should we assign a suitable weight w j for the attribute(s) A j to reflect its (their) importance to the user? 2. How do we assign a preference score v ij for an attribute value t ij ? For example, when assigning the score for the attribute value "Year = 2005" in t 1 , should the score be larger than the score assigned for attribute value "Year = 2002" in t 2 and how much larger is reasonable? The first problem will be discussed in Section 4. The second problem will be discussed in Section 5. Having assigned a preference score v ij (1jm) to each attribute-value of t i and a weight w j to the attribute A j , the value preference scores v ij are summed to obtain the ranking score s i for t i to reflect the attribute importance for the user. That is: = = m j ij j i v w s 1 The overall architecture of a system employing QRRE is shown in Figure 1. Such a system includes two components: pre-processing component and online processing component. The pre-processing component collects statistics about the Web database D using a set Figure 1: Architecture of a system employing Query Result Ranking for E-commerce (QRRE). 577 of selected queries. Two kinds of histograms are built in the preprocessing step: single-attribute histograms and bi-attribute histograms. A single-attribute histogram is built for each attribute A j . A bi-attribute histogram is built for each non-Price attribute (i.e., A j in which i&gt;1) using the Price attribute A 1 . The online-processing component ranks the query results given the user query q. After getting the query results T from the Web database D for q, a weight is assigned for each attribute by comparing its data distribution in D and in the query results T. At the same time, the preference score for each attribute value in the query result is determined using the information from the bi-attribute histograms. The attribute weights and preference scores are combined to calculate the ranking score for each tuple in the query result. The tuples' ranking scores are sorted and the top K tuples with the largest ranking scores are presented to the user first. ATTRIBUTE WEIGHT ASSIGNMENT In the real world, different users have different preferences. Some people prefer luxury cars while some people care more about price than anything else. Hence, we need to surmise the user's preference when we make recommendations to the user as shown by Example 4 in Section 3. The difficulty of this problem lies in trying to determine what a user`s preference is (i.e., which attributes are more important) when no user feedback is provided. To address this problem, we start from the query the user submitted. We assume that the user's preference is reflected in the submitted query and, hence, we use the query as a hint for assigning weights to attributes. The following example provides the intuition for our attribute weight assignment approach. Example 5: Consider the query q with condition "Year &gt; 2005", which denotes that the user prefers a relatively new car. It is obvious that the specified attribute Year is important for the user. However, all the tuples in the query result T satisfy the query condition. Hence, we need to look beyond the specified attribute and speculate further about what the user's preferences may be from the specified attribute. Since the user is interested in cars that are made after 2005, we may speculate that the user cares about the Mileage of the car. Considering the distribution of Mileage values in the database, cars whose model year is greater than 2005 usually have a lower mileage when compared to all other cars. In contrast, attribute Location is less important for the user and its distribution in cars whose model year is greater than 2005 may be similar to the distribution in the entire database. According to this intuition, an attribute A j that correlates closely with the query will be assigned a large weight and vice verse. Furthermore, as Example 3 in Section 1 shows, the correlation of A j and the query can be measured by the data distribution difference of A j in D and in T. It should be noted that the specified attribute is not always important, especially when the condition for the specified attribute is not selective. For example, for a query with condition "Year &gt; 1995 and Make = BMW", the specified attribute Year is not important because almost all tuples in the database satisfy the condition "Year &gt; 1995" and the Year distribution in D and in T is similar. A natural measure of the distribution difference of A j in D and in T is the Kullback-Leibler distance or Kullback-Leibler (KL) divergence [13]. Suppose that A j is a categorical attribute with value set {a j1 , a j2 , ..., a jk }. Then the KL-divergence of A j from D to T is: = = = = = k l jl j jl j jl j KL T a A prob D a A prob D a A prob T D D 1 ) | ( ) | ( log ) | ( ) || ( (1) in which prob(A j =a jl | D) refers to the probability that A j = a jl in D and prob(A j =a jl | T) refers to the probability that A j = a jl in T. If A j is a numerical attribute, its value range is first discretized into a few value sets, where each set refers to a category, and then the KL-divergence of A j is calculated as in (1). 4.1 Histogram Construction To calculate the KL-divergence in equation (1) we need to obtain the distribution of attribute values over D. The difficulty here is that we are dealing with an autonomous database and we do not have full access to all the data. In [24], the attribute value distribution over a collection of data crawled from D is used to estimate the actual attribute value distribution over D. However, it is highly likely that the distribution of the crawled data can be different from that of D because the crawled data may be biased to the submitted queries. In this paper, we propose a probing-and-count based method to build a histogram for an attribute over a Web database 1 . We assume that the number of query results is available in D for a given query. After submitting a set of selected queries to D, we can extract the number of query results, instead of the actual query results, to get the attribute value distribution of A i . An equi-depth histogram [27] is used to represent the attribute value distribution, from which we will get the probability required in Equation (1). The key problem in our histogram construction for A i is how to generate a set of suitable queries to probe D. Figure 2 shows the algorithm for building a histogram for attribute A i . For each attribute A i , a histogram is built in the preprocessing stage. We assume that one attribute value of A i is enough to be a query for D. If A i is a categorical attribute, each category of A i is used as a query to get its occurrence count (Lines 2-3). If A i is a numerical attribute, an equal-depth histogram is built for A i . We first decide the occurrence frequency threshold t for each bucket by dividing |D|, namely, the number of tuples in D, with the minimum bucket number n that will be created for a numerical attribute A i . In our experiments, n is empirically set to be 20. Then we probe D using a query with condition on A i such that low A i &lt;up and get c, the number of instances in that range (Line 8). If c is smaller than t, a bucket is added for it in H Di (Line 10) and another query probe is prepared (Line 11). Otherwise, we update the query probe condition on A i by reducing the size of the bucket (Line 13) and a new iteration begins. The iteration continues until each value in the value range is in a bucket. It is obvious that there are some improvements that can be made to the algorithm to accelerate the histogram construction. The improvements are not described here because histogram construction is not the major focus of this paper. Considering that only a single-attribute histogram is constructed, the process should complete quickly. 1 Although both our histogram construction method and the histogram construction methods in [1] and [5] are probing-based , they have different goals. The goal in [1] and [5] is to build a histogram that precisely describes the regions on which the queries concentrate, while our purpose is to build a histogram that summarizes the data distribution of D as precisely as possible with a number of query probes. 578 A histogram H Ti also needs to be built for A i over T (the result set) to get its probability distribution over T. For each bucket of H Di , a bucket with the same bucket boundary is built in H Ti while its frequency is counted in T. 4.2 Attribute Weight Setting After getting the histogram of A i over D and T, the histograms are converted to a probability distribution by dividing the frequency in each bucket of the histogram by the bucket frequency sum of the histogram. That is, the probability distribution of A i for D, P Di , is | | | | D c p Dk Di = in which c Dk is the frequency of the k th bucket in H Di . The probability distribution of A i for T, P Ti , is | | | | T c p Tk Ti = in which c Tk is the frequency of the k th bucket in H Ti . Next, for the i th attribute A i , we assign its importance w i as = = m j Tj Dj Ti Di i P P KL P P KL w 1 ) , ( ) , ( The attribute weight assignment is performed not only on the unspecified attributes, but also on the specified attributes. If a specified attribute is a point condition, its attribute weight will be the same for all tuples in the query result. If a specified attribute is a range condition, its attribute weight will be different for the tuples in the query result. Example 6 illustrates this point. Example 6: Consider a query q with condition "Make = 2004 and Price &lt; 10000". In q, since the specified attribute Make is a point attribute, the attribute weight assigned to it is useless because all the query results have the same value for Make. On the other hand, since the attribute Price is a range attribute, the price of different tuples is an important factor to consider during ranking. 4.3 Examples of Attribute Weight Assignment In our experiments, we found that the attribute weight assignment was intuitive and reasonable for the given queries. Table 2 shows the attribute weight assigned to different attributes corresponding to different queries in our experiments for the carDB. Given a query with condition "Mileage &lt; 20000", which means that the user prefers a new car, as expected the attribute "Mileage" is assigned a large weight because it is a specified attribute and the attribute "Year" is assigned a large weight too. The attribute "Model" is assigned a large weight because a new car usually has a model that appears recently. In contrast, Consider the query with condition "Make = BMW & Mileage &lt; 100000". The sub-condition "Mileage &lt; 100000" possesses a very weak selective capability because almost all tuples in the database satisfy it. The buyer is actually just concerned about the Make and the Model of the car. As expected, the attribute Make and Model are assigned large weights, while Year and Mileage are no longer assigned large weights. Table 2: Attribute weight assignments for two queries. Mileage &lt; 20000 Make = BMW & Mileage &lt; 100000 Year 0.222 0.015 Make 0.017 0.408 Model 0.181 0.408 Price 0.045 0.120 Mileage 0.533 0.04 Location 0.0003 0.002 ATTRIBUTE PREFERENCE SCORE ASSIGNMENT In addition to the attributes themselves, different values of an attribute may have different attractions for the buyer. For example, a car with a low price is obviously more attractive than a more expensive one if other attribute values are the same. Similarly, a car with low mileage is also obviously more desirable. Given an attribute value, the goal of the attribute preference score assignment module is to assign a preference score to it that reflects its desirableness for the buyer. To facilitate the combination of scores of different attribute values, all scores assigned for different attribute values are in [0, 1]. Instead of requiring human involvement for attribute value assignment, given a normal E-commerce context, we make the following two intuitive assumptions: 1. Price assumption: A product with a lower price is always more desired by buyers than a product with a higher price if the other attributes of the two products have the same values. For example, if all other attribute values are the same, a cheaper car is preferred over a more expensive car. 2. Non-Price assumption: A non-Price attribute value with higher desirableness for the user corresponds to a higher price. For example, a new car, which most buyers prefer, is usually more Input: Attribute A i and its value range Web database D with the total number of tuples | D | Minimum bucket number n Output: A single-attribute histogram H Di for A i Method: 1. If A i is a categorical attribute 2. For each category a ij of A i , probe D using a query with condition "A i =a ij " and get its occurrence count c 3. Add a bucket (a ij , c) into H D i 4. If A i is a numerical value attribute with value range [a low , a up ) 5. t = |D| / n 6. low = a low , up = a up 7. Do 8. probe D with a query with condition "low A i &lt;up" and get its occurrence count c 9. if c t 10. Add a bucket (low, up, c) into H D i 11. low = up, up = a up 12. else 13. up = low + (up - low) / 2 14. While low &lt; a up 15. Return H D i Figure 2: Probing-based histogram construction algorithm. 579 expensive than an old car. Likewise, a luxury car is usually more expensive than an ordinary car. With the above two assumptions, we divide the attributes into two sets: Price attribute set, which only includes the attribute Price, and non-Price attribute set, which includes all attributes except Price. The two sets of attributes are handled in different ways. According to the Price assumption, we assign a large score for a low price and a small score for a high price. To avoid requiring human involvement to assign a suitable score for a Price value, the Price distribution in D is used to assign the scores. Given a Price value t, a score v t is assigned to it as the percentage of tuples whose Price value is bigger than t i in D: | | D S v t t = in which S t denotes the number of tuples whose Price value is bigger than t. In our experiments, the histogram for the attribute Price A 1 , whose construction method is described in Section 4, is used for the Price preference score assignment. Figure 3 shows the algorithm used to assign a score v for a Price value t using the Price histogram. Given the Price histogram H D1 , the frequency sum is fist calculated (Line 1). Then we count the number S t of tuples whose Price value is bigger than t. For each bucket in H D1 , if the lower boundary of the bucket is bigger than t, it means that all the tuples for this bucket have a Price value bigger than t and the frequency of this bucket is added to S t (Line 4). If t is within the boundary of the bucket, we assume that the Price has a uniform distribution in the bucket and a fraction of the frequency in this bucket is added to S t (Line 5). If the upper boundary of the bucket is smaller than t, it means that all the tuples for this bucket have a Price value lower than t and the bucket is ignored. Finally the ratio is generated by dividing S t with the frequency sum. For a value a i of a non-Price attribute A i , the difficulty of assigning it a score is two fold: 1. How to make the attribute preference score assignment adaptive for different attributes? Our goal is to have an intuitive assignment for each attribute without human involvement. The difficulty is that different attributes can have totally different attribute values. 2. How to establish the correspondence between different attributes? For example, how can we know that the desirableness of "Year = 2005" is the same as the desirableness of "Mileage = 5000" for most users? We solve the problem in two steps. First, based on the non-Price assumption, we can convert a non-Price value a i to a Price attribute value t i : If A i is a categorical attribute, t i is the average price for all tuples in D such that A i =a i . If A i is a numerical attribute, v i is the average price for all tuples in D such that a i -d &lt; A i &lt; a i +d where d is used to prevent too few tuples or no tuple being collected if we just simply set A i =a i . In our experiments, a bi-attribute histogram (A 1 , A i ) is used when a i is converted to a Price value. The bi-attribute histograms are built in the pre-processing step in a way similar to the histogram construction described in Section 4. Second, after converting all non-Price attribute values to Price values, we use a uniform mechanism to assign them a preference score. We assign a large score for a large Price value according to the non-Price assumption. That is, given a converted Price value t i , a preference score v i is assigned to it as the percentage of Price values that is smaller than t i in D. The algorithm for the converted Price preference score assignment can be easily adapted from the algorithm in Figure 3. 5.1 Examples of Attribute Preference Score Assignment Table 3 shows the average Price and assigned score for different Make values for the carDB database used in our experiments. It can be seen that the prices for different car makes fit our intuition well. Luxury cars are evaluated to have a higher price than standard cars and, consequently, are assigned a larger preference score. We found that the attribute preference assignments for other attributes in carDB are intuitive too. Table 3: Make-Price-Score correspondence. Make Average Price Score Mitsubishi 12899 0.183 Volkswagen 16001 0.372 Honda 16175 0.373 Toyota 16585 0.387 Acura 20875 0.599 BMW 33596 0.893 Benz 37930 0.923 EXPERIMENTS In this section, we describe our experiments, report the QRRE experimental results and compare them with some related work. We first introduce the databases we used and the related work for comparison. Then we informally give some examples of query result ranking to provide some intuition for our experiments. Next, a more formal evaluation of the ranking results is presented. Finally, the running time statistics are presented. 6.1 Experimental Setup To evaluate how well different ranking approaches capture a user's preference, five postgraduate students were invited to participate in the experiments and behave as buyers from the E-commerce databases. Input: Price histogram H D1 ={(c 1 ,low 1 ,up 1 ), ..., (c m , low m , up m )} Price value t Output: Price score v Method: 1. = i i c sum 2. S t = 0 3. For i =1..m 4. if (low m &gt; t ) S t = S t + c i 5. if (low m &lt; t &lt; up m ) S t =S t + c i * (t -low m )/( up m - low m ) 6. v = S t / sum 7. return v Figure 3: Price value score assignment algorithm. 580 6.1.1 Databases For our evaluation, we set up two databases from two domains in E-commerce . The first database is a used car database carDB(Make, Model, Year, Price, Mileage, Location) containing 100,000 tuples extracted from Yahoo! Autos. The attributes Make, Model, Year and Location are categorical attributes and the attributes Price and Mileage are numerical attributes. The second database is a real estate database houseDB(City, Location, Bedrooms, Bathrooms, Sq Ft, Price) containing 20,000 tuples extracted from Yahoo! Real Estate. The attributes City, Location, Bedrooms and Bathrooms are categorical attributes and the attributes Sq Ft and Price are numerical attributes. To simulate the Web databases for our experiments we used MySQL on a P4 3.2-GHz PC with 1GB of RAM . We implemented all algorithms in JAVA and connected to the RDBMS by DAO. 6.1.2 Implemented Algorithms Besides QRRE described above, we implemented two other ranking methods, which are described briefly below, to compare with QRRE. RANDOM ranking model: In the RANDOM ranking model, the tuples in the query result are presented to the user in a random order. The RANDOM model provides a baseline to show how well QRRE can capture the user behavior over a random method. Probabilistic Information Retrieval (PIR) ranking model: A probabilistic information retrieval (PIR) technique, which has been successfully used in the Information Retrieval field, is used in [8] for ranking query results. This technique addresses the same problem as does QRRE. In PIR, given a tuple t, its ranking score is given by the following equation: = Y y X x Y y D y x p W y x p D y p W y p t Score ) , | ( ) , | ( ) | ( ) | ( ) ( in which X is the specified attributes, Y is the unspecified attributes, W is a past query workload and p denotes the probability. As mentioned in Section 2, PIR work focuses on point queries without considering range queries. Therefore, when applying the PIR ranking model, the numerical attributes Price and Mileage in carDB and Sq Ft and Price in houseDB are discretized into meaningful ranges as categories, which in reality requires a domain expert. In PIR, a workload is required to obtain the conditional probability used to measure the correlation between specified attribute values present in the query and the unspecified attributes. In our experiments, we requested 5 subjects to behave as different kinds of buyers, such as rich people, clerks, students, women, etc. and post queries against the databases. We collected 200 queries for each database and these queries are used as the workload W for the PIR model. 6.2 Examples of Query Result Ranking When we examine the query result rankings, we find that the ranking results of both QRRE and PIR are much more reasonable and intuitive than that of RANDOM. However, there are some interesting examples that show that the QRRE rankings are superior to those of PIR. We found that the ranking result of QRRE is more reasonable than that of PIR in several ways: QRRE can discover an assumption that is implicitly held by a buyer. For example, for a query with condition "Mileage &lt; 5000". QRRE ranks cars with Year = 2006 as the top recommendation. Intuitively, this is because a 2006 model year car usually has lower mileage and this is what the user is looking for. However, PIR is unable to identify the importance of Year because most users assume that Mileage itself is enough to represent their preference and, consequently, the relationship between Year and Mileage is not reflected in the workload. In PIR, given a numerical attribute, its value range needs to be discretized into meaningful categories and the values within a category are assumed to be the same during ranking. For example, if we assign a car with "Mileage &lt; 10000" to be a category "Mileage = small", then PIR will treat "Mileage = 2000" to be the same as "Mileage = 9000", which is obviously unreasonable. In contrast, QRRE will identify that "Mileage = 2000" is more desirable than "Mileage = 9000". QRRE considers the value difference of the specified attributes of the tuples in the query result, while PIR ignores the difference. For example, for a query with condition "Make = Mercedes-Benz and Model = ML500 and Year &gt; 2003", QRRE usually ranks the cars that are made in this year first. However, PIR does not take the Year difference in the query result records into consideration during ranking. Likewise, QRRE often produces a ranking better than does PIR for houseDB. The actual evaluation in the following section confirms these observations. 6.3 Ranking Evaluation We now present a more formal evaluation of the query result ranking quality. A survey is conducted to show how well each ranking algorithm captures the user's preference. We evaluate the query results in two ways: average precision and user preference ranking. 6.3.1 Average Precision In this experiment, each subject was asked to submit three queries for carDB and one query for houseDB according to their preference. Each query had on average 2.2 specified attributes for carDB and 2.4 specified attributes for houseDB. We found that all the attributes of carDB and houseDB were specified in the collected queries at least once. On average, for carDB, each query had a query result of 686 tuples, with the maximum being 4,213 tuples and the minimum 116 tuples. It can be seen that the many-query-result problem is a common problem in reality. Each query for houseDB has a query result of 166 tuples on average. Since it is not practical to ask the subjects to rank the whole query result for a query, we adopt the following strategy to compare the performance of different ranking approaches. For each implemented ranking algorithm, we collected the first 10 tuples that it recommended. Hence, thirty tuples are collected in total. If there is overlap among the recommended tuples from different algorithms, we extract more tuples using the RANDOM algorithm so that thirty unique tuples are collected in total. Next, for each of the fifteen queries, each subject was asked to rank the top 10 tuples as the relevant tuples that they preferred most from the thirty unique tuples collected for each query. During ranking, they were asked to behave like real buyers to rank the records according to their preferences. 581 Table 4: Average precision for different ranking methods for carDB. QRRE PIR RANDOM q1 0.72 0.52 0.08 q2 0.62 0.62 0.06 q3 0.72 0.22 0.06 q4 0.52 0.64 0.04 q5 0.84 0.78 0.06 q6 0.68 0.36 0.04 q7 0.92 0.46 0.02 q8 0.88 0.64 0.06 q9 0.78 0.62 0.04 q10 0.74 0.64 0.04 q11 0.56 0.66 0.06 q12 0.86 0.76 0.08 q13 0.84 0.36 0.02 q14 0.58 0.38 0.04 q15 0.76 0.66 0.06 Average 0.735 0.555 0.048 We use the Precision/Recall metrics to evaluate how well the user's preference is captured by the different ranking algorithms. Precision is the ratio obtained by dividing the number of retrieved tuples that are relevant by the total number of retrieved tuples. Recall is the ratio obtained by dividing the number of relevant tuples by the number of tuples that are retrieved. In our experiments, both the relevant tuples and the retrieved tuples are 10, which make the Precision and Recall to be equal. Table 4 shows the average precision of the different ranking methods for each query. It can be seen that both QRRE and PIR consistently have a higher precision than RANDOM. For 11 queries out of 15, the precision of QRRE is higher than that of PIR. The precision of QRRE is equal to that of PIR for two queries and is lower than that of PIR for the remaining two queries. QRRE's average precision is 0.18 higher than that of PIR. QRRE has a precision higher than 0.5 for each query while PIR has a precision as low as 0.22 for q3. It should be noted that there is some overlap between the top-10 ranked results of QRRE and top-10 ranked results of PIR for most queries. Figure 4 and Figure 5 show the average precision of the three ranking methods graphically for both carDB and houseDB. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Query A ver ag e P r eci s i o n QRRE PIR RANDOM Figure 4: Average prevision for different ranking methods for carDB. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 2 3 4 5 Query A v er ag e P r e c i s i o n QRRE PIR RANDOM Figure 5: Average precision for different ranking methods for houseDB. 6.3.2 User Preference Ranking In this experiment, 10 queries were collected from the 5 subjects for carDB and 5 queries were collected for houseDB. After getting the query results, they were ranked using the three ranking methods. The ranking results were then provided to the subjects in order to let them select which result they liked best. Table 5 and Table 6 show the user preference ranking (UPR) of the different ranking methods for each query for carDB and houseDB, respectively. It can be seen that again both QRRE and PIR greatly outperform RANDOM. In most cases, the subjects preferred the ranking results of QRRE to that of PIR. Table 5: UPR for different ranking methods for carDB. QRRE PIR RANDOM q1 0.8 0.2 0 q2 1 0 0 q3 0.6 0.4 0 q4 0.4 0.6 0 q5 0.4 0.6 0 q6 0.8 0.2 0 q7 1 0 0 q8 0.6 0.2 0.2 q9 0.8 0.2 0 q10 0.8 0.2 0 Average 0.72 0.26 0.02 Table 6: UPR for different ranking methods for houseDB. QRRE PIR RANDOM q1 0.4 0.4 0.2 q2 0.8 0.2 0 q3 1 0 0 q4 0.4 0.6 0 q5 0.6 0.4 0 Average 0.66 0.32 0.02 582 While these preliminary experiments indicate that QRRE is promising and better than the existing work, a much larger scale user study is necessary to conclusively establish this finding. 6.4 Performance Report Using QRRE, histograms need to be constructed before the query results can be ranked. The histogram construction time depends on the number of buckets and the time to query the web to get the number of occurrences for each bucket. However, in most cases the histogram usually does not change very much over time and so needs to be constructed only once in a given time period. The query result ranking in the online processing part includes four modules: the attribute weight assignment module, the attribute-value preference score assignment module, the ranking score calculation module and the ranking score sorting module. Each of the first three modules has a time complexity of O(n) after constructing the histogram, where n is the number of query results, and the ranking score sorting module has a time complexity of O(nlog(n)). Hence, the overall time complexity for the online processing stage is O(nlog(n)). Figure 6 shows the online execution time of the queries over carDB as a function of the number of tuples in the query result. It can be seen that the execution time of QRRE grows almost linearly with the number of tuples in the query result. This is because ranking score sorting is fairly quick even for a large amount of data and thus most of the running time is spent in the first three modules. 0 5000 10000 15000 20000 25000 30000 35000 40000 45000 50000 0 500 1000 1500 2000 2500 3000 Tuple Number E x e c ut i on Ti m e ( m s ) Figure 6: Execution times for different numbers of query results for carDB. CONCLUSION In this paper, a novel automated ranking approach for the many-query -result problem in E-commerce is proposed. Starting from the user query, we assume that the specified attributes are important for the user. We also assume that the attributes that are highly correlated with the query also are important to the user. We assign a weight for each attribute according to its importance to the user. Then, for each value in each tuple of the query result, a preference score is assigned according to its desirableness in the E-commerce context, where users are assumed to more prefer products with lower prices. All preference scores are combined according to the attribute weight assigned to each attribute. No domain knowledge or user feedback is required in the whole process. Preliminary experimental results indicate that QRRE captures the user preference fairly well and better than existing works. We acknowledge the following shortcoming of our approach, which will be the focus for our future research. First, we do not deal with string attributes, such as book titles or the comments for a house, contained in many Web databases. It would be extremely useful to find a method to incorporate string attributes into QRRE. Second, QRRE has only been evaluated on small-scale datasets. We realize that a large, comprehensive benchmark should be built to extensively evaluate a query result ranking system, both for QRRE and for future research. Finally, QRRE has been specifically tailored for E-commerce Web databases. It would be interesting to extend QRRE to also deal with non-E-commerce Web databases. ACKNOWLEDGMENTS This research was supported by the Research Grants Council of Hong Kong under grant HKUST6172/04E. REFERENCES [1] A. Aboulnaga and S. Chaudhuri. "Self-tuning Histograms: Building Histograms Without Looking at Data," Proc. of the ACM SIGMOD Conf., 181-192, 1999. [2] S. Agrawal, S. Chaudhuri and G. Das. "DBXplorer: A System for Keyword Based Search over Relational Databases," Proc. of 18 th Intl. Conf. on Data Engineering, 5-16, 2002. [3] R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval, Addison-Wesley, 1999. [4] A. Balmin, V. Hristidis and Y. Papakonstantinou. "ObjectRank: Authority-Based Keyword Search in Databases," Proc. of the 30 th Intl. Conf. on Very Large Databases, 564-575, 2004. [5] G. Bhalotia, C. Nakhe, A. Hulgeri, S. Chakrabarti and S. Sudarshan. "Keyword Searching and Browsing in Databases using BANKS," Proc. of 18 th Intl. Conf. on Data Engineering, 431-440, 2002. [6] N. Bruno, S. Chaudhuri and L. Gravano. "STHoles: A Multidimensional Workload-aware Histogram," Proc. of the ACM SIGMOD Conf., 211-222, 2001. [7] K. Chakrabarti, S. Chaudhuri and S. Hwang. "Automatic Categorization of Query Results," Proc. of the ACM SIGMOD Conf., 755-766, 2004. [8] S. Chaudhuri, G. Das, V. Hristidis and G. Weikum. "Probabilistic Ranking of Database Query Results," Proc. of the Intl. Conf. on Very Large Databases, 888-899, 2004. [9] K. Chakrabarti, K. Porkaew and S. Mehrotra. "Efficient Query Refinement in Multimedia Databases," Proc. of 16 th Intl. Conf. on Data Engineering, 196, 2000. [10] W. Cohen. "Integration of Heterogeneous Databases Without Common Domains Using Queries Based on Textual Similarity," Proc. of the ACM SIGMOD Conf., 201-212, 1998. [11] W. Cohen. "Providing Database-like Access to the Web Using Queries Based on Textual Similarity," Proc. of the ACM SIGMOD Conf., 558-560,1998. [12] W.B. Croft and J. Lafferty. Language Modeling for Information Retrieval. Kluwer 2003. [13] R.O. Duda, P.E. Hart and D.G. Stork, Pattern Classification. John Wiley & Sons, USA, 2001. [14] N. Fuhr. "A Probabilistic Framework for Vague Queries and Imprecise Information in Databases," Proc. of the 16 th Intl. Conf. on Very Large Databases, 696-707, 1990. 583 [15] N. Fuhr. "A Probabilistic Relational Model for the Integration of IR and Databases," Proc. of the ACM SIGIR Conf., 309-317, 1993. [16] F. Geerts, H. Mannila and E. Terzi. "Relational Link-based Ranking," Proc. of the 30 th Intl. Conf. on Very Large Databases, 552-563, 2004. [17] V. Hristidis and Y. Papakonstantinou. "DISCOVER: Keyword Search in Relational Databases," Proc. of the 28 th Intl. Conf. on Very Large Databases, 670-681, 2002. [18] G. Koutrika and Y.E. Ioannidis. "Personalization of Queries in Database Systems," Proc. of 20 th Intl. Conf. on Data Engineering, 597-608, 2004. [19] G. Koutrika and Y.E. Ioannidis. "Constrained Optimalities in Query Personalization," Proc. of the ACM SIGMOD Conf., 73-84 , 2005. [20] Y.E. Ioannidis. "The History of Histograms (abridged)," Proc. of the 29 th Intl. Conf. on Very Large Databases, 19-30, 2003. [21] W. Kieling. "Foundations of Preferences in Database Systems," Proc. of the 28 th Intl. Conf. on Very Large Databases, 311-322, 2002. [22] R. Kooi. The Optimization of Queries in Relational Databases. PhD Thesis, Case Western Reserve University, 1980. [23] I. Muslea and T. Lee. "Online Query Relaxation via Bayesian Causal Structures Discovery," Proc. of the AAAI Conf., 831-836 , 2005. [24] U. Nambiar and S. Kambhampati. "Answering Imprecise Queries over Autonomous Web Databases," Proc. of 22 nd Intl. Conf. on Data Engineering, 45, 2006. [25] Z. Nazeri, E. Bloedorn and P. Ostwald. "Experiences in Mining Aviation Safety Data," Proc. of the ACM SIGMOD Conf., 562-566, 2001. [26] M. Ortega-Binderberger, K. Chakrabarti and S. Mehrotra. "An Approach to Integrating Query Refinement in SQL," Proc. Intl. Conf. on Extending Data Base Technology, 15-33, 2002. [27] G. Piatetsky-Sharpiro and C. Connell. "Accurate Estimation of the Number of Tuples Satisfying a Condition," Proc. of the ACM SIGMOD Conf., 256--276, 1984. [28] Y. Rui, T.S. Huang and S. Merhotra. "Content-Based Image Retrieval with Relevance Feedback in MARS," Proc. IEEE Intl. Conf. on Image Processing, 815-818, 1997. [29] G. Salton, A. Wong and C.S. Yang. "A Vector Space Model for Information Retrieval," Communications of the ACM 18(11), 613-620, 1975. [30] K. Sparck Jones, S. Walker and S.E. Robertson. "A Probabilistic Model of Information Retrieval: Development and Comparative Experiments - Part 1," Inf. Process. Management 36(6), 779-808, 2000. [31] K. Sparck Jones, S. Walker and S.E. Robertson. "A Probabilistic Model of Information Retrieval: Development and Comparative Experiments - Part 2," Inf. Process. Management 36(6), 809-840, 2000. [32] E.M. Voorhees. "The TREC-8 Question Answering Track Report," Proc. of the 8 th Text Retrieval Conf, 1999. [33] L. Wu, C. Faloutsos, K. Sycara and T. Payne. "FALCON: Feedback Adaptive Loop for Content-Based Retrieval," Proc. of the 26 th Intl. Conf. on Very Large Databases, 297-306, 2000. 584

many query result problem;rank the query results;query result ranking;QRRE;algorithms;experimentation;attribute value;Attribute weight assignment;Query result ranking;attribute preference;design;PIR;e-commerce web databases;human factors;E-commerce

161

Query Type Classification for Web Document Retrieval

The heterogeneous Web exacerbates IR problems and short user queries make them worse. The contents of web documents are not enough to find good answer documents. Link information and URL information compensates for the insufficiencies of content information. However, static combination of multiple evidences may lower the retrieval performance . We need different strategies to find target documents according to a query type. We can classify user queries as three categories, the topic relevance task, the homepage finding task, and the service finding task. In this paper, a user query classification scheme is proposed. This scheme uses the difference of distribution, mutual information , the usage rate as anchor texts, and the POS information for the classification. After we classified a user query, we apply different algorithms and information for the better results. For the topic relevance task, we emphasize the content information, on the other hand, for the homepage finding task, we emphasize the Link information and the URL information. We could get the best performance when our proposed classification method with the OKAPI scoring algorithm was used.

INTRODUCTION The Web is rich with various sources of information. It contains the contents of documents, web directories, multi-media data, user profiles and so on. The massive and heterogeneous web document collections as well as the unpredictable querying behaviors of typical web searchers exacerbate Information Retrieval (IR) problems. Retrieval approaches based on the single source of evidence suffer from weakness that can hurt the retrieval performance in certain situations [5]. For example, content-based IR approaches have a difficulty in dealing with the diversity in vocabulary and the quality of web documents, while link-based approaches can suffer from an incomplete or noisy link structure . Combining multiple evidences compensates for the weakness of a single evidence [17]. Fusion IR studies have repeatedly shown that combining multiple sources of evidence can improve retrieval performance [5][17]. However, previous studies did not consider a user query in combining evidences [5][7][10][17]. Not only documents in the Web but also users' queries are diverse. For example, for user query `Mutual Information' , if we count on link information too highly, well-known site that has `mutual funds' and `information' as index terms gets the higher rank. For user query `Britney's Fan Club' , if we use content information too highly, yahoo or lycos's web directory pages get the higher rank, instead of the Britney's fan club site. Like these examples, combining content information and link information is not always good. We have to use different strategies to meet the need of a user. User queries can be classified as three categories according to their intent [4]. topic relevance task (informational) homepage finding task (navigational) service finding task (transactional) The topic relevance task is a traditional ad hoc retrieval task where web documents are ranked by decreasing likelihood of meeting the information need provided in a user query [8]. For example, `What is a prime factor?' or `prime factor' is a query of the topic relevance task. The goal of this query is finding the meaning of `prime factor'. The homepage finding task is a known-item task where the goal is to find the homepage (or site entry page) of the site described in a user query. Users are interested in finding a certain site. For example, `Where is the site of John Hopkins Medical Institutions ?' or `John Hopkins Medical Institutions' is a query of the homepage finding task. The goal of this query is finding the entry page of `John Hopkins Medical Institutions'. The service finding task is a task where the goal is to find 64 web documents that provide the service described in a user query. For example, `Where can I buy concert tickets?' or `buy concert tickets' is a query of the service finding task. The goal of this query is finding documents where they can buy concert tickets. Users may want different documents with the same query. We cannot always tell the class of a query clearly. But we can tell most people want a certain kind of documents with this query. In this paper, we calculate the probability that the class of a user query is the topic relevance task or the homepage finding task. Based on this probability, we combine multiple evidences dynamically. In this paper, we consider the topic relevance task and the homepage finding task only. Because the proposed method is based on the difference of databases, we can apply the same method to classify the service finding task. In this paper, we present a user query classification method and a combining method for each query type. In section 2, we describe various types of information (Content, Link, and URL information). Section 3 lists the differences of search tasks and the properties of Content, Link, and URL information . In section 4, we present the model of a query classification . In section 5, we experiment with our proposed model. Conclusion is described in section 6. MULTIPLE SOURCES OF INFORMATION In this section, we explain various sources of information for the web document retrieval. There are three types of information , Content information, Link information, and URL information. 2.1 Content Information There are multiple types of representations for a document . These representations typically contain titles, anchor texts, and main body texts [5]. A title provides the main idea and the brief explanation of a web document. An anchor text provides the description of linked web documents and files. An anchor text often provides more accurate description of a web document than the document itself. We usually use tf and df to calculate the relevance of a given web documents [1]. tf is the raw frequency of a given term inside a document. It provides one measure of how well that term describes the document contents. df is the number of documents in which the index term appears. The motivation for using an inverse document frequency is that terms that appear in many documents are not very useful for distinguishing a relevant document from a non-relevant one. There are various scoring algorithms that use tf and df . These scoring algorithms include the normalization and the combination of each factor, tf and df . 2.2 Link Information A hyperlink in a web document is a kind of citation. The essential idea is that if page u has a link to page v, then the author of u is implicitly assigning some importance to page v. Since we can represent the Web as a graph, we can use graph theories to help us make a search engine that returns the most important pages first. The PageRank or P R(A) of a page A is given as follows [13]. P R(A) = (1 - d) + (1) d(P R(T 1 )/C(T 1 ) + . . . + P R(T n )/C(T n )) We assume page A has pages T 1 . . . T n that point to it. The parameter d is a damping factor that can be set between 0 and 1. Also C(A) is defined as the number of links going out of a page A. P R(A) can be calculated using a simple iterative algorithm, and corresponds to the principal eigenvector of the normalized link matrix of the Web [3]. 2.3 URL Information The URL string of a site entry page often contains the name or acronym of the corresponding organization. Therefore , an obvious way of exploiting URL information is trying to match query terms and URL terms. Additionally, URLs of site entry pages tend to be higher in a server's directory tree than other web documents, i.e. the number of slashes (`/') in an entry page URL tends to be relatively small. Kraaij et al. suggested 4 types of URLs [16]. root: a domain name (e.g. http://trec.nist.gov) subroot: a domain name followed by a single directory (e.g. http://trec.nist.gov/pubs/) path: a domain name followed by an arbitrarily deep path (e.g. http://trec.nist.gov/pubs/trec9/papers) file: anything ending in a filename other than `in-dex .html' (e.g. http://trec.nist.gov/pubs/trec9/t9proc.html) Kraaij et al. estimated a prior probability (URLprior ) of being an entry page on the basis of the URL type for all URL types t (root, subroot, path, and file). 2.4 Combination of Information We can combine results of each search engine or scores of each measure to get better results. Croft proposed the IN-QUERY retrieval system, based on the inference network, to combine multiple evidences [5]. The inference network model is a general model for combining information. It is data-level fusion. The model is based on probabilistic updating of the values of nodes in the network, and many retrieval techniques and information can be implemented by config-uring the network properly. Several researchers have experimented with linearly combining the normalized relevance scores (s i ) given to each document [7][10][16]. score(d) = i i s i (d) (2) It requires training for the weight i given to each input system. For example, we can get a better result by combining content information and URL type information with the following weight [16]. score(d) = 0.7 content + 0.3 URLprior (3) TOPIC RELEVANCE TASK AND HOMEPAGE FINDING TASK In this section, we show properties of Content information, Link information, and URL information in each search task. Besides, we will propose the method for linearly combining information for each task. 65 We use TREC data collection, to show the differences of each search task. We made a simple search engine that use the variation of the OKAPI scoring function [15]. Given a query Q, the scoring formula is: score = t (QD d ) T F d,t IDF t (4) T F d,t = 0.4 + 0.6 tf d,t tf d,t + 0.5 + 1.5 doclen d avg doclen (5) IDF t = log( N + 0.5 df t )/log(N + 1) (6) N is the number of documents in the collection. tf d,t is the number of occurrences of an index term t in a document d, and df t is the number of documents in which t occurs. We use the data for the web track, the 10-gigabyte WT10g collection [2], distributed by CSIRO [6]. We use TREC-2001 topic relevance task queries (topics 501-550) for the topic relevance task, and 145 queries for the homepage finding task [8]. For the homepage finding task, NIST found a homepage within WT10g and then composed a query designed to locate it. We used the anchor text representation (Anchor) and the common content text representation (Common) for indexing . Every document in the anchor text representation has anchor texts and the title as content, and excludes a body text. Consequently the anchor text representation has brief or main explanations of a document. We used two other evidences for a scoring function besides the OKAPI score. One is URLprior for URL information and the other is PageRank for Link information. We linearly interpolated Content information (OKAPI score), URLprior, and PageRank. We call this interpolation as CMB . rel(d) = 0.65 Content Information + (7) 0.25 URL Information + 0.1 Link Information We used `and' and `sum' operators for matching query terms [1]. `and' operator means that the result document has all query terms in it. `sum' operator means that a result document has at least one query term in it. Table 1 shows the average precision of the topic relevance task and the MRR of the homepage finding task [8]. The first column in the table 1 means the method that we used for indexing and scoring. For example, `Anchor and CMB' means that we used the anchor text representation for indexing , `and' operator for query matching, and the OKAPI score, PageRank and URLprior for scoring. The average precision is defined as the average of the precision obtained at the rank of each relevant document. P avg = 1 |R| d R R r(d) r(d) (8) R is the set of all relevant documents and R r(d) is the set of relevant documents with rank r(d) or better. MRR (Mean Reciprocal Rank) is the main evaluation measure for the homepage finding task. MRR is based on the rank of the first correct document (answer i rank) according to the Table 1: Topic Relevance Task vs. Homepage Finding Task Topic Homepage model P avg MRR Anchor and 0.031 0.297 Anchor and CMB 0.031 0.431 Anchor sum 0.034 0.351 Anchor sum CMB 0.034 0.583 Common and 0.131 0.294 Common and CMB 0.122 0.580 Common sum 0.182 0.355 Common sum CMB 0.169 0.673 MAX 0.226 0.774 AVG 0.145 0.432 following formula: M RR = 1 #queries #queries i =1 1 answer i rank (9) M AX represents the best score of a search engine that submitted in TREC-2001. AV G represents the average score of all search engines that submitted in TREC-2001. We got the better result with the common content text representation than the anchor text representation in the topic relevance task. A title and anchor texts do not have enough information for the topic relevance task. On the other hand, we could get the similar performance with the anchor text representation in the homepage finding task. URL information and Link information are good for the homepage finding task but bad for the topic relevance task. In the topic relevance task, we lost our performance by combining URL and Link information. The query of the topic relevance task usually consists of main keywords that are relevant to some concept or the explanation of what they want to know. However, we cannot assume that other people use same expressions and keywords to explain what a user wants to know. Therefore we could not get a good result with `and' operator in the topic relevance task. But on the other hand the query of the homepage finding task consists of entity names or proper nouns. Therefore we could have good results with `and' operator when we can have a result document. However, the MRR of `Anchor and CMB' is lower than that of `Common sum CMB' in the homepage finding task. `Anchor and CMB' method did not retrieve a document for 31 queries. To compensate for this sparseness problem, we combined the results of `Anchor and CMB' and `Common sum CMB' . This combined result showed 0.730 in the homepage finding task. When we combined the results of `Anchor and ' and `Common sum' , it showed 0.173 in the topic relevance task. This implies that the result documents with `and' operator are good and useful in the homepage finding task. We can conclude that we need different retrieval strategies according to the category of a query. We have to use the field information (title, body, and anchor text) of each term, and combine evidences dynamically to get good results. In the topic relevance task, the body text of a document is good for indexing, `sum' operator is good for query term matching, 66 and combining URL and Link information are useless. On the other hand, in the homepage finding task, anchor texts and titles are useful for indexing, `and' operator is also good for query term matching, and URL and Link information is useful. By combining results from main body text and anchor texts and titles we can have the better performance. USER QUERY CLASSIFICATION In this section, we present the method for making a language model for a user query classification. 4.1 Preparation for Language Model We may use the question type of a query to classify the category of a user query. For example, "What is a two electrode vacuum tube?" is a query of the topic relevance task. "Where is the site of SONY?" is a query of the homepage finding task. We can assume the category of a query with an interrogative pronoun and cue expressions (e.g. `the site of'). However, people do not provide natural language queries to a search engine. They usually use keywords for their queries. It is not easy to anticipate natural language queries. In this paper, we assume that users provide only main keywords for their queries. We define a query Q as the set of words. Q = {w 1 , w 2 , . . . , w n } (10) To see the characteristics of each query class, we use two query sets. For the topic relevance task, TREC-2000 topic relevance task queries (topics 451-500) are used. For the homepage finding task, queries for randomly selected 100 homepages 1 are used. We call them QU ERY T -T RAIN and QU ERY H -T RAIN . We divided WT10g into two sets, DB T OP IC and DB HOM E . If the URL type of a document is `root' type, we put this document to DB HOM E . Others are added to DB T OP IC . According to the report of [16], our division method can get site entry pages with 71.7% precision. Additionally we put virtual documents into DB HOM E with anchor texts. If a linked document is in DB T OP IC , then we make a virtual document that consists of anchor texts and put it into DB HOM E . If a linked document is in DB HOM E , then we add anchor texts to the original document. Usually a site entry page does not have many words. It is not an explanatory document for some topic or concept, but the brief explanation of a site. We can assume that site entry pages have the different usage of words. If we find distinctive features for site entry pages, then we can discriminate the category of a given query. #DB T OP IC and #DB HOM E mean the number of documents in the DB T OP IC and DB HOM E respectively. However , most documents in the DB HOM E have a short length, we normalized the number of documents with the following equation. #DB T OP IC = # of documents in DB T OP IC (11) #DB HOM E = # of documents in DB HOM E (12) avg doclength HOM E avg doclength T OP IC 1 available at http://www.ted.cmis.csiro.au/TRECWeb/Qrels/ 4.2 Distribution of Query Terms `Earthquake' occurs more frequently in DB T OP IC . But `Hunt Memorial Library' shows the high relative frequency in DB HOM E . General terms tend to have same distribution regardless of the database. If the difference of distribution is larger than expected, this tells whether a given query is in the topic relevance task class or the homepage finding task class. We can calculate the occurrence ratio of a query with the following equation [11]. Dist(w 1 , . . . , w n ) = n C(w 1 , . . . , w n ) n i =1 C(w i ) (13) C(w) is the number of documents that have w as an index term. df of w is used for C(w). C(w 1 , . . . , w n ) is the number of documents that have all w 1 , . . . , w n as index terms. To see the distribution difference of a query, we use the following ratio equation. dif f Dist (Q) = Dist HOM E (Q) Dist T OP IC (Q) (14) If a query has only one term, we use the chi-square [11]. We make a 2-by-2 table for the given word `w'. word=w word = w DB T OP IC a b DB HOM E c d a + b = #DB T OP IC and c + d = #DB HOM E . `a' is the frequency of the word `w' in the DB T OP IC and `c' is the frequency of the word `w' in the DB HOM E . The chi-square value shows the dependence of the word `w' and DB. If the chi-square value of the word `w' is high, then `w' is a special term of DB T OP IC or DB HOM E . We classify these words that have a high chi-square value according to the df . If `w' has a high df then the word `w' is the topic relevance task query. Otherwise `w' is the homepage finding task query. For example, `f ast' shows the high chi-square value, since it is used a lot to modify proper names. However, one word `f ast' is not the proper name. We classify a word that has a high chi-square and a high df into the topic relevance task. If the chi-square value of the word `w' is low, then `w' is a general term. Fig.2 shows the results of dif f Dist of queries that have at least two query terms. The mean values of QU ERY T -T RAIN 's dif f Dist and QU ERY H -T RAIN 's dif f Dist are 0.5138 and 1.1 respectively. As the value of dif f Dist of a given query is higher, we can have confidence that the query has special terms. On the other hand, if the score of dif f Dist is near the mean value of QU ERY T -T RAIN , it means the query has general terms, not a special expression. We calculate the possibility that a given query is in each class with the mean value and the standard deviation. However, there are queries that show high dif f DIST in QU ERY T -T RAIN . For example, `Jenniffer Aniston' and `Chevrolet Trucks' showed 2.04 and 0.76 respectively. Usually proper names showed high dif f DIST values. If a proper name is frequently used in the DB HOM E , then we can think of it as the name of the site. 4.3 Mutual Information There are two or more words that co-occur frequently. These words may have syntactic or semantic relations to 67 if length(Q)=1 then calculate the 2 of Q if 2 &gt; 18 then if df of a query &gt; 65 the topic relevance task else the homepage finding task else the topic relevance task else calculate distributions of a query in each database calculate dif f Dist(Q) if dif f Dist(Q) &gt; the homepage finding task else unknown Figure 1: The Algorithm of Distribution Difference Method 0% 10% 20% 30% 40% 50% 60% 70% 0.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5 Ratio of Distribution Difference P e r c e n t a g e o f O b se r v at i o n s QUERY-TOPIC-TRAIN QUERY-HOMEPAGE-TRAIN Figure 2: Distribution of Queries each other. We say these words have some dependency. For example, `tornadoes formed' shows similar dependency regardless of the database. But `Fan Club' has a high dependency in DB HOM E set. This means that `tornadoes formed' is a general usage of words but `Fan Club' is a special usage in DB HOM E . Therefore, the dependency of `Fan Club' can be the key clue of guessing the category of a user query. If the difference of dependency of each term is larger than expected , this tells whether a given query is the topic relevance task or the homepage finding task. For two variables A and B, we can calculate the dependency with mutual information , I(A; B) [9]. We use the pointwise mutual information I(x, y) to calculate the dependency of terms in a query [11]. I(A; B) = H(A) + H(B) - H(A, B) = a,b p(a, b)log p(a, b) p(a)p(b) (15) I(x, y) = log p(x, y) p(x)p(y) (16) We extend pointwise mutual information for three variables . We use the set theory to calculate the value of an 0% 5% 10% 15% 20% 25% 30% 0.25 0.75 1.25 1.75 2.25 2.75 3.25 3.75 4.25 4.75 5.25 Ratio of MI Difference P e r c e n t a g e o f O b s e r v a t i o n QUERY-TOPIC-TRAIN QUERY-HOMEPAGE-TRAIN Figure 3: Mutual Information of Queries intersection part, like two variables problem. I(A; B; C) = H(A, B, C) - H(A) - H(B) - H(C) + I(A; B) + I(B; C) + I(C; A) = a,b,c p(a, b, c)log p(a, b)p(b, c)p(c, a) p(a, b, c)p(a)p(b)p(c) (17) I(x, y, z) = log p(x, y)p(y, z)p(z, x) p(x, y, z)p(x)p(y)p(z) (18) In principle, p(x, y) means the probability that x and y are co-occurred in a specific distance [11]. Usually x and y are consecutive words. Since the number of words and documents are so huge in IR domain, it is not easy to keep statistics . Our measure assume that x and y are co-occurred in a document. We use df of a given term to calculate the number of documents that contain a term. Like the distribution difference measure, we use the ratio difference equation to see the difference of MI. If pointwise mutual information is below zero then we use zero. dif f M I (Q) = M I HOM E (Q) M I T OP IC (Q) (19) Fig.3 shows the results of dif f M I . The mean values of QU ERY T -T RAIN 's dif f M I and QU ERY H -T RAIN 's dif f M I are 1.9 and 2.7 respectively. For example, the topic relevance task query `mexican food culture' showed 1.0, but the homepage finding task query `Newave IFMO' showed 7.5. QU ERY H -T RAIN gets a slightly high standard deviation. It means that the query of QU ERY H -T RAIN has different MI in DB HOM E . As the value of dif f M I is higher, we can have confidence that the query has a special dependency. We calculate the possibility that a given query is in each class with the mean value and the standard deviation. 4.4 Usage Rate as an Anchor Text If query terms appear in titles and anchor texts frequently, this tells the category of a given query is the homepage finding task. Titles and anchor texts are usually entity names or proper nouns, the usage rate shows the probability that given terms are special terms. use Anchor (w 1 , . . . , w n ) = (20) C SIT E AN CHOR (w 1 , . . . , w n ) - C SIT E (w 1 , . . . , w n ) C SIT E (w 1 , w 2 , . . . , w n ) 68 C SIT E (w) means the number of site entry documents that have w as an index term. C SIT E AN CHOR (w) means the number of site entry documents and anchor texts that have w as an index term. 4.5 POS information Since the homepage finding task queries are proper names, they do not usually contain a verb. However, some topic relevance task queries include a verb to explain what he or she wants to know. For example, `How are tornadoes formed?' or briefly `tornadoes formed' contain a verb `formed'. If a query has a verb except the `be' verb, then we classified it into the topic relevance task. 4.6 Combination of Measures The difference of distribution method can apply more queries than the difference of MI. The usage rate as anchor texts and the POS information show small coverage. However , four measures cover different queries. Therefore, we can have more confidence and more coverage by combining these measures. We use a different combination equation as the number of query terms. If the query has 2 and 3 terms in it, we use pointwise mutual information also. S(Q) = diff Dist (Q) + diff M I (Q) + (21) use Anchor (Q) + P OS inf o (Q) We choose , , , and with train data (QU ERY T -T RAIN and QU ERY H -T RAIN ). If `S(Q)' score is not high or low enough, then we make no decision. EXPERIMENTS In this section, we show the efficiency of a user query classification. 5.1 Query Classification We used four query sets for experimenting our query classification method. QU ERY T -T RAIN and QU ERY H -T RAIN are used for training (TRAIN). TREC-2001 topic relevance task queries (Topic 501-550) and TREC-2001 homepage finding task queries (1-145) are used for testing (TEST). We call two test sets as QU ERY T -T EST and QU ERY H -T EST . We used WT10g for making a classification model. We classified queries with our proposed method. If the score `S(Q)' is high enough to tell that a given query is in the topic relevance task or the homepage finding task query, then we assigned the query type to it. For other cases, we did not classify a query category. Table 2 shows the classification result of our proposed language model. Table 2: Query Classification Result QUERY TRAIN TEST Measure Precision Recall Precision Recall Dist. 77.3% 38.7% 82.1% 28.2% MI 90.9% 20.0% 78.2% 29.9% Anchor 73.6% 35.3% 82.4% 35.9% POS 100% 9.3% 96.4% 13.8% All 81.1% 57.3% 91.7% 61.5% By combining each measure, we could apply our method to more queries and increase precision and recall. Our pro-Table 3: Average Precision of the Topic Relevance Task model OKAPI TF-IDF KL DIR MIXFB KL D Lemur 0.182 0.170 0.210 0.219 MLemur 0.169 0.159 0.200 0.209 Table 4: MRR of the Homepage Finding Task model OKAPI TF-IDF KL DIR MIXFB KL D Lemur 0.355 0.340 0.181 0.144 MLemur 0.673 0.640 0.447 0.360 posed method shows the better result in the test set. This is due to the characteristics of the query set. There are 7 queries that have a verb in QU ERY T -T RAIN and 28 queries in QU ERY T -T EST . We can assume that the POS information is good information. The main reason of misclassification is wrong division of WT10g. Since our method usually gives the high score to the proper name, we need correct information to distinguish a proper name from a site name. We tried to make DB HOM E automatically. However, some root pages are not site entry pages. We need a more sophisticated division method. There is a case that a verb is in the homepage finding task query. `Protect & Preserve' is the homepage finding task query but `protect' and `preserve' are verbs. However, `Protect' and `Preserve' start with a capital letter. We can correct wrong POS tags. There are queries in QU ERY T -T EST that look like queries of QU ERY H -T EST . For example, `Dodge Recalls' is used to find documents that report on the recall of any dodge automobile products. But user may want to find the entry page of `Dodge recall'. This is due to the use of main keywords instead of a natural language query. There are 6 queries in QU ERY T -T EST and 6 queries in QU ERY H -T EST that do not have a result document that has all query terms in it. We could not use our method to them. WT10g is not enough to extract probability information for these two query sets. To make up this sparseness problem, we need a different indexing terms extraction module . We have to consider special parsing technique for URL strings and acronyms in a document. Also we need a query expansion technique to get a better result. 5.2 The Improvement of IR Performance We used the Lemur Toolkit [12] to make a general search engine for the topic relevance task. The Lemur Toolkit is an information retrieval toolkit designed with language modeling in mind. The Lemur Toolkit supports several retrieval algorithms. These algorithms include a dot-product function using TF-IDF weighting algorithm, the Kullback-Leibler (KL) divergence algorithm, the OKAPI retrieval algorithm, the feedback retrieval algorithm and the mixture model of Dirichlet smoothing, MIXFB KL D [14]. For the homepage finding task, we add the URLprior probability of a URL string to the Lemur Toolkit. Besides Link information, we add the PageRank of a document. We normalized PageRank values, so the max value is 100 and the min value is 0. First we extracted top 1,000 results with the Lemur Toolkit. 69 Table 5: The Retrieval Performance with Classification Method OKAPI TF-IDF MIXFB KL D Measure DEFAULT TOPIC HOME TOPIC HOME TOPIC HOME Dist. TOPIC 0.178 0.469 0.168 0.447 0.216 0.226 Dist. HOME 0.174 0.666 0.164 0.633 0.212 0.359 MI TOPIC 0.179 0.465 0.168 0.445 0.218 0.233 MI HOME 0.169 0.673 0.159 0.640 0.209 0.360 Anchor TOPIC 0.176 0.513 0.165 0.489 0.215 0.232 Anchor HOME 0.169 0.666 0.159 0.633 0.209 0.359 POS TOPIC 0.182 0.355 0.170 0.340 0.219 0.144 POS HOME 0.173 0.673 0.163 0.640 0.212 0.354 All TOPIC 0.180 0.552 0.168 0.528 0.217 0.280 All HOME 0.173 0.666 0.163 0.633 0.212 0.353 Then we combined URL information and Link information to reorder results with the equation Eq. 7. We presented top 1,000 documents as the answer in the topic relevance task, and 100 documents in the homepage finding task. We call this modified Toolkit as MLemur Toolkit. Table 3 and 4 show results of the topic relevance task and the homepage finding task that use the Lemur Toolkit and the MLemur Toolkit. MIXFB KL D showed the good result in the topic relevance task but showed the poor result in the homepage finding task. We can say that a good information retrieval algorithm for the topic relevance task is not always good for the homepage finding task. We chose three algorithms , the OKAPI , the TF-IDF , and the MIXFB KL D that got the best and worst score in each task, for the test of performance improvement by query type classification. Table 5 shows the change of performance. `DEFAULT' means the default category for an unclassified query. Digits in the TOPIC column and the HOME column are average precision and MRR respectively. From the result, the OKAPI algorithm and the homepage finding task as a default class method shows the good performance. 5.3 Discussion To classify a query type, we need the document frequency of a query term in each database. This lowers the system efficiency . However, we may create two databases as proposed in this paper for indexing. We retrieve two result document sets from each database and classify a query type at the same time. And then according to the category of a query, merge two results. From table 1, merging the results of the anchor text representation and the common content representation shows good performance. We need more work to unify the query classification work and the document retrieval. In this paper, we proposed a user query classification method for the topic relevance task and the homepage finding task. The queries of the homepage finding task usually consist of entity names or proper nouns. However queries of the service finding task have verbs for the service definition. For example, "Where can I buy concert tickets?" has `buy' as the service definition. To find these cue expressions, we need more sophisticated analysis of anchor texts. Since the service in the Web is provided as a program, there is a trigger button. Mostly these trigger buttons are explained by anchor texts. We have to distinguish an entity name and an action verb from anchor texts. We have to change measures for the query classification from a word unit to entity and action units. User query classification can be applied to various areas. MetaSearch is the search algorithm that combines results of each search engine to get the better result [7]. [10] proposed CombMNZ, Multiply by NonZeros, is better than other scoring algorithm, CombSUM , Summed similarity over systems. But if we consider the homepage finding task, we are in a different situation. Table 6 and 7 show the improvement of performance of MetaSearch algorithms. We had an experiment with random samplings of 2, 3, 4, and 5 engine results. The score is the average improvement of 100 tests. CombMNZ was good for the topic relevance task, but CombSUM was good for the homepage finding task. It also tells, we need different strategies for MetaSearch as the class of a query. Table 6: Performance of MetaSearch in the Topic Relevance Task engine # 2 3 4 5 CombSUM -2.4% 4.4% 3.7% 4.8% CombMNZ -1.2% 5.7% 5.3% 5.8% Table 7: Performance of Metasearch in the Homepage Finding Task engine # 2 3 4 5 CombSUM -4.5% 0.7% -0.9% 0.8% CombMNZ -6.0% -0.4% -4.5% -2.4% CONCLUSIONS We have various forms of resources in the Web, and consequently purposes of user queries are diverse. We can classify user queries as three categories, the topic relevance task, the homepage finding task, and the service finding task. Search engines need different strategies to meet the purpose of a user query. For example, URL information and Link information are bad for the topic relevance task, but on the other hand, they are good for the homepage finding task. We made two representative databases, DB HOM E 70 and DB T OP IC , for each task. To make databases, we divided text collection by the URL type of a web document. If the URL of a document contains a host name only, then we put it into DB HOM E . Also we make a virtual document with an anchor text and put it into DB HOM E . Other documents are put into DB T OP IC . If given query's distributions in DB HOM E and DB T OP IC are different, then this tells a given query is not a general word. Therefore, we can assume the category of a given query is in the homepage finding task. Likewise, the difference of dependency, Mutual Information, and the usage rate as anchor texts tell whether a given query is in the homepage finding task or not. We tested the proposed classification method with two query sets, QU ERY T -T EST and QU ERY H -T EST . The usage rate as anchor texts and the POS information show small coverage. On the other hand, distribution difference and dependency showed good precision and coverage. Also each classifier applied to different queries. We could get the better precision and recall by combining each classifier. We got 91.7% precision and 61.5% recall. After we classified the category of a query, we used different information for a search engine. For the topic relevance task, Content information such as TFIDF is used. For the homepage finding task, Link information and URL information besides content information are used. We tested our dynamic combining method. From the result, our classification method showed the best result with the OKAPI scoring algorithm. ACKNOWLEDGMENTS We would like to thank Jamie Callan for providing useful experiment data and the Lemur toolkit. REFERENCES [1] R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. ACM PRESS BOOKS, 1999. [2] P. Bailey, N. Craswell, and D. Hawking. Engineering a multi-purpose test collection for web retrieval experiments. Information Processing and Management, to appear. [3] S. Brin and L. Page. The anatomy of a large-scale hypertextual Web search engine. Computer Networks and ISDN Systems, 30(1-7):107117, 1998. [4] A. Broder. A taxonomy of web search. SIGIR Forum, 36(2), 2002. [5] W. B. Croft. Combining approaches to information retrieval. In Advances in Information Retrieval: Recent Research from the Center for Intelligent Information Retrieval, pages 136. Kluwer Academic Publishers, 2000. [6] CSIRO. Web research collections - trec web track. www.ted.cmis.csiro.au /TRECWeb/, 2001. [7] E. Fox and J. Shaw. Combination of multiple searches. In Text REtrieval Conference (TREC-1), pages 243252, 1993. [8] D. Hawking and N. Craswell. Overview of the trec-2001 web track. In Text REtrieval Conference (TREC-10), pages 6167, 2001. [9] E. Jaynes. Information theory and statistical mechanics. Physics Review, 106(4):620630, 1957. [10] J. H. Lee. Analyses of multiple evidence combination. In Proceedings of the 20th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 267276, 1997. [11] C. D. Manning and H. Schutze. Foundations of Statistical Natural Language Processing. The MIT Press, 1999. [12] P. Ogilvie and J. Callan. Experiments using the lemur toolkit. In Text REtrieval Conference (TREC-10) http://www-2.cs.cmu.edu/ lemur, pages 103108, 2001. [13] L. Page, S. Brin, R. Motwani, and T. Winograd. The pagerank citation ranking: Bringing order to the web. Technical report, Stanford Digital Library Technologies Project, 1998. [14] J. M. Ponte. Language models for relevance feedback. In W. B. Croft, editor, Advances in Information Retrieval: Recent Research from the Center for Intelligent Information Retrieval, pages 7395. Kluwer Academic Publishers, 2000. [15] S. E. Robertson, S. Walker, S. Jones, M. Hancock-Beaulieu, and M. Gatford. Okapi at trec-3. In Text REtrieval Conference (TREC-2), pages 109126, 1994. [16] T. Westerveld, W. Kraaij, and D. Hiemstra. Retrieving web pages using content, links, urls and anchors. In Text REtrieval Conference (TREC-10), pages 663672, 2001. [17] K. Yang. Combining text and link-based retrieval methods for web ir. In Text REtrieval Conference (TREC-10), pages 609618, 2001. 71

URL Information;web document;URL;improvement;frequency;task;information;model;rate;IR;Combination of Multiple Evidences;Link Information;query;Query Classification

162

Querying Bi-level Information

In our research on superimposed information management, we have developed applications where information elements in the superimposed layer serve to annotate, comment, restructure, and combine selections from one or more existing documents in the base layer. Base documents tend to be unstructured or semi-structured (HTML pages, Excel spreadsheets, and so on) with marks delimiting selections. Selections in the base layer can be programmatically accessed via marks to retrieve content and context. The applications we have built to date allow creation of new marks and new superimposed elements (that use marks), but they have been browse-oriented and tend to expose the line between superimposed and base layers. Here, we present a new access capability, called bi-level queries, that allows an application or user to query over both layers as a whole. Bi-level queries provide an alternative style of data integration where only relevant portions of a base document are mediated (not the whole document) and the superimposed layer can add information not present in the base layer. We discuss our framework for superimposed information management, an initial implementation of a bi-level query system with an XML Query interface, and suggest mechanisms to improve scalability and performance.

INTRODUCTION You are conducting background research for a paper you are writing. You have found relevant information in a variety of sources: HTML pages on the web, PDF documents on the web and on your SIGMOD anthology of CDs, Excel spreadsheets and Word documents from your past work in a related area, and so on. You identify relevant portions of the documents and add annotations with clarifications, questions, and conclusions. As you collect information, you frequently reorganize the information you have collected thus far (and your added annotations) to reflect your perspective. You intentionally keep your information structure loose so you can easily move things around. When you have collected sufficient information, you import it, along with your comments, in to a word-processor document. As you write your paper in your word-processor, you revisit your sources to see information in its context. Also, as you write your paper you reorganize its contents, including the imported information, to suit the flow. Occasionally, you search the imported annotations, selections, and the context of the selections. You mix some of the imported information with other information in the paper and transform the mixture to suit presentation needs. Most researchers will be familiar with manual approaches to the scenario we have just described. Providing computer support for this scenario requires a toolset with the following capabilities: 1. Select portions of documents of many kinds (PDF, HTML, etc.) in many locations (web, CD, local file system, etc.), and record the selections. 2. Create and associate annotations (of varying structure) with document selections. 3. Group and link document selections and annotations, reorganize them as needed, and possibly even maintain multiple organizations. 4. See a document selection in its context by opening the document and navigating to the selected region, or access the context of a selection without launching its original document. 5. Place document selections and annotations in traditional documents (such as the word-processor document that contains your paper). 6. Search and transform a mixture of document selections, annotations, and other information. Systems that support some subset of these capabilities exist, but no one system supports the complete set. It is hard to use a collection of systems to get the full set of features because the systems do not interoperate well. Some hypertext systems can create multiple organizations of the same information, but they tend to lack in the types of source, granularity of information, or the location of information consulted. For example, Dexter [6] requires all information consulted to be stored in its proprietary database. Compound document systems can address sub-documents , but they tend to have many display constraints. For example, OLE 2 [9] relies on original applications to render information. Neither type of system supports querying a mixture of document selections and annotations. Superimposed information management is an alternative solution for organizing heterogeneous in situ information, at document and sub-document granularity. Superimposed information (such as annotations) refers to data placed over existing information sources (base information) to help organize, access, connect and reuse information elements in those sources [8]. In our previous work [12], we have described the Superimposed Pluggable Architecture for Contexts and Excerpts (SPARCE), a middleware for superimposed information management, and presented some superimposed applications built using SPARCE. Together they support Capabilities 1 through 4. In this paper, we show how SPARCE can be used to support Capability 6. Details of support for Capability 5 are outside the scope of this paper. Before we proceed with the details of how we support Capability 6, we introduce a superimposed application called RIDPad [12]. Figure 1 shows a RIDPad document that contains information selections and annotations related to the topic of information integration. The document shown contains eight items: CLIO, Definition, SchemaSQL, Related Systems, Goal, Model, Query Optimizer, and Press. These items are associated with six distinct base documents of three kinds--PDF, Excel, and HTML. An item has a name, a descriptive text, and a reference (called a mark) to a selection in a base document. For example, the item labeled `Goal' contains a mark into a PDF document. The boxes labeled Schematic Heterogeneity and Garlic are groups. A group is a named collection of items and other groups. A RIDPad document is a collection of items and groups. RIDPad affords many operations for items and groups. A user can create new items and groups, and move items between groups. The user can also rename, resize, and change visual characteristics such as color and font for items and groups. With the mark associated with an item, the user can navigate to the base layer if necessary, or examine the mark's properties and browse context information (such as containing paragraph) from within RIDPad via a reusable Context Browser we have built. The operations RIDPAD affords are at the level of items and groups. However, we have seen the need to query and manipulate a RIDPad document and its base documents as a whole. For example, possible queries over the RIDPad document in Figure 1 include: Q1: List base documents used in this RIDPad document. Q2: Show abstracts of papers related to Garlic. Q3: Create an HTML table of contents from the groups and items. Query Q1 examines the paths to base documents of marks associated with items in the RIDPad document. Q2 examines the context of marks of items in the group labeled `Schematic Heterogeneity.' Q3 transforms the contents of the RIDPad document to another form (table of contents). In general, queries such as these operate on both superimposed information and base information. Consequently, we call them bi-level queries. Figure 1: A RIDPad document. There are many possible choices on how to present the contents of superimposed documents (such as the RIDPad document in Figure 1) and base documents for querying. We could make the division between the superimposed and base documents obvious and let the user explicitly follow marks from superimposed information to base information. Instead, our approach is to integrate a superimposed document's contents and related base information to present a uniform representation of the integrated information for querying. The rest of this paper is organized as follows: Section 2 provides an overview of SPARCE. Section 3 provides an overview of bi-level query systems and describes a nave implementation of a bi-level query system along with some example bi-level queries. Section 4 discusses some applications and implementation alternatives for bi-level query systems. Section 5 briefly reviews related work. Section 6 summarizes the paper. We use the RIDPad document in Figure 1 for all examples in this paper. SPARCE OVERVIEW The Superimposed Pluggable Architecture for Contexts and Excerpts (SPARCE) facilitates management of marks and context information in the setting of superimposed information management [12]. A mark is an abstraction of a selection in a base document. Several mark implementations exist, typically one per base type (PDF, HTML, Excel, and so on). A mark implementation chooses an addressing scheme appropriate for the base type it supports. For example, an MS Word mark implementation uses the starting and ending character index of a text selection, whereas an MS Excel mark uses the row and column names of the first and last cell in the selection. All mark implementations provide a common interface to address base information, regardless of base type or access protocol they 8 support. A superimposed application can work uniformly with any base type due to this common interface. Context is information concerning a base-layer element. Presentation information such as font name, containment information such as enclosing paragraph and section, and placement information such as line number are examples of context information. An Excerpt is the content of a marked base-layer element. (We treat an excerpt also as a context element.) Figure 2 shows the PDF mark corresponding to the item `Goal' (of the RIDPad document in Figure 1) activated. The highlighted portion is the marked region. Table 1 shows some of the context elements for this mark. Figure 2: A PDF mark activated. Figure 3 shows the SPARCE architecture reference model. The Mark Management module is responsible for operations on marks (such as creating and storing marks). The Context Management module retrieves context information. The Superimposed Information Management module provides storage service to superimposed applications. The Clipboard is used for inter-process communication. Table 1: Some context elements of a PDF mark. Element name Value Excerpt provide applications and users with ... Garlic system Font name Times New Roman Enclosing paragraph Loosely speaking, the goal ... Section Heading Garlic Overview SPARCE uses mediators [13] called context agents to interact with different base types. A context agent is responsible for resolving a mark and returning the set of context elements appropriate to that mark. A context agent is different from mediators used in other systems because it only mediates portions of base document a mark refers to. For example, if a mark refers to the first three lines of a PDF document, the mark's context agent mediates those three lines and other regions immediately around the lines. A user could retrieve broader context information for this mark, but the agent will not do so by default. Figure 3: SPARCE architecture reference model. A superimposed application allows creation of information elements (such as annotations) associated with marks. It can use an information model of its choice (SPARCE does not impose a model) and the model may vary from one application to another. For example, RIDPad uses a group-item model (simple nesting), whereas the Schematics Browser, another application we have built, uses an ER model [2, 12]. The superimposed model may be different from any of the base models. A detailed description of SPARCE is available in our previous work [12]. BI-LEVEL QUERY SYSTEM A bi-level query system allows a superimposed application and its user to query the superimposed information and base information as a whole. User queries are in a language appropriate to the superimposed model. For example, XQuery may be the query language if the superimposed model is XML (or a model that can be mapped to XML), whereas SQL may be the query language if superimposed information is in the relational model. Figure 4: Overview of a bi-level query system. Figure 4 provides an overview of a bi-level query system. An oval in the figure represents an information source. A rectangle denotes a process that manipulates information. Arrows indicate data flow. The query processor accepts three kinds of information--superimposed, mark, and context. Model transformers transform information from the three sources in to model(s) appropriate for querying. One of these transformers, the context transformer, is responsible for transforming context information . We restrict bi-level query systems to use only one superimposed model at a time, for practical reasons. Choosing a query language and the model for the result can be hard if superimposed models are mixed. Base Info 1 Base Info n Context Agents Model Transformers Mark Info Superimposed Info Query Processor Superimposed Application Superimposed Information Management Mark Management Context Management Clipboard Base Application Result Query 9 3.1 Implementation We have implemented a nave bi-level query system for the XML superimposed model. We have developed a transformer to convert RIDPad information to XML. We have developed a context transformer to convert context information to XML. We are able to use mark information without any transformation since SPARCE already represents that information in XML. User queries can be in XPath, XSLT, and XQuery. We use Microsoft's XML SDK 4.0 [10] and XQuery demo implementation [11] to process queries. We use three XML elements to represent RIDPad information in XML-&lt ;RIDPadDocument&gt; for the document, &lt;Group&gt; for a group, and &lt;Item&gt; for an item. For each RIDPad item, the system creates four children nodes in the corresponding &lt;Item&gt; element. These children nodes correspond to the mark, container (base document where the mark is made), application, and context. We currently transform the entire context of the mark. The XML data is regenerated if the RIDPad document changes. Figure 5: Partial XML data from a RIDPad document. Figure 5 shows partial XML data generated from the RIDPad document in Figure 1. It contains two &lt;Group&gt; elements (corresponding to the two groups in Figure 1). The `Garlic' element contains four &lt;Item&gt; elements (one for each item in that group in Figure 1). There is also an &lt;Item&gt; element for the group-less item CLIO. The &lt;Item&gt; element for `Goal' is partially expanded to reveal the &lt;Mark&gt; , &lt; Container&gt; , &lt; Application&gt; , and &lt; Context&gt; elements it contains. Contents of these elements are not shown. 3.2 Example Bi-level Queries We now provide bi-level query expressions for the queries Q1 to Q3 listed in Section 1. Q1: List base documents used in this RIDPad document. This query must retrieve the path to the base document of the mark associated with each item in a RIDPad document. The following XQuery expression does just that. The Location element in the Container element contains the path to the document corresponding to the mark associated with an item. &lt;Paths&gt; {FOR $l IN document(&quot;source&quot;)//Item/Container/Location RETURN &lt;Path&gt;{$l/text()}&lt;/Path&gt; } &lt;/Paths&gt; Q2: Show abstracts of papers related to Garlic. This query must examine the context of items in the group labeled `Garlic.' The following XPath expression suffices. This expression returns the text of a context element whose name attribute is `Abstract', but only for items in the required group. //Group[@name='Garlic']/Item/Context//Elemen t[@name='Abstract']/text() Q3: Create an HTML table of contents from the groups and items. We use an XSLT style-sheet to generate a table of contents (TOC) from a RIDPad document. Figure 6 shows the query in the left panel and its results in the right panel. The right panel embeds an instance of MS Internet Explorer. The result contains one list item (HTML LI tag) for each group in the RIDPad document. There is also one list sub-item (also an HTML LI tag) for each item in a group. The group-less item CLIO is in the list titled `Other Items.' A user can save the HTML results, and open it in any browser outside our system. Figure 6: RIDPAD document transformed to HTML TOC. The HTML TOC in Figure 6 shows that each item has a hyperlink (HTML A tag) attached to it. A hyperlink is constructed using a custom URL naming scheme and handled using a custom handler. Custom URLs are one means of implementing Capability 5 identified in Section 1. DISCUSSION The strength of the current implementation is that it retrieves context information for only those parts of base documents that the superimposed document refers to (via marks). Interestingly, the same is also its weakness: it retrieves context information for all parts of the base documents the superimposed document refers to, regardless of whether executing a query requires those elements . For example, only Query Q2 looks at context information (Q1 looks only at container information, Q3 looks at superimposed information and mark information). However, the XML data generated includes context information for all queries. Generating data in this manner is both inefficient and unnecessary-information may be replicated (different items may use the same mark), and context information can be rather large (the size of the complete context of a mark could exceed the size of its docu-10 ment), depending on what context elements a context agent provides. It is possible to get the same results by separating RIDPad data from the rest and joining the various information sources. Doing so preserves the layers, and potentially reduces the size of data generated. Also, it is possible to execute a query in-crementally and only generate or transform data that qualifies in each stage of execution. Figure 7 gives an idea of the proposed change to the schema of the XML data generated. Comparing with the Goal Item element of Figure 5, we see that mark, container, application, and context information are no longer nested inside the Item element. Instead, an &lt;Item&gt; element has a new attribute called markID . In the revised schema, the RIDPad data, mark, container, application, and context information exist independently in separate documents, with references linking them. With the revised schema, no context information would be retrieved for Query Q1. Context information would be retrieved only for items in the `Schematic Heterogeneity' group when Q2 is executed. Figure 7: XML data in the revised schema. Preserving the layers of data has some disadvantages. A major disadvantage is that a user will need to use joins to connect data across layers. Such queries tend to be error-prone, and writing them can take too much time and effort. A solution would be to allow a user to write bi-level queries as they currently do (against a schema corresponding to the data in Figure 5), and have the system rewrite the query to match the underlying XML schema (as in Figure 7). That is, user queries would actually be expressed against a view of the actual data. We are currently pursuing this approach to bi-level querying. Our current approach of grabbing context information for all marks could be helpful in some cases. For example, if a query workload ends up retrieving context of all (or most) marks, the current approach is similar to materializing views, and could lead to faster overall query execution. The current implementation does not exploit relationships between superimposed information elements. For example, Figure 8 shows the RIDPad document in Figure 1 enhanced with two relationships `Uses' and `Addresses' from the item CLIO. A user may exploit these relationships, to pose richer queries and possibly recall more information. For example, with the RIDPad document in Figure 8, a user could now pose the following queries: What system does CLIO use? How is CLIO related to SchemaSQL? Our initial use anticipated for bi-level queries was to query superimposed and base information as a whole, but we have noticed that superimposed application developers and users could use the capability to construct and format (on demand) superimposed information elements themselves. For example, a RIDPad item's name may be a section heading. Such a representation of an item could be expressed as the result of a query or a transformation. Figure 8: A RIDPad document with relationships. Bi-level queries could also be used for repurposing information. For example, Query Q3 could be extended to include the contents of items (instead of just names) and transform the entire RIDPad document to HTML (like in Figure 6). The HTML version can then be published on the web. We have demonstrated bi-level queries using XML query languages, but superimposed applications might benefit from other query languages. The choice of the query language depends largely on the superimposed information model (which in turn depends on the task at hand). More than one query language may be appropriate for some superimposed information models, in some superimposed applications. For example, both CXPath [3] and XQuery may be appropriate for some applications that use the XML superimposed model. The base applications we have worked with so far do not themselves have query capabilities. If access to context or a selection over context elements can be posed as a query in a base application, we might benefit from applying distributed query-processing techniques. Finally, the scope of a bi-level query is currently the superimposed layer and the base information accessible via the marks used. Some applications might benefit from including marks generated automatically (for example, using IR techniques) in the scope of a query. RELATED WORK SPARCE differs from mediated systems such as Garlic [4] and MIX [1]. Sources are registered with SPARCE simply by the act of mark creation in those sources. Unlike in Garlic there is no need to register a source and define its schema. Unlike MIX, SPARCE does not require a DTD for a source. 11 METAXPath [5] allows a user to attach metadata to XML elements . It enhances XPath with an `up-shift' operator to navigate from data to metadata (and metadata to meta-metadata, and so on). A user can start at any level, but only cross between levels in an upwards direction. In our system, it is possible to move both upwards and downwards between levels. METAXPath is designed to attach only metadata to data. A superimposed information element can be used to represent metadata about a base-layer element, but it has many other uses. CXPath [3] is an XPath-like query language to query concepts, not elements. The names used in query expressions are concept names, not element names. In the CXPath model there is no document root--all concepts are accessible from anywhere. For example, the CXPath expression `/Item' and `Item' are equivalent . They both return all Item elements when applied to the XML data in Figure 5. The `/' used for navigation in XPath follows a relationship (possibly named) in CXPath. For example, the expression "/Item/{Uses}Group" returns all groups that are related to an item by the `Uses' relationship when applied to an XML representation of the RIDPad in Figure 8. CXPath uses predefined mappings to translate CXPath expressions to XPath expressions. There is one mapping for each concept name and for each direction of every relationship of every XML source. In our system, we intend to support multiple sources without predefined mappings , but we would like our query system to operate at a conceptual level like CXPath does. As discussed in Section 4, preserving the layers of data, yet allowing a user to express queries as if all data is in one layer means queries are expressed against views. Information Manifold [7] provides useful insight in to how heterogeneous source may be queried via views. That system associates a capability record with each source to describe its inputs, outputs, and selection capabilities . We currently do not have such a notion in our system, but we expect to consider source descriptions in the context of distributed query processing mentioned in Section 4. SUMMARY Our existing framework for superimposed applications supports examination and manipulation of individual superimposed and base information elements. More global ways to search and manipulate information become necessary as the size and number of documents gets larger. A bi-level query system is a first step in that direction. We have an initial implementation of a query system, but still have a large space of design options to explore. ACKNOWLEDGMENTS This work was supported in part by US NSF Grant IIS-0086002. We thank all reviewers. REFERENCES [1] Baru, C., Gupta, A., Ludscher, B., Marciano, R., Papakonstantinou, Y., Velikhov, P., and Chu, V. XML-Based Information Mediation with MIX. In Proceedings of the SIGMOD conference on Management of Data (Philadelphia, June, 1999). ACM Press, New York, NY, 1999, 597-599. [2] Bowers, S., Delcambre, L. and Maier, D. Superimposed Schematics: Introducing E-R Structure for In-Situ Information Selections. In Proceedings of ER 2002 (Tampere, Finland, October 7-11, 2002). Springer LNCS 2503, 2002. 90104. [3] Camillo, S.D., Heuser, C.A., and Mello, R. Querying Heterogeneous XML Sources through a Conceptual Schema. In Proceedings of ER 2003 (Chicago, October 13-16, 2003). Springer LNCS 2813, 2003. 186199. [4] Carey, M.J., Haas, L.M., Schwarz, P.M., Arya, M., Cody, W.F., Fagin, R., Flickner, M., Luniewski, A.W., Niblack, W., Petkovic, D., Thomas, J., Williams, J.H., and Wimmers, E.L. Towards heterogeneous multimedia information systems: The Garlic approach. IBM Technical Report RJ 9911, 1994. [5] Dyreson, C.E., Bohlen, M.H., and Jensen, C.S. METAXPath. In Proceedings of the International Conference on Dublin Core and Metadata Applications (Tokyo, Japan, October 2001). 2001, 17-23. [6] Halasz, F.G., and Schwartz, F. The Dexter Hypertext Reference Model. Communications of the ACM, 37, 2, 30-39 . [7] Levy, A.Y., Rajaraman, A., and Ordille, J.J. Querying heterogeneous information sources using source descriptions. In Proceedings of VLDB (Bombay, India 1996). 251-262. [8] Maier, D., and Delcambre, L. Superimposed Information for the Internet. In Informal Proceedings of WebDB '99 (Philadelphia, June 3-4, 1999). 1-9. [9] Microsoft. COM: The Component Object Model Specification, Microsoft Corporation. 1995. [10] Microsoft. MS XML 4.0 Software Development Kit. Microsoft Corporation. Available online at http://msdn.microsoft.com/ [11] Microsoft. XQuery Demo. Microsoft Corporation. Available online at http://xqueryservices.com/ [12] Murthy, S., Maier, D., Delcambre, L., and Bowers, S. Putting Integrated Information in Context: Superimposing Conceptual Models with SPARCE. In Proceedings of the First Asia-Pacific Conference of Conceptual Modeling (Dunedin, New Zealand, Jan. 22, 2004). 71-80. [13] Wiederhold, G. Mediators in the architecture of future information systems. IEEE Computer, 25, 3 (March 1992). 3849. 12

Bi-level queries;implementation;system;Superimposed information management;SPARCE;superimposed;document;management;RIDPAD;query;information;Information integration;METAXPath;hyperlink

163

Ranking Flows from Sampled Traffic

Most of the theoretical work on sampling has addressed the inversion of general traffic properties such as flow size distribution , average flow size, or total number of flows. In this paper, we make a step towards understanding the impact of packet sampling on individual flow properties. We study how to detect and rank the largest flows on a link. To this end, we develop an analytical model that we validate on real traces from two networks. First we study a blind ranking method where only the number of sampled packets from each flow is known. Then, we propose a new method, protocol-aware ranking, where we make use of the packet sequence number (when available in transport header) to infer the number of non-sampled packets from a flow, and hence to improve the ranking. Surprisingly, our analytical and experimental results indicate that a high sampling rate (10% and even more depending on the number of top flows to be ranked) is required for a correct blind ranking of the largest flows. The sampling rate can be reduced by an order of magnitude if one just aims at detecting these flows or by using the protocol-aware method.

INTRODUCTION The list of the top users or applications is one of the most useful statistics to be extracted from network traffic. Network operators use the knowledge of the most popular destinations to identify emerging markets and applications or to locate where to setup new Points of Presence. Content delivery networks use the popularity of sites to define caching and replication strategies. In traffic engineering, the identification of heavy hitters in the network can be used to treat and route them differently across the network [20, 17, 10]. Keeping track of the network prefixes that generate most traffic is also of great importance for anomaly detection . A variation in the pattern of the most common applications may be used as a warning sign and trigger careful inspection of the packet streams. However, the ability to identify the top users in a packet stream is limited by the network monitoring technology. Capturing and processing all packets on high speed links still remains a challenge for today's network equipment [16, 9]. In this context, a common solution is to sample the packet stream to reduce the load on the monitoring system and to simplify the task of sorting the list of items. The underlying assumption in this approach is that the sampling process does not alter the properties of the data distribution. Sampled traffic data is then used to infer properties of the original data (this operation is called inversion). The inversion of sampled traffic is, however, an error-prone procedure that often requires a deep study of the data distribution to evaluate how the sampling rate impacts the accuracy of the metric of interest. Although the inversion may be simple for aggregate link statistics (e.g., to estimate the number of packets transmitted on a link, it is usually sufficient to multiply the number of sampled packets by the inverse of the sampling rate), it is much harder for the properties of individual connections or "flows" [9, 11, 8]. For these reasons, in this paper, we address this simple, and so far unanswered, question: which sampling rate is needed to correctly detect and rank the flows that carry the most packets? We define the problem as follows. Consider a traffic monitor that samples packets independently of each other with probability p (random sampling) and classifies them into sampled flows. At the end of the measurement period, the monitor processes the list of sampled flows, ranks them based on their size in packets, and returns an ordered list of the t largest flows. We are interested in knowing (i) whether the ordered list contains all the actual largest flows in the original packet 188 stream (detection), and (ii) if the items in the list appear in the correct order (ranking). We build an analytical model and define a performance metric that evaluates the accuracy of identification and ranking of the largest flows. We consider a flow to consist of a single TCP connection. However, our results are general and can be applied to alternative definitions of flow, as well. We evaluate two approaches to sort the list of flows: (i) Blind, where the sampled flows are ranked just based on their sampled size. This method can be applied to any definition of flow. (ii) Protocol-aware, where we make use of additional information in the packet header (e.g., the sequence number in TCP packets) to infer the number of non-sampled packets between sampled ones. This method can only be applied to flow definitions that preserve the protocol level details. The contributions of this work are the following: (1) We perform an analytical study of the problem of ranking two sampled flows and compute the probability that they are misranked. We propose a Gaussian approximation to make the problem numerically tractable. (2) We introduce the protocol-aware ranking method that uses protocol level information to complement the flow statistics and render the detection and ranking of the largest flows more accurate. (3) Based on the model for the ranking of two flows, we propose a general model to study the detection and ranking problem, given a generic flow size distribution. We define a performance metric and evaluate the impact of several metric's parameter on the accuracy of the ranking. (4) We validate our findings on measurement data using publicly-available packet-level traces. Our results indicate that a surprisingly high sampling rate is required to obtain a good accuracy with the blind approach (10% and even more depending on the number of flows of interest). As for the protocol-aware approach, it allows to reduce the required sampling rate by an order of magnitude compared to the blind approach. The paper is structured as follows. Next, we discuss the related literature. In Section 3 and 4, we present our model. Section 5 analyzes the model numerically and Section 6 validates it on real packet-level traces. Section 7 concludes the paper and provides perspectives for our future research. RELATED WORK The inversion of sampled traffic has been extensively studied in the literature. The main focus has been on the inversion of aggregate flow properties such as flow size distribution [9, 11], average flow size or total number of flows [8] on a given network link. Duffield et al. [8] study the problem of flow splitting and propose estimators for the total number of flows and for the average flow size in the original traffic stream. [9, 11] study the inversion of the flow size distribution with two different methods. They both show that the major difficulty comes from the number of flows that are not sampled at all and that need to be estimated with an auxiliary method. As an auxiliary method, [8, 9] propose the use of the SYN flag in the TCP header to mark the beginning of a flow. [9] shows that periodic and random sampling provide roughly the same result on high speed links, and so random sampling can be used for mathematical analysis due to its appealing features. [4] finds the sampling rate that assures a bounded error on the estimation of the size of flows contributing to more than some predefined percentage of the traffic volume. [14] studies whether the number of sampled packets is a good estimator for the detection of large flows without considering its impact on the flow ranking. Given the potential applications of finding the list of top users, it does not come as a surprise that there has been a significant effort in the research community to find ways to track frequent items in a data stream [5, 7, 3, 10]. However, this problem has usually been addressed from a memory requirement standpoint. All the works in the literature assume that if the algorithm and the memory size is well chosen, the largest flows can be detected and ranked with a high precision . However, in the presence of packet sampling, even if the methods rank correctly the set of sampled flows, there is no guarantee that the sampled rank corresponds to the original rank. The problem we address in this paper complements these works as it focuses on the impact of sampling on the flow ranking. BASIC MODEL RANKING TWO FLOWS In this section, we study the probability to misrank two flows of original sizes S 1 and S 2 in packets. This probability is the basis for the general model for detecting and ranking the largest flows that we will present later. Indeed, the detection and ranking of the largest flows can be transformed into a problem of ranking over a set of flow pairs. Without loss of generality, we assume S 1 &lt; S 2 . We consider a random sampling of rate p. Let s 1 and s 2 denote the sizes in packets of both flows after sampling. The two sampled flows are misranked if (i) s 1 is larger than s 2 , or (ii) both flows are not sampled, i.e., their sampled sizes equal to zero. By combining (i) and (ii), one can see that the necessary condition for a good ranking is to sample at least one packet from the larger flow (i.e., the smaller of the two flows can disappear after sampling). The probability to misrank the two flows can then be written as P m (S 1 , S 2 ) = P {s 1 s 2 }. For the case S 1 = S 2 , we consider the two flows as misranked if s 1 = s 2 , or if both flows are not sampled at all, i.e. s 1 = s 2 = 0. We compute and study the misranking probability of two flows of given sizes in the rest of this section. First, we consider the blind ranking method where only the number of sampled packets from a flow is known. For this method, we express the misranking probability as a double sum of binomials, then we present a Gaussian approximation to make the problem tractable numerically. Second, we consider the protocol-aware ranking method for which we calculate a numerical-tractable closed-form expression of the misraking probability. Note that the misranking probability is a symmetric function, i.e., P m (S 1 , S 2 ) = P m (S 2 , S 1 ). 3.1 Blind ranking With this method, s 1 and s 2 represent the number of sampled packets from flows S 1 and S 2 . Under our assumptions , these two variables are distributed according to a binomial distribution of probability p. Hence, we can write for S 1 &lt; S 2 , P m (S 1 , S 2 ) = P {s 1 s 2 } = S 1 i=0 b p (i, S 1 ) i j=0 b p (j, S 2 ). (1) b p (i, S) is the probability density function of a binomial distribution of probability p, i.e., the probability of obtaining i successes out of S trials. We have b p (i, S) = S i p i (1 - p) S-i for i = 0, 1, ..., S, and b p (i, S) = 0 for i &lt; 0 and i &gt; S. The 189 probability to misrank two flows of equal sizes is given by P {s 1 = s 2 or s 1 = s 2 = 0} = 1 - P {s 1 = s 2 = 0} = 1 S 1 i=1 b 2 p (i, S 1 ). Unfortunately, the above expression for the misranking probability is numerically untractable since it involves two sums of binomials. For large flows of order S packets, the number of operations required to compute such a probability is on the order of O(S 3 ), assuming that the complexity of the binomial computation is on the order of O(S). The problem becomes much more complex if one has to sum over all possible flow sizes (i.e., O(S 5 )). For this reason, we propose next a Gaussian approximation to the problem of blind ranking that is accurate and easy to compute. We use this approximation to study the ranking performance as a function of the sampling rate and the flow sizes. 3.1.1 Gaussian approximation to blind ranking Consider a flow made of S packets and sampled at rate p. The sampled size follows a binomial distribution. However , it is well known that the binomial distribution can be approximated by a Normal (or Gaussian) distribution when p is small and when the product pS is on the order of one (flows for which, on average, at least few packets are sampled ) [21, pages 108109]. We assume that this is the case for the largest flows, and we consider the sampled size of a flow as distributed according to a Normal distribution of average pS and of variance p(1 - p)S. Using this approximation , one can express the misranking probability for the blind ranking problem in the following simple form. Proposition 1. For any two flows of sizes S 1 and S 2 packets (S 1 = S 2 ), the Gaussian approximation gives, P m (S 1 , S 2 ) 1 2 erf c |S 2 - S 1 | 2(1/p - 1)(S 1 + S 2 ) , (2) where erfc(x) = ( 2 ) x e -u 2 du is the complementary error cumulative function. Proof: Consider two flows of sizes S 1 and S 2 in packets such that S 1 &lt; S 2 . Their sampled versions s 1 and s 2 both follow Normal distributions of averages pS 1 and pS 2 , and of variances p(1 - p)S 1 and p(1 - p)S 2 . We know that the sum of two Normal variables is a Normal variable. So the difference s 1 - s 2 follows a Normal distribution of average p(S 1 - S 2 ) and of variance p(1 - p)(S 1 + S 2 ). We have then this approximation for the misranking probability: P m (S 1 , S 2 ) = P {s 1 - s 2 0} P V &gt; p(S 2 - S 1 ) p(1 - p)(S 1 + S 2 ) = 1 2 erfc S 2 - S 1 2(1/p - 1)(S 1 + S 2 ) . (3) V is a standard Normal random variable. Given the symmetry of the misranking probability, one can take the absolute value of S 2 - S 1 in (3) and get the expression stated in the proposition, which is valid for all S 1 and S 2 . For S 1 = S 2 , one can safely approximate the misranking probability to be equal to 1. This approximation is however of little importance given the very low probability of having two flows of equal sizes, especially when they are large. 3.2 Protocol-aware ranking Packets can carry in their transport header an increasing sequence number. A typical example is the byte sequence number in the TCP header. Another example could be the sequence number in the header of the Real Time Protocol (RTP) [19]. One can use this sequence number, when available , to infer the number of non-sampled packets (or bytes in the case of TCP) between sampled ones, and hence to improve the accuracy of ranking. The size of the sampled flow in this case is no longer the number of packets collected, but rather the number of packets that exist between the first and last sampled packets from the flow. Although this solution is limited to flows whose packet carry a sequence number, we believe that the study of this ranking method is important given the widespread use of the TCP protocol. Our objective is to understand how the use of protocol-level information can supplement the simple, and more general, blind method and if it is worth the additional overhead it introduces (i.e., storing two sequence numbers per flow record). In the following, we calculate the misranking probability of two flows of given sizes when using the protocol-aware method. This probability will be used later in the general ranking problem. The main contribution of this section is a closed-form expression for the misranking probability that is numerically tractable, without the need for any approximation . Let S be the size of a flow in packets. Let s b , s b = 1, 2, ..., S, denote the (packet) sequence number carried by the first sampled packet, and let s e , s e = S, S - 1, ..., s b , denote the sequence number carried by the last sampled packet. Given s b and s e , one can estimate the size of the sampled flow in packets to s = s e - s b + 1. The error in this estimation comes from the non-sampled packets that are transmitted before s b and after s e . We give next the distribution of s, which is needed for the computation of the misranking probability, then we state our main result. Before presenting the analysis, note that this new flow size estimator only counts the packets that are transmitted with distinct sequence numbers. In the case of TCP, this corresponds to the number of bytes received at the application layer, rather then the number of bytes carried over the network . It is equivalent to assuming that the probability of sampling a retransmitted (or duplicated) packet is negligible . This is a reasonable assumption if the loss rate is low. We will address this aspect in more detail in Section 6. Consider a flow of size S 2 in packets. Using the above definition for s, the sampled flow has a size of i packets, i 2, with probability: P {s = i} = S-i+1 k=1 P {s b = k} P {s e = k + i - 1} . We have P {s b = k} = (1 - p) k-1 p, and P {s e = k + i - 1} = (1 - p) S-k-i+1 p. This gives P {s = i} = S-i+1 k=1 (1 - p) k-1 p(1 - p) S-k-i+1 p = p 2 (1 - p) S-i (S - i + 1). (4) As for i = 0, we have P {s = 0} = (1 - p) S for S 1. And for i = 1, we have P {s = 1} = p(1 - p) S-1 S for S 1. It is easy to prove that the cumulative distribution of s is the 190 following for all values of S: P {s i = 0} = p(1 - p) S-i (S - i + 1) + (1 - p) S-i+1 . (5) We come now to the misranking probability, which we recall is a symmetric function. For S 1 &lt; S 2 , we have P m (S 1 , S 2 ) = P {s 2 s 1 } = S 1 i=0 P {s 1 = i} i j=0 P {s 2 = j} . (6) And for S 1 = S 2 , we have P m (S 1 , S 2 ) = 1 S 1 i=1 P {s 1 = i} 2 . (7) Our main result is the following. Proposition 2. For S 1 &lt; S 2 , the misranking probability is equal to P m (S 1 , S 2 ) = (1 - p) S 1 (1 - p) S 2 + p(1 - p) S 1 -1 S 1 [p(1 - p) S 2 -1 S 2 + (1 - p) S 2 ] + p 3 2 F (1 - p, 1 - p) xy + p 2 F (1 - p, 1 - p) x , where F (x, y) = xy S 2 -S 1 +1 + ... + x S 1 -1 y S 2 -1 = xy S 2 -S 1 +1 (1 - (xy) S 1 -1 )/(1 - xy). For S 1 = S 2 = S, the misranking probability is equal to P m (S, S) = 1 - p 2 (1 - p) 2(S-1) S 2 - p 4 2 G(1 - p, 1 - p) xy , where G(x, y) = xy + x 2 y 2 + x S-1 y S-1 = (xy - (xy) S )/(1 - xy). Proof: One can validate the results by plugging (4) and (5) into (6) and (7). Note that the main gain of writing the misraking probability in such a condensed form is a complexity that drops from O(S 3 ) in (6) to O(S) in our final result. This gain comes from the closed-form expression for the cumulative distribution in (5), and from introducing the two functions F (x, y) and G(x, y). These two latter functions transform two series whose complexity is O(S 2 ) into a closed-form expression whose complexity is O(S). We solve the derivatives in the above equations using the symbolic toolbox of matlab, which gives explicit expressions for the misranking probability. These expressions are simple to compute, but span on multiple lines, so we omit them for lack of space. 3.3 Analysis of the misranking probability 3.3.1 The blind case We use the Gaussian approximation to study how the misranking probability varies with the sampling rate and with the sizes of both flows, in particular their difference. The study of the impact of the flow sizes is important to understand the relation between flow size distribution and ranking of the largest flows. The misranking probability is a decreasing function of the sampling rate. It moves to zero when p moves to 1 and to 0.5 when p approaches zero 1 . Therefore, there exists one sampling rate that leads to some desired misranking probability, and any lower sampling rate results in larger error. We study now how the misranking probability varies with the sizes of both flows. Take S 1 = S 2 - k, k a positive integer. From (2) and for fixed k, the misranking probability increases with S 1 and S 2 (erfc(x) is an increasing function in x). This indicates that it is more difficult to rank correctly two flows different by k packets as their sizes increase in absolute terms. The result is different if we take the size of one flow equal to &lt; 1 times the size of the second, i.e., S 1 = S 2 . Here, (S 1 - S 2 )/S 1 + S 2 is equal to S 1 (1 )/1 + , which increases with S 1 . Hence, the misranking probability given in (2) decreases when S 1 increases. We conclude that, when the two flow sizes maintain the same proportion, it is easier to obtain a correct ranking when they are large in absolute terms. We can now generalize the result above. One may think that the larger the flows, the better the ranking of their sampled versions. Our last two examples indicate that this is not always the case. The ranking accuracy depends on the relative difference of the flow sizes. In general, to have a better ranking, the difference between the two flow sizes must increase with the flow sizes and the increase must be larger than a certain threshold. This threshold is given by (2): the difference must increase at least as the square root of the flow sizes. This is an interesting finding. In the context of the general ranking problem, it can be interpreted as follows. Suppose that the flow size has a cumulative distribution function y = F (x). As we move to the tail of the distribution 2 , the size of the flows to be ranked increases. The ranking performance improves if the difference between flow sizes increases faster than x. This is equivalent to saying that dx/dy should increase with x faster than x. All common distributions satisfy this condition, at least at their tails. For example, with the exponential distribution we have dx/dy e x (1/ is the average), while for the Pareto distribution we have dx/dy x +1 ( is the shape). 3.3.2 The protocol-aware case The first difference with the blind case is in the estimation error (S - s = s b - 1 + S - s e ), which can be safely assumed to be independent of the flow size for large flows (only dependent on p). This means that if two large flows keep the same distance between them while their sizes increase, their ranking maintains the same accuracy. Their ranking improves if the difference between their sizes increases as well, and it deteriorates if the difference between their sizes decreases . So in contrast to the blind case, the threshold for the ranking here to improve is that the larger flow should have its size increasing a little faster than the smaller one. In the context of the general ranking problem where flow sizes are distributed according to a cumulative distribution function y = F (x), and when the top flows become larger, the protocol-aware ranking improves if the derivative dx/dy increases with x. This is equivalent to saying that the function F (x) should be concave, which is satisfied by most common distributions at their tail. For blind ranking, concavity was 1 The Gaussian approximation does not account for the case p = 0 where the misranking probability should be equal to 1 based on our definition. 2 Because we are more and more focusing on large flows or because the number of available flows for ranking increases. 191 not enough to obtain a better ranking; the derivative dx/dy had to increase faster than x. So in conclusion, the condition to have a better ranking when we move to the tail of the flow size distribution is less strict with the protocol-aware method, which is an indication of its good performance. The second difference with the blind case is in the relation between the ranking accuracy and the sampling rate. Consider two large flows of sizes S 1 and S 2 in packets, and let s 1 and s 2 denote their sampled sizes. The coefficient of variation of the difference s 2 - s 1 is an indication on how well the ranking performs (a small coefficient of variation results in better ranking 3 ). It is easy to prove that this coefficient of variation scales as 1/p for protocol-aware ranking and as 1/p for blind ranking. This is again an important finding. It tells that when the sampling rate is very small, blind ranking could (asymptotically) perform better than protocol-aware ranking. Our numerical and experimental results will confirm this finding. GENERAL MODEL DETECTING AND RANKING THE LARGEST FLOWS We generalize the previous model from the ranking of two flows to the detection and ranking of the top t flows, t = 1, 2, . . . , N . The misranking probability P m (S 1 , S 2 ) pre-viously calculated is the basis for this generalization. Let N t denote the total number of flows available in the measurement period before sampling. We want the sampled list of top t flows to match the list of top t flows in the original traffic. Two criteria are considered to decide whether this match is accurate. First, we require the two lists to be identical . This corresponds to the ranking problem. The second, less constrained, criterion requires the two lists to contain the same flows regardless of their relative order within the list. This corresponds to the detection problem. For both problems, the quality of the result is expressed as a function of the sampling rate p, the flow size distribution, the number of flows to rank t, and the total number of flows N . 4.1 Performance metric In order to evaluate the accuracy of detection and ranking , we need to define a performance metric that is easy to compute and that focuses on the largest flows. A flow at the top of the list can be misranked with a neighboring large flow or a distant small flow. We want our metric to differentiate between these two cases and to penalize more the latter one; a top-10 flow replaced by the 100-th flow in the sampled top list is worse than the top-10 flow being replaced by the 11-th flow. We also want our metric to be zero when the detection and ranking of the top flows are correct. We introduce our performance metric using the ranking problem. The performance metric for the detection problem is a straightforward extension. Let's form all flow pairs where the first element of a pair is a flow in the top t and the second element is anywhere in the sorted list of the N original flows. The number of these pairs is equal to N - 1 + N - 2 + + N - t = (2N - t - 1)t/2. We then count the pairs in this set that are misranked after sampling and we take the sum as our metric for ranking accuracy. This 3 For S 1 &lt; S 2 , we are interested in P {s 1 s 2 }. According to Tchebychev inequality, this probability can be supposed to behave like VAR[s 1 - s 2 ]/E [s 1 - s 2 ] 2 , which is the square of the coefficient of variation. sum indicates how good the ranking is at the top of the list. It is equal to zero when the ranking is correct. When the ranking is not correct, it takes a value proportional to the original rank of the flows that have taken a slot in the top-t list. For example, if the top flow is replaced by its immediate successor in the list, the metric will return a ranking error of 1. Instead, if the same flow is replaced by a distant flow, say the 100-th, the metric will return an error of 99. Also, note that our metric does not account for any misranking of flows outside the list of top t flows. For any two flows n and m, such that n &gt; m &gt; t, the fact that n takes the position of m does not add anything to our performance metric since our metric requires at least one element of a flow pair to be in the original list of top t flows. In the detection problem, we are no longer interested in comparing flow pairs whose both elements are in the top t list. We are only interested in the ranking between flows in the top t list and those outside the list. Therefore, our detection metric is defined as the number of misranked flow pairs, where the first element of a pair is in the list of top t flows and the second element is outside this list (non top t). The above metrics return one value for each realization of flow sizes and of sampled packets. Given that we want to account for all realizations, we define the performance metrics as the number of misranked flow pairs averaged over all possible values of flow sizes in the original list of N flows and over all sampling runs. We deem the ranking/detection as acceptable when our metric takes a value below one (i.e., on average less than one flow pair is misranked). In addition to the above, our metrics have the advantage of being easily and exactly calculable. Performance metrics based on probabilities (e.g.,[12]) require lot of assumptions that make them only suitable for computing bounds, but not exact values. 4.2 Computation of the performance metric for the ranking problem Consider a flow of i packets belonging to the list of top t flows in the original traffic (before sampling). First, we compute the probability that this flow is misranked with another flow of general size and general position. Denote this probability by P mt (i), where m stands for misranking and t for top. Then, we average over all values of i to get P mt4 . This latter function gives us the probability that, on average, the top t-th flow is misranked with another flow. Thus, our performance metric, which is defined as the average number of misranked flow pairs where at least one element of a pair is in the top t, is equal to (2N - t - 1)t P mt /2. Next, we compute the value of P mt . Let p i denote the probability that the size of a general flow is equal to i packets, and P i denote the flow size complementary cumulative distribution, i.e., P i = j=i p j . For a large number of flows N and a high degree of multiplexing, we consider safe to assume that flow sizes are independent of each other (see [2] for a study of the flow size correlation on a OC-12 IP backbone link). A flow of size i belongs to the list of top t flows if the number of flows in the original total list, with a size larger than i, is less or equal than t - 1. Since each flow can be larger than i with probability P i independently of the other flows, we can write the probability that a flow of size i belongs to the list of the top t flows 4 Note that the distribution of the size of a flow at the top of the list is different from that of a generic flow. 192 as P t (i, t, N ) = t-1 k=0 b P i (k, N - 1), where b P i (k, N - 1) is the probability to obtain k successes out of N - 1 trials , P i being the probability of a success. The probability that the t-th largest flow has a size of i packets is equal to P t (i) = p i P t (i, t, N )/ P t (t, N ). P t (t, N ) is the probability that a flow of general size is among the top t in the original total list, which is simply equal to t/N . Using the above notation, one can write the misranking probability between a top t flow of original size i packets and any other flow as follows P mt (i) = 1 P t (i, t, N ) i-1 j=1 p j P t (i, t, N - 1)P m (j, i)+ j=i p j P t (i, t - 1, N - 1)P m (i, j) . (8) In this expression, we sum over all possible original sizes of the other flow (the variable j) and we separate the case when this other flow is smaller than i from the case when it is larger than i 5 . P m (i, j) is the misranking probability of two flows of sizes i and j packets, which we calculated in the previous section for the two ranking methods. P mt is then equal to i=1 P t (i)P mt (i). For protocol-aware ranking, P m (i, j) is given explicitly in Proposition 2 and can be easily computed. For blind ranking, we use the Gaussian approximation summarized in Proposition 2, which we recall holds when at least one of the two flows to be compared is large. 4.3 Computation of the performance metric for the detection problem Consider the probability that a flow among the top t is swapped with a flow that does belong to the top t. Let P mt denote this probability. Following the same approach described in Section 4, we can write P mt = 1 P t i=1 i-1 j=1 p i p j P t (j, i, t, N )P m (j, i). To get this expression for P mt , we sum over all possible values for the size of the flow in the top t (index i) and all possible values for the size of the other flow not among the top t (index j). In this expression, p i and p j represent the probability that the size of a flow is equal to i or j packets, respectively. P m (j, i) is the probability that two flows of sizes i and j are misranked it is given by the Gaussian approximation described in Proposition 1 for the blind method and the result stated in Proposition 2 for the protocol-aware method. P t (j, i, t, N ) is the joint probability that a flow of size i belongs to the list of the top t flows while another flow of size j does not belong to it (i.e., it is in the bottom N - t flows). P t is the joint probability that a flow of any size belongs to the list of the top t flows while another flow of any size does not belong to this list. It is equal to t(N - t)/(N (N - 1)). We now compute P t (j, i, t, N ) for j &lt; i, i.e., the probability that flow i belongs to the top list while flow j does not. The number of flows larger than i should be smaller than t, while the number of flows larger than j should be larger than t. The probability that a flow size is larger than 5 In the case j i, at most t - 2 flows can be larger than i packets if we want the flow of size i to be in the top t. Trace Jussieu Abilene Link speed GigE (1 Gbps) OC-48 (2.5 Gbps) Duration 2 hours 30 minutes TCP connections 11M 15M Packets 112M 125M Table 1: Summary of the traces i is P i = k=i p k . The probability that it is larger than j is P j = k=j p k . The probability that a flow size is between j and i given that it is smaller than i is (P j - P i )/(1 - P i ). We call it P j,i . It follows that: P t (j, i, t, N ) = t-1 k=0 b P i (k, N - 2) N -k-2 l=t-k-1 b P j,i (l, N - k - 2). The first sum accounts for the probability to see less than t flows above i packets. The second sum accounts for the probability to see more than t flows above j given that k flows (k &lt; t) were already seen above i. For t = 1, P t (j, i, t, N ) is no other than P t (i, t, N - 1), and both P mt and P mt are equal (i.e., the ranking and the detection problems are the same). Once P mt is computed, we multiply it by the total number of flow pairs whose one element is in the top t and the other one is not. This total number is equal to t(N -t). Our metric for the detection problem is the result of this multiplication. As for the ranking problem, we want this metric to be less than one for the detection of the top t flows to be accurate. NUMERICAL RESULTS We analyze now the accuracy of identifying and ranking the largest flows in a packet stream for both the blind and protocol-aware methods. Our metrics require the following input: p i , the flow size distribution and N , the total number of flows observed on the link during the measurement period. To derive realistic values for these two quantities, we consider two publicly available packet-level traces. The first trace is Abilene-I collected by NLANR [15] on an OC-48 (2.5 Gbps) link on the Abilene Network [1]. The second trace has been collected by the Metropolis project [13] on a Gigabit Ethernet access link from the Jussieu University campus in Paris to the Renater Network [18]. Table 1 summarizes the characteristics of the two traces. We model the flow size distribution in the traces with Pareto. We opted for Pareto since it is known to be appropriate to model flow sizes in the Internet due to its heavy tailed feature [6]. Note that it is not our goal to find an accurate approximation of the distribution of flow sizes in our traces, but rather to find a general, well-known, distribution that approaches the actual flow size. In this section we analyze a wide range of parameters while Section 6 focuses on the performance we observe in the two packet-level traces. The Pareto distribution is continuous with a complementary cumulative distribution function given by P {S &gt; x} = (x/a) . &gt; 0 is a parameter describing the shape of the distribution and a &gt; 0 is a parameter describing its scale. The Pareto random variable takes values larger than a, and has an average value equal to a/( - 1). The tail of the Pareto distribution becomes heavier as decreases. We use our traces to derive an indicative value of the shape parameter . To this end, we compute the empirical complementary cumulative distribution of flow sizes and we 193 10 0 10 1 10 2 10 3 10 4 10 5 10 -8 10 -7 10 -6 10 -5 10 -4 10 -3 10 -2 10 -1 10 0 Flow size in Kbytes Complementary CDF Abilene trace x -2 10 0 10 1 10 2 10 3 10 4 10 5 10 -7 10 -6 10 -5 10 -4 10 -3 10 -2 10 -1 10 0 Flow size in Kbytes Complementary CDF Jussieu trace x -1.5 Figure 1: Empirical flow size distribution plot it on a log-log scale. A heavy-tailed distribution of shape parameter decays linearly on a log-log scale at rate -. The empirical distributions are shown in Figure 1. The plots show that equal to 2 suits the Abilene trace and equal to 1.5 suits the Jussieu one. This means that the flow size distribution has a heavier tail in the Jussieu trace. Then, we compute the average flow size in packets to get the starting point a for the Pareto distribution. As an average flow size we measure 5.76 Kbytes and 7.35 packets on the Abilene trace, and 9.22 Kbytes and 9.9 packets on the Jussieu trace. The total number of flows N is set by taking a measurement interval equal to one minute, then multiplying this interval by the average arrival rate of flows per second on each trace. This gives N = 487 Kflows for the Abilene trace and N = 103 Kflows for the Jussieu one. In the rest of this section, all figures plot the ranking metric versus the packet sampling rate p on a log-log scale. We vary p from 0.1% to 50%. Each figure shows different lines that correspond to different combinations of t, , and N . We are interested in the regions where the value of the metric is below one, indicating that the ranking is accurate on average. To ease the interpretation of results in the figures, we plot the horizontal line of ordinate 1. 5.1 Blind ranking 5.1.1 Impact of the number of flows of interest The first parameter we study is t, the number of largest flows to rank. The purpose is to show how many flows can be detected and ranked correctly for a given sampling rate. We set , N , and the average flow size to the values described before. The performance of blind ranking the top t flows is shown in Figure 2 for both traces. We observe that the larger the number of top flows of interest, the more difficult it is to detect and rank them correctly. In particular, with a sampling rate on the order of 1%, it is possible to rank at most the top one or two flows. As we focus at larger values of t, the required sampling rate to get a correct ranking increases well above 10%. Note that with a sampling rate on the order of 0.1%, it is almost impossible to detect even the largest flow. We also observe that the ranking on the Jussieu trace behaves slightly better than that on the Abilene trace. The Jussieu trace has a heavier tail for its flow size distribution, and so the probability to get larger flows at the top of the list is higher, which makes the ranking more accurate. This will be made clear next as we will study the impact of the shape parameter . 5.1.2 Impact of the flow size distribution We consider the blind ranking of the top 10 flows varying 10 -1 10 0 10 1 10 -2 10 0 10 2 10 4 10 6 Packet sampling rate (%) Average number of misranked flow pairs Abilene trace, N=487K, beta=2, blind ranking t=25 t=10 t=5 t=2 t=1 10 -1 10 0 10 1 10 -2 10 0 10 2 10 4 10 6 Packet sampling rate (%) Average number of misranked flow pairs Jussieu trace, N=103K, beta=1.5, blind ranking t=25 t=10 t=5 t=2 t=1 Figure 2: Performance of blind ranking varying the number t of top flows of interest the shape parameter for the Pareto distribution among five distinct values: 3, 2.5, 2, 1.5 and 1.2. Note that for 2 the Pareto distribution is known to be heavy tailed (infinite variance). The other parameters of the model (N and the average flow size) are set as before. The values taken by our metric are shown in Figure 3 for both traces. We can make the following observations from the figure: Given a sampling rate, the ranking accuracy improves as becomes smaller, i.e., the tail of the flow size distribution becomes heavier. Indeed, when the distribution tail becomes heavier, the probability to obtain larger flows at the top of the list increases, and since it is simpler to blindly rank larger flows (for distributions satisfying the square root condition, see Section 3.1.1), the ranking becomes more accurate. The ranking is never correct unless the sampling rate is very high. In our setting, one needs to sample at more than 50% to obtain an average number of misranked flow pairs below one for a value of equal to 1.5 (i.e, heavy tailed distribution), and at more than 10% for a value of equal to 1.2 (i.e., pronounced heavy tailed distribution). For larger values of (i.e., lighter tail), the sampling rate needs to be as high as 100%. 5.1.3 Impact of the total number of flows Another important parameter in the ranking problem is N , the total number of flows available during the measurement period. When N increases, the flows at the top of the list should become larger, and therefore as we saw in Section 3.1.1, the blind ranking accuracy should improve 194 10 -1 10 0 10 1 10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6 10 7 Packet sampling rate (%) Average number of misranked flow pairs Abilene trace, N = 487K, t = 10 flows, blind ranking beta=3 beta=2.5 beta=2 beta=1.5 beta=1.2 10 -1 10 0 10 1 10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6 10 7 Packet sampling rate (%) Average number of misranked flow pairs Jussieu trace, N = 103K, t = 10 flows, blind ranking beta=3 beta=2.5 beta=2 beta=1.5 beta=1.2 Figure 3: Performance of blind ranking varying the shape parameter of the flow size distribution for flow size distributions satisfying the square root condition (in particular the Pareto distribution we are considering here). N varies with the utilization of the monitored link the higher the utilization, the larger the number of flows. N can also vary with the duration of the measurement period the longer we wait before ranking and reporting results, the larger the number of flows. We study the impact of N on the blind ranking accuracy . We take the same value of N used in the previous sections and computed over one minute measurement period (487 Kflows for the Abilene trace and 103 Kflows for the Jussieu trace), then we multiply it by some constant factor ranging from 0.5 (2 times fewer flows) to 5 (5 times more flows). Results are shown in Figure 4. The lines in the figures correspond to a factor value equal to: 0.5, 1, 2.5, and 5. In these figures, we consider the ranking of the top 10 flows with the values of and average flow size set from the traces. Clearly, the ranking accuracy improves as N increases. However, in our setting, this improvement is still not enough to allow a perfect ranking. One can always imagine increasing N (e.g., by increasing the measurement period) until the top t flows are extremely large and hence, perfectly detected and ranked. 5.2 Protocol-aware ranking Protocol-aware ranking takes advantage of the information carried in the transport header of the sampled packets to infer the number of non-sampled packets of a flow. We use our model to check whether this improvement exists and to evaluate it. Remember that we are always in the context of low retransmission and duplication rates, which is neces-10 -1 10 0 10 1 10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6 Packet sampling rate (%) Average number of misranked flow pairs Abilene trace, beta=2, t = 10, blind ranking N=244K N=487K N=1.2M N=2.4M 10 -1 10 0 10 1 10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6 Packet sampling rate (%) Average number of misranked flow pairs Jussieu trace, beta = 2, t = 10, blind ranking N=52K N=103K N=258K N=515K Figure 4: Performance of blind ranking varying the total number of flows sary to remove the discrepancy between carried data volume (throughput) and application data volume (goodput). Using the previous values for N , and average flow size, we reproduce Figure 2, but this time for the protocol-aware case. This leads to Figure 5, which illustrates the impact of the number of largest flows to rank. For lack of space, we omit the other figures. We compare this new figure to its counterpart in the blind case. We make the following two observations: (i) The protocol-aware method improves the accuracy of the largest flows ranking by an order of magnitude for high sampling rates (above 1%). For example, for the Abilene trace, a sampling rate on the order of 50% was necessary to detect and rank the largest 5 flows with the blind method. Now, with the protocol-aware method, a sampling rate on the order of 5% is sufficient. The same conclusion applies to the Jussieu trace. A sampling rate on the order of 10% is needed. With the protocol-aware method, it becomes on the order of 1%. (ii) The protocol-aware method does not improve the performance when applied at low sampling rates (above 1%). This can be clearly seen if we compare the plots between both figures for sampling rates below 1%. This results confirms our observations in Section 3.3.2. 5.3 Largest flows detection To illustrate the difference between ranking and detection, we consider the same scenario as in Section 5.1.1. We plot the detection metric as a function of the sampling rate for different values of t (the number of top flows of interest) and for both Abilene and Jussieu traces. This gives Figure 6 for 195 10 -1 10 0 10 1 10 -2 10 0 10 2 10 4 10 6 Packet sampling rate (%) Average number of misranked flow pairs Abilene trace, N=487K, beta=2, protocol-aware ranking t=25 t=10 t=5 t=2 t=1 10 -1 10 0 10 1 10 -2 10 0 10 2 10 4 10 6 Packet sampling rate (%) Average number of misranked flow pairs Jussieu trace, N=103K, beta=1.5, protocol-aware ranking t=25 t=10 t=5 t=2 t=1 Figure 5: Performance of protocol-aware ranking varying the number t of top flows of interest blind ranking and Figure 7 for protocol-aware ranking. A comparison between these results and their counterparts in Figure 2 and 5, respectively, shows a significant improvement in the detection case for both ranking methods. All plots are shifted down by an order of magnitude. For example , in the case of blind ranking, the required sampling rate to correctly rank the top 5 flows was around 50% for the Abilene trace and 10% for the Jussieu trace. Now, with blind detection, it is around 10% and 3%, respectively. Another example is with the protocol-aware method where a sampling rate around 10% was required to rank the largest 10 flows (Figure 5), whereas now, a sampling rate around 1% is sufficient to only detect them. The same gain can be observed if we reconsider the other scenarios in Section 5.1 (not presented here for lack of space). Also, note how in the detection case the protocol aware method allows a better accuracy for high sampling rates when compared to the blind method. For low sampling rates (e.g., below 1%), the accuracy does not improve. EXPERIMENTAL RESULTS In this section we present the results of running random sampling experiments directly on the packet traces. We use the traces described in Section 5 and compute the performance metrics defined in Section 4.1. In our traces we consider only TCP packets. Since TCP sequence numbers count bytes, we express the flow sizes in bytes instead of packets throughout this section. Our experiments are meant to address four major issues that arise when we move from the analytical study to a real network setting: (i) how to deal with invalid TCP sequence 10 -1 10 0 10 1 10 -2 10 0 10 2 10 4 10 6 Packet sampling rate (%) Average number of misranked flow pairs Abilene trace, N=487K, beta=2, blind detection t=25 t=10 t=5 t=2 t=1 10 -1 10 0 10 1 10 -2 10 0 10 2 10 4 10 6 Packet sampling rate (%) Average number of misranked flow pairs Jussieu trace, N=103K, beta=1.5, blind detection t=25 t=10 t=5 t=2 t=1 Figure 6: Only detecting the largest flows: Performance of blind ranking varying the number t of top flows of interest numbers in the packet stream; (ii) the importance of flow size distributions and duration of the measurement interval; (iii) the impact of packet loss rates on individual flows lost packets trigger retransmissions by the TCP senders; (iv) the variability of the detection/ranking performance across multiple bins and packet sampling patterns. 6.1 Implementation of protocol-aware ranking The protocol-aware method depends on TCP sequence numbers to perform the ranking. For a given flow, it keeps track of the lowest and highest sequence number observed (taking care of packets that wrap around the sequence number space), s b and s e respectively. Note that an actual implementation of this method would just require two 32 bit fields per flow to store the two sequence numbers. At the end of the measurement period, we compute the difference between the highest and lowest sequence numbers for each sampled flow, and we use the obtained values to rank flows. We then compare this ranking with the one obtained by counting all the bytes each flow transmits in the original non sampled traffic. In order to discard invalid packets carrying incorrect sequence numbers that would corrupt the ranking, we implement a simple heuristic to update s e and s b . A sampled packet with sequence number S &gt; s e causes an update s e S if S - s e &lt; ( MTU)/p. The same rule applies to the updates of s b . This way we set a limit on the maximum distance in the sequence space between two sampled pack-196 10 -1 10 0 10 1 10 -2 10 0 10 2 10 4 10 6 Packet sampling rate (%) Average number of misranked flow pairs Abilene trace, N=487K, beta=2, protocol-aware detection t=25 t=10 t=5 t=2 t=1 10 -1 10 0 10 1 10 -2 10 0 10 2 10 4 10 6 Packet sampling rate (%) Average number of misranked flow pairs Jussieu trace, N=103K, beta=1.5, protocol-aware detection t=25 t=10 t=5 t=2 t=1 Figure 7: Only detecting the largest flows: Performance of protocol-aware ranking varying the number t of top flows of interest ets. This distance is inversely proportional to the sampling rate and depends on the Maximum Transmission Unit. Furthermore, we use the parameter that allows to make this threshold more or less "permissive" in order to account for the randomness of the sampling process and for other transport-layer events (e.g., packet retransmissions when the TCP window is large). We have run several experiments with different values of and the results have shown little sensitivity to values of &gt; 10. All the results in this Section are derived with = 100. 6.2 Flow size distribution and measurement interval As shown in Figure 1, flow size distributions do not follow a perfect Pareto. Furthermore, the measurement interval itself plays a major role in shaping the distribution: it caps the size of the largest flows, that is not unbounded but now depends on the link speed. Indeed, network operators often run measurements using a "binning" method, where packets are sampled for a time interval, classified into flows, ranked, and then reported. At the end of the interval, the memory is cleared and the operation is repeated for the next measurement interval. With this binning method, all flows active at the end of the measurement interval are truncated, so that not all sampled packets of the truncated flow are considered at the same time for the ranking. The truncation may, therefore, penalize large flows and alter the tail of the flow size distribution (where flows are of large size and probably last longer than the measurement interval). Each experiment consists of the following. We run ran-10 -1 10 0 10 1 10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6 packet sampling rate % average number of misranked flow pairs Jussieu trace, blind ranking top 1 top 2 top 5 top 10 top 25 10 -1 10 0 10 1 10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6 packet sampling rate % average number of misranked flow pairs Jussieu trace, protocol-aware ranking top 1 top 2 top 5 top 10 top 25 Figure 8: Performance of blind and protocol-aware ranking on Jussieu trace (60s measurement interval ). dom sampling on the packet traces and classify the sampled packets into flows. At the end of each measurement interval (set to 1 or 5 minutes), we collect the flows and rank them by the number of bytes sampled for each flow. We compare the ranking before and after sampling using our performance metric (Section 4.1). For each sampling rate we conduct 15 runs and we calculate averages. The results of the experiments confirm the numerical results of the previous section. In the interest of space, we plot the results of two representative experiments on which we make several observations. The difference between numerical and experimental results, especially at low sampling rates, is caused by the non perfect match of the empirical flow size distribution with Pareto (Figure 1). Figure 8 shows the performance of ranking flows on the Jussieu trace when the measurement bin is 60s. We consider a wide range of sampling rates from 0.1% to 50% and study the performance when ranking the top 1, 2, 5, 10 and 25 flows in the packet stream. The top graph in Figure 8 is derived using the blind method while the bottom graph shows the performance of the protocol-aware methods. These results are very similar to the numerical results. For sampling rates above 1%, protocol-aware ranking gives approximately an order of magnitude gain on the performance when compared to blind ranking. When the sampling rate is lower than 1%, however, the performance of the two methods is similar. Overall, the blind method requires a sampling rate of 10% to correctly identify the largest flow in the packet stream. The same sampling rate allows to correctly rank the largest 5 flows when using the protocol-aware method. 197 6.3 Impact of loss rate In the analysis of the protocol-aware method in Section 3.2, we made the assumption of negligible number of retransmissions for all the flows in the packet stream. A retransmitted packet may cause inconsistency between the blind and protocol-aware method depending on the location of the monitoring point. Indeed, the blind method counts the total number of bytes sent by the flow while the protocol-aware method considers only the data sent by the transport layer. Therefore, if the packet is lost before the monitoring point, the blind and protocol-aware method will have a consistent view of the number of bytes sent. Instead, if the packet is lost after the monitoring point, the blind method may count this packet twice. The impact of packet losses on the detection and ranking of the largest flows depends on the metric used to estimate the size of the flows. If flow sizes are estimated according to the total number of bytes sent (i.e., the throughput), then the protocol-aware method may incur in an underestimation error that is independent of the sampling rate (it will occur even if all packets are sampled!). On the other hand, if the flow sizes are estimated according to the transport data sent (i.e., the goodput), then the blind method may incur in an overestimation error independently of the sampling rate. To illustrate the effect of packet loss rates, we plot in Figure 9 the performance of detecting the largest flows in the Abilene trace when the measurement bin is 5 minutes and the flow sizes are measured using the total number of bytes sent over the link. The top graph shows the performance of the blind method, while the bottom graph presents the results for the protocol-aware method. We can make the following observations: The protocol-aware method keeps performing better than the blind method when the sampling rate is above 1%. At lower sampling rates, the blind method performs better although it presents very large errors. For sampling rates above 2%, the curve relative to the detection of the top-25 flows in the protocol-aware method flattens to a value around 70. This is due to the presence of a few flows that experience a high loss rate when compared to other flows. Increasing the sampling rate does not help the protocol-aware method in detecting the largest flows when the volume of bytes sent is used to define the flow size. However , the protocol-aware method can correctly detect the top-25 flows when their size is defined in terms of transport data (see Figure 10). In summary, the network operator has to choose the metric of interest that depends on the application. For example , for anomaly detection or traffic engineering, a metric that counts the number of bytes sent may be more appropriate . Instead, for dimensioning caches and proxies, the metric that considers the size of the objects transferred may be preferred. This latter metric suits more the protocol-aware method. 6.4 Variability of the results A last important aspect that we need to address is the variability of the results across multiple measurement intervals and different realizations of the sampling process. Indeed, moving from one measurement interval to another, 10 -1 10 0 10 1 10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6 packet sampling rate % average number of misranked flow pairs Abilene trace, blind detection top 1 top 2 top 5 top 10 top 25 10 -1 10 0 10 1 10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6 packet sampling rate % average number of misranked flow pairs Abilene trace, protocol-aware detection top 1 top 2 top 5 top 10 top 25 Figure 9: Performance of blind (top) and protocol-aware (bottom) detection on Abilene trace (300s measurement interval). the composition of flows varies and with it the flow size distribution . Moreover, the sampling process may "get lucky" in certain cases and provide good results. The opposite is also possible. Figure 11 shows the average performance over 15 sampling experiments of the detection of the top-10 flows in the Abilene trace over the 5-minute measurement intervals. The error bars indicate the standard deviation across the 15 experiments. As usual, the top graph refers to the blind method, while the bottom graph presents the protocol-aware method results. As we can see the average performance shows limited variability. A sampling rate of 0.1% gives poor results for all bins, while increasing the sampling rates consistently helps. With a sampling rate of 10% the performance metric (i.e., average number of misranked flow pairs) for the blind method is always below 100 while the protocol-aware method is always below 1. Looking at the standard deviation, we observe large values for the blind method and much smaller values for the protocol-aware method. This indicates that the blind method is more sensitive to the sampling process than the protocol-aware method. The explanation is given in Section 3.3.2 where we showed that that the blind method presents a larger error for large flow sizes (expect when the sampling rate is very low). CONCLUSIONS We study the problem of detection and ranking the largest flows from a traffic sampled at the packet level. The study is 198 10 -1 10 0 10 1 10 -1 10 0 10 1 10 2 10 3 10 4 10 5 10 6 packet sampling rate % average number of misranked flow pairs Abilene trace, protocol-aware detect top 1 top 2 top 5 top 10 top 25 Figure 10: Performance of protocol-aware detection on Abilene trace (300s measurement interval) when using actual amount of data sent by the transport layer application. done with stochastic tools and real packet-level traces. We find that the ranking accuracy is strongly dependent on the sampling rate, the flow size distribution, the total number of flows and the number of largest flows to be detected and ranked. By changing all these parameters, we conclude that ranking the largest flows requires a high sampling rate (10% and even more). One can reduce the required sampling rate by only detecting the largest flows without considering their relative order. We also introduce a new method for flow ranking that exploits the information carried in transport header. By analysis and experimentation, we demonstrate that this new technique allows to reduce the required sampling rate by an order of magnitude. We are currently exploring two possible future directions for this work. First, we want to study the accuracy of the ranking when the sampled traffic is fed into one of the mechanisms proposed in [10, 12] for sorting flows with reduced memory requirements. Second, we are exploring the use of adaptive schemes that set the sampling rate based on the characteristics of the observed traffic. Acknowledgements We wish to thank NLANR [15], Abilene/Internet2 [1] and the Metropolis project [13] for making available the packet traces used in this work. REFERENCES [1] Abilene: Advanced networking for leading-edge research and education. http://abilene.internet2.edu. [2] C. Barakat, P. Thiran, G. Iannaccone, C. Diot, and P. Owezarski. Modeling Internet backbone traffic at the flow level. IEEE Transactions on Signal Processing (Special Issue on Signal Processing in Networking), 51(8):21112124, Aug. 2003. [3] M. Charikar, K. Chen, and M. Farach-Colton. Finding frequent items in data streams. In Proceedings of ICALP, 2002. [4] B. Y. Choi, J. Park, and Z. Zhang. Adaptive packet sampling for flow volume measurement. Technical Report TR-02-040, University of Minnesota, 2002. [5] G. Cormode and S. Muthukrishnan. What's hot and what's not: Tracking most frequent items dynamically. In Proceedings of ACM PODS, June 2003. [6] M. Crovella and A. Bestravos. Self-similarity in the World Wide Web traffic: Evidence and possible causes. IEEE/ACM Transactions on Networking, 5(6):835846, Dec. 1997. 0 200 400 600 800 1000 1200 1400 1600 1800 10 -2 10 0 10 2 10 4 10 6 10 8 time (sec) average number of swapped flow pairs Abilene trace, detection sampling 0.1% sampling 1% sampling 10% 0 200 400 600 800 1000 1200 1400 1600 1800 10 -2 10 0 10 2 10 4 10 6 10 8 time (sec) average number of swapped flow pairs Abilene trace, detection sampling 0.1% sampling 1% sampling 10% Figure 11: Performance of blind (top) and protocol-aware (bottom) detection over multiple 300s intervals (Abilene trace). Vertical bars show the standard deviation over multiple experiments. [7] E. Demaine, A. Lopez-Ortiz, and I. Munro. Frequency estimation of internet packet streams with limited space. In Proceedings of 10th Annual European Symposium on Algorithms, 2002. [8] N. G. Duffield, C. Lund, and M. Thorup. Properties and prediction of flow statistics from sampled packet streams. In Proceedings of ACM Sigcomm Internet Measurement Workshop, Nov. 2002. [9] N. G. Duffield, C. Lund, and M. Thorup. Estimating flow distributions from sampled flow statistics. In Proceedings of ACM Sigcomm, Aug. 2003. [10] C. Estan and G. Varghese. New directions in traffic measurement and accounting. In Proceedings of ACM Sigcomm, Aug. 2002. [11] N. Hohn and D. Veitch. Inverting sampled traffic. In Proceedings of ACM Sigcomm Internet Measurement Conference, Oct. 2003. [12] J. Jedwab, P. Phaal, and B. Pinna. Traffic estimation for the largest sources on a network, using packet sampling with limited storage. Technical Report HPL-92-35, HP Laboratories, Mar. 1992. [13] Metropolis: METROlogie Pour l'Internet et ses services. http://www.laas.fr/ owe/METROPOLIS/metropolis eng.html. [14] T. Mori, M. Uchida, R. Kawahara, J. Pan, and S. Goto. Identifying elephant flows through periodically sampled packets. In Proceedings of ACM Sigcomm Internet Measurement Conference, Oct. 2004. [15] NLANR: National Laboratory for Applied Network Research. http://www.nlanr.net. [16] Packet Sampling Working Group. Internet Engineering Task Force. http://www.ietf.org/html.charters/psamp-charter.html. [17] K. Papagiannaki, N. Taft, and C. Diot. Impact of flow dynamics on traffic engineering design principles. In Proceedings of IEEE Infocom, Hong Kong, China, Mar. 2004. [18] Renater. http://www.renater.fr. [19] H. Schulzrinne, S. Casner, R. Frederick, and V. Jacobson. RTP: A transport protocol for real-time applications. RFC 1889, Jan. 1996. [20] A. Shaikh, J. Rexford, and K. G. Shin. Load-sensitive routing of long-lived IP flows. In Proceedings of ACM Sigcomm, Sept. 1999. [21] M. Spiegel. Theory and Problems of Probability and Statistics. McGraw-Hill, 1992. 199

largest flow detection and ranking;validation with real traces;Packet sampling;performance evaluation

164

Ranking Target Objects of Navigational Queries

Web navigation plays an important role in exploring public interconnected data sources such as life science data. A navigational query in the life science graph produces a result graph which is a layered directed acyclic graph (DAG). Traversing the result paths in this graph reaches a target object set (TOS). The challenge for ranking the target objects is to provide recommendations that reflect the relative importance of the retrieved object, as well as its relevance to the specific query posed by the scientist. We present a metric layered graph PageRank (lgPR) to rank target objects based on the link structure of the result graph. LgPR is a modification of PageRank; it avoids random jumps to respect the path structure of the result graph. We also outline a metric layered graph ObjectRank (lgOR) which extends the metric ObjectRank to layered graphs. We then present an initial evaluation of lgPR. We perform experiments on a real-world graph of life sciences objects from NCBI and report on the ranking distribution produced by lgPR. We compare lgPR with PageRank. In order to understand the characteristics of lgPR, an expert compared the Top K target objects (publications in the PubMed source) produced by lgPR and a word-based ranking method that uses text features extracted from an external source (such as Entrez Gene) to rank publications.

INTRODUCTION The last few years have seen an explosion in the number of public Web accessible data sources, Web services and semantic Web applications. While this has occurred in many domains, biologists have taken the lead in making life science data public, and biologists spend a considerable amount of time navigating through the contents of these sources, to obtain information that is critical to their research. Providing meaningful answers to queries on life science data sources poses some unique challenges. First, information about a scientific entity, e.g., genes, proteins, sequences and publications, may be available in a large number of autonomous sources and several sources may provide different descriptions of some entity such as a protein. Second, the links between scientific objects (links between data entries in the different sources) are important in this domain since they capture significant knowledge about the relationship and interactions between these objects. Third, interconnected data entries can be modeled as a large complex graph. Queries could be expressed as regular expression navigational queries and can more richly express a user's needs, compared to simpler keyword based queries. Consider the following navigational query: Retrieve publications related to the gene 'tnf ' that are reached by traversing one intermediate (protein or sequence) entry. This query expresses the scientist's need to expand a search for gene related publications beyond those publications whose text directly addresses the 'tnf' gene, while still limiting the search to publications that are closely linked to gene entries. Consider gene sources OMIM Gene and Entrez Gene, protein sources NCBI Protein and SwissProt, sequences in NCBI Nucleotide and biomedical publications in PubMed. Figure 1 represents the results of evaluating this navigational query against these sources. The result is a layered DAG; we refer to it as a result graph (RG). All paths in this directed result graph (RG) start with data entries in the sources OMIM Gene or Entrez Gene; this is the first layer. They visit one intermediate data entry in sources NCBI Protein, Swiss Prot or NCBI Nucleotide (second layer) and they terminate in a publication data entry in PubMed (final layer). The query returns all objects in PubMed that are reached 27 &quot;tnf&quot; keyword &quot;tnf&quot; keyword NCBI Nucleotide Swiss Prot PubMed NCBI Protein OMIM Gene Entrez Gene Figure 1: An example of a result graph (RG) by traversing results paths; these PubMed entries are re-ferred to as the target object set (TOS) reached by traversing the result paths of the RG. In contrast, a keyword based query would not have been able to specify the set of target publications. Navigational queries, the RG and the target object set (TOS) that answers the query are defined in the paper. It is difficult for a user to explore all target objects in a reasonable amount of time and it is important to provide a ranking of the TOS. As is well known, word based ranking methods are very good at identifying the most relevant results, typically using features extracted from the contents of the target objects. For example [13] produces a ranking of documents in PubMed that are most relevant to a gene. In contrast, PageRank [11] focuses on the importance of the target object and importance is transferred from other important objects via the link structure. A recent technique ObjectRank [1] addresses both relevance and importance; it exploits schema knowledge to determine the correct authority transfer between important pages. We note that there is also research on ranking paths [2]. For term-based query dependent ranking, we refer to [3, 12]. The focus of this paper is to produce a ranking method to select the best target objects in the RG that answer the navigational query. Our ideal ranking must identify target objects that are both relevant and important. The ranking must also be query dependent since we must guarantee that the target objects that are ranked indeed occur in the RG and answer the navigational query. Further, both relevance and importance must be determined with respect to the objects in TOS, rather than with respect to all the data entries (as is the case with PageRank). We propose two ranking metrics for the layered graph RG; they are layered graph PageRank (lgPR) and layered graph ObjectRank (lgOR). lgPR extends PageRank by distinguishing different roles (intermediate node, answer node) which can be played by the same node in the result graph. It does not perform random jumps so as to respect the RG. Our second metric lgOR is an extension to ObjectRank; due to space limitations we only discuss it briefly. We report on our preliminary evaluation of lgPR on a real dataset from NCBI/NIH. For some navigational queries, we apply lgPR to the corresponding RG and use the ranking distribution for lgPR to illustrate that lgPR indeed discriminates among the TOS objects. We also apply the original PR metric to the object graph of life science data (against which we evaluate the query). We compare with applying lgPR to the actual RG to illustrate that lgPR and PR produce dissimilar rankings. Finally, we report on an initial user experiment. We consider a set of complex queries typical of a scientist searching for gene related PubMed publications, and the Top K results of a word based ranking technique (Iowa) that has been shown to be accurate in answering gene queries [13]. We compare the Iowa Top K publications with the lgPR Top K publications, for some sample gene related queries, using criteria that reflect both relevance and importance. We use these criteria to understand the characteristics of lgPR. The paper is organized as follows: Section 2 describes the data model, navigational query language and layered DAG result graph. Section 3 presents PageRank, lgPR, ObjectRank , and lgOR. 4 reports on preliminary results of an experimental study with NCBI data and concludes. DATA MODEL We briefly describe a data model and navigational query language for the life science graph. Details in [6, 9, 14]. 2.1 Data Model for the Life Science Graph The data model comprises three levels: ontology, source and data (Figure 2). At the ontology level, a domain ontol-Gene Marker Publication Nucleotide Protein Disease OMIM Gene \NCBI Gene PubMed OMIM Disease Swiss Prot NCBI Protein NCBI Nucleotide UniSTS Ontology Level Source Level Mappings Data Level Mappings Figure 2: A Data Model for the Life Science Graph ogy describes the universe of discourse, e.g., a gene, a pro-28 tein, etc., and the relationships among them. An ontology graph OG = (C, L C ) models the domain ontology, where nodes in C represent classes, and edges in L C correspond to relationships among classes. For example, genes and publications are classes in OG and the association discuss relates publications with genes. In this paper, we only consider one type of link, isRelatedTo, to capture the semantics of a relationship; therefore, we omit all link labels. At the source level, a source graph SG = (S, L S ) describes data sources and links that implement logical classes (C) and associations (L C ) in OG, respectively. For example, PubMed and Entrez Gene are sources that implement the logical classes publications and genes, respectively. A mapping defines logical classes in C in terms of the sources in S that implement them. A link between sources represents a hyper-link, a service or an application that connects these two sources. At the data level, a Data Graph is a graph (D, L D ), where D is a set of data entries and L D is a set of references between entries. A mapping m S establishes which data entries in D are published by source S. 2.2 Navigational Query Language We define a query as a path expression over the alphabet C in OG, where each class occurrence can optionally be an-notated with a Boolean expression. The simplest Boolean expression is the comparison of a Field to a particular value. In this paper, a field can be either source or Object content, and the relational operators can be "=" for source and "contains" for Object content. A condition over source and the relational operator "=", (source = "name-of-source"), restricts the query to some specific sources that implement the class. A condition on Object content and the relational operator "contains", specifies the set of keywords that must occur within objects in the Data Graph. The symbol is a wild-card matching any class and the "." represents any relationship. The query: Retrieve publications that are related to the gene "tnf or aliases in human" in OMIM or Enrtez Gene, and are reached by traversing one intermediate resource, is expressed in the navigational query language as follows: Q = Gene[Object content contains {"tnf" and aliases in human} and source = OMIM or Entrez Gene] Publication The answer to a query Q is defined at the three levels of the data model. It comprises three sets of paths: OG (Q), SG (Q) and DG (Q). The meaning of query Q with respect to the ontology graph OG, OG (Q), is the set of simple paths in OG that correspond to words in the language induced by the regular expression Q. The meaning of the query with respect to the source graph SG, SG (Q), is the set of all simple paths in SG that correspond to mappings of the paths in OG (Q). Finally, the answer for query Q with respect to the data graph DG, DG (Q), is the set of simple paths in DG that are the result of mapping the paths in SG (Q) using mapping function m S . A simple path does not repeat (revisit) the same class, data source or data entry (in the same path). The queries that are presented in this section are typical queries posed by researchers. At present, there are no query evaluation engines to answer navigational queries and researchers must rely on manual navigation via browsers or they must write scripts; the latter involves labor to keep writing the scripts and the scripts may be inefficient in answering these queries. 2.3 Result Graph The union of paths in DG (Q) is the result graph RG. We note that for our query language, all the paths that satisfy a query are of the same length, i.e., all the paths in the sets OG (Q), SG (Q) and DG (Q) are of the same length. We model a result graph RG Q = (D RG , L RG ) for a query Q, as a layered directed acyclic graph comprising k layers, L 1 , ..., L k , where k is determined by the query. The set of nodes D RG corresponds to the union of the data entries that appear in the paths in DG (Q). L RG represents the links among these data entries. A layer L i is composed of the union of the data entries in the paths DG (Q) that appear in the i-th position of the paths. The data entries in the k-th layer are called the target objects and they form the target object set (TOS) of the RG. Note that since the result graph has multiple paths, and since a source may occur in different layers of these paths, the same data entry may appear multiple times in the different layers, depending on its connectivity to other data entries. In this case, each occurrence of the data entry is represented independently within each layer/path in which it occurs. The result graph framework distinguishes the different roles (intermediate node, answer node) which can be played by the same node in the result graph. Figure 1 is a layered RG for the following query: Retrieve publications related to the gene "tnf " traversing one intermediate source; it has three layers. The first layer corresponds to the genes in the sources OMIM Gene and Entrez Gene that are related to the keyword "tnf". The second layer are the entries in the sources NCBI Protein, Swiss Prot or NCBI Nucleotide that are reached by objects in the first layer. Finally , the target objects in the third layer (TOS) are the publications in PubMed that are linked to the objects in the second layer. RANKING METRICS We briefly describe the PageRank metric [11] and then discuss our metric lgPR for layered DAGs. We briefly discuss the ObjectRank metric [1] and our extension lgOR. 3.1 PageRank PageRank assumes that links between pages confer authority . A link from page i to page j is evidence that i is suggesting that j is important. The importance from page i that is contributed to page j is inversely proportional to the outdegree of i. Let N i be the outdegree of page i. The corresponding random walk on the directed web graph can be expressed by a transition matrix A as follows: A[i, j] = 1 N i if there is an edge from i to j 0 otherwise Let E be an arbitrary vector over the webpages, representing the initial probability of visiting a page. Let d be the probability of following a link from a page and let (1 -d) be the probability of a random jump to a page. The PageRank ranking vector R = dA R + (1 - d)E. R converges for the web graph with any E, since generally the web graph is aperiodic and irreducible[5, 10]. PageRank cannot be directly applied to a layered graph. A Markov Chain is irreducible if and only if the graph contains only one strongly connected component. RG is not 29 outgoing links with respect to the query. There are several potential ways to extend PageRank for RG. First, one can ignore links that point to pages without outgoing edges since these pages do not affect the ranking of other pages [11]. However we are specifically interested in obtaining a ranking for the TOS or the objects in the last layer of the layered result graph RG with no outgoing links, we cannot ignore these pages. Another possibility is modifying the transition matrix probability so that one takes a random jump from a node in the TOS [5]. This will ensure that the graph will be irreducible and aperiodic. However, this would arbitrarily modify RG whose structure is determined by the query; modifying RG will not assure that it answers the query. To summarize, the extensions to PageRank in the literature cannot be applied to the problem of ranking the target object set TOS of RG. 3.2 Layered Graph PageRank(lgPR) We describe layered graph PageRank to rank the TOS. 3.2.1 The Metric Table 1 lists the symbols used to compute lgPR. Symbol Meaning RG(V RG , E RG ) Result Graph, a layered DAG, with objects V RG and edges E RG e E RG an edge in E RG R ranking vector for objects in RG R ini initial ranking vector A lg the transition matrix for objects in RG k the number of layers in the result graph OutDeg RG (u p ) outdegree from object u at layer p (across multiple link types to objects in layer p + 1 Table 1: Symbols used by lgPR The layered DAG result graph RG is represented by a transition matrix A lg to be defined next. Note that an object in the object graph may occur in multiple paths of the result graph, in different layers; it will be replicated in the transition matrix for each occurrence. Each object u at layer p will have an entry in the transition matrix to some object v at layer q. We denote the occurrence of them as u p and v q respectively. The ranking vector R is defined by a transition matrix A lg and initial ranking vector R ini , is as follows: R = A k -1 lg R ini = ( k -1 l=1 A lg ) R ini We pick R ini as follows: the entry for an object in R ini is 1 if this object is a link in start layer and 0 otherwise. The transition matrix A lg is computed as follows: A lg [i p u , j q v ] = 1 OutDeg RG (u p ) if OutDeg RG (u p ) &gt; 0 and e(u p , v q ) E RG , 0 otherwise. Note that we define the outdegree of each object in RG to only consider those edges that actually occur in RG and link to objects in the next layer. This reflect the probability that a user follows an object path in the RG. In contrast, PageRank considers all outgoing edges from a page. Unlike PageRank, lgPR differentiates the occurrence of a data entry in different layers, as well as the links to entries in subsequent layers; lgPR is thus able to reflect the role of objects and links (from the entire graph of data entries) in answering a navigational query. Suppose an object a occurs in an intermediate layer as well as in the TOS of the RG. It is possible that a is able to convey authority to other objects in the TOS. However, a may not rank very high in the TOS for this query. This characteristic is unique to lgPR. Thus, the score associated with the object is query dependent to reflect the role played by the object in the result graph. 3.2.2 Convergence Property This transition matrix A lg is neither irreducible nor aperiodic as all rows for target objects contain only 0's. The matrix A is a nilpotent matrix and the number of layers is the index. We provide two defintions (details in [8]). Definition 3.1. A square matrix A is a nilpotent matrix, if there exists some positive integer k such that A k = 0 but A k -1 = 0. Integer k is known as the index of A. Definition 3.2. Let k be the index of A. {A k -1 x, A k -2 x, ..., Ax, x } form a Jordan Chain, where x is any vector such that A k -1 x = 0. A characteristic of a nilpotent matrix is that its only eigenvalue is 0. The consequence is that any vector x is an eigenvector of A as long as Ax = 0. From the previous definition {A k -1 lg R ini , A k -2 lg R ini , ..., A lg R ini , A lg } forms a Jordan Chain, since A k -1 lg R ini = 0. We show following two lemmas without providing proof in this paper. Lemma 3.3. Jordan chain {A k -1 lg R ini , A k -2 lg R ini , ..., A lg R ini , A lg } is a linearly independent set. Lemma 3.4. {A k -1 lg R ini , A k -2 lg R ini , ..., A lg R ini , A lg } consists of a sequence of ranking vectors. In R ini , only objects in layer 0 have non-zero scores; In ranking vector A m lg R ini , only objects in layer m receive non-zero scores. The final ranking vector by lgPR is the first eigenvector in the Jordan Chain, given the above initial ranking vector R ini and the transition matrix A lg . While the traditional PageRank algorithm converges on a ranking in multiple iterations , lgPR can be computed in exactly k - 1 iterations. Note that because RG is a layered DAG, we can use link matrices, each of which represents links between neighboring layers, instead of the single transition matrix A lg for the entire graph. We also use keywords to filter query answers at each iteration. 3.3 Layered Graph ObjectRank(lgOR) PR is computed a priori on the complete data graph and is independent of the RG. A recent technique ObjectRank [1] extends PageRank to consider relevance of query keywords. It exploits schema knowledge to determine the correct authority transfer in a schema graph. In ObjectRank, the authority flows between objects according to semantic connections . It does so by determining an authority weight for each edge in their schema graph. The ranking is (keyword) query dependent. Due to space limitations, we do not provide the details of the ObjectRank metric. Instead, we briefly describe how irreducible since the last layer in RG contains nodes with no 30 the transition matrix for lgPR can be extended to consider the authority weights associated with edges that occur in RG. Consider a metric layered graph ObjectRank(lgOR). The difference from lgPR is the transition matrix A OG . It is as follows: A[i p u , j q v ] = (e E RG ) if e(u p , v q ) E RG , 0 otherwise. (e E RG ) = (e ESG ) OutDeg(u p ,e ESG ) if OutDeg(u p , e E SG ) &gt; 0 0 if OutDeg(u p , e E SG ) = 0 Let the edge between u p and v q map to an edge E SG in the SG. (E SG ) represents the authority transfer weight associated with E SG . OutDeg(u p , e E SG ) is the outdegree in RG of type E SG . As discussed in [1], the success of ObjectRank depends on correctly determining the authority weight to be associated with each link. Figure 3 (next section) illustrates the source graph that we use in our evaluation of navigational queries. For lgOR to be successful, an authority weight may have to be associated with each link in each result path (type) in the RG. Experiments with users to determine the correct authority weights for lgOR is planned for future work. Currently the importance is computed after query evaluation . We compute result graph first, then ranking, for the reason that the transition matrix is defined in terms of outdegree in the RG. This motivates further research of combination of two problems, whose ideal solution is to ranking objects during query evaluation. EXPERIMENTS ON LGPR We report on experiments on real world data. We show that the lgPR ranking distribution has the ability to differentiate among the target objects of the RG and it is different from PageRank. A user compared the Top K results of lgPR and a word based ranking (Iowa) [13], using criteria that reflect both importance and relevance, to determine their characteristics. 4.1 Experiment Setting NCBI/NIH is the gatekeeper for biological data produced using federal funds in the US 1 . We consider a source graph SG of 10 data sources and 46 links. Figure 3 presents the source graph used in this task.We used several hundred keywords to sample data from these sources (the EFetch utility) and followed links to the other sources (the ELink utility). We created a data graph of approximately 28.5 million objects and 19.4 million links. We note that several objects are machine predicted objects so it is not uncommon that they have no links. The object identifiers for the data entries (nodes of the data graph) and the pair of object identifiers (links) were stored in a DB2 relational database. Table 2 identifies the queries and keywords that were used in this experiment. The symbols g, p, n, s refer to classes gene, publication, nucleotide and SNP, respectively. Note that is the wild card and can match all the classes and sources (in the source graph). For each navigational query, the source paths that answer the query were determined using an algorithm described in 1 www.ncbi.nlm.nih.gov Class PubMed PmId Title Class Author Name Class Journal Name ClassYear Year (1,*) (1,1) (1,*) Class Lash Terms TermId Descrip (1,*) (0,*) Class Entrez Gene EGId (1,*) (1,1) Class Geo GeoId ClassOMIM CddId Class UniSTS USId Class UniGene UGId Class Entrez Protein EPId Class dbSNP dbSID Class CDD CddId Class Entrez Nucleotide NuId (1,*) Figure 3: Source Graph for User Evaluation Queries g.n.p, g.s.p, g.n.s.p, g.s.n.p, g.s.g.n.p, g.s.n.g.p, g. .p, g. . .p, g. . . .p "parkinson disease", "aging","cancer" Keywords "diabetes", "flutamide", "stress" "degenerative joint","tnf","insulin" "fluorouracil", "osteoarthritis","sarcoma" Table 2: Experiment setting [14]. Evaluating the paths in the data graph for each source path was implemented by SQL queries. Since a result graph RG could involve multiple source paths whose computation may overlap we applied several multiple query optimization techniques. The SQL queries were executed on DB2 Enterprise Server V8.2 installed on a 3.2 GHz Intel Xeon processor with 1GB RAM. The execution time for these queries varied considerably, depending on the size and shape of RG. If we consider the query g.n.p with keyword "degenerative joint" used to filter 'g', one source path was ranked in approximately 1 second. However, the query (g. . . .p) with the keyword aging used to filter 'g' created a very large result graph and the execution time for this was approximately 2000 seconds. Typically the We note that computing the high scoring TOS objects of the RG efficiently is a related but distinct optimization problem. 4.2 lgPR Distribution We report on the query (g. .p), i.e., paths from genes to publications via one intermediate source. Figures 4 and 5 report on the distribution of scores produced by the lgPR metric for the target objects in TOS for some representative queries. The first 10 bars represent scores in the range (0.00-0.01) to (0.09-0.1) and the last bar represents the range (0.1-1.0). Fig 4 shows that a small number of objects have very high score and the majority have a low score. As expected, many queries and keywords produced distributions that were similar to Figure 4. Most of the objects in TOS, in this case approx. 12,000 objects, had a very low score, and less than 200 object had a score in the range (0.1-1.0). However, we made an interesting observation that some queries produced distributions that were similar to Figure 5. In this case, while many of the results (approx. 120) had low scores in the range (0.00-0.02), 46 objects had scores in the range (0.1-1.0) and 120 objects had scores in between. 31 .00-.01 .01-.02 .02-.03 .03-.04 .04-.05 .05-.06 .06-.07 .07-.08 .08-.09 .09-.10 .10-1.00 0 2000 4000 6000 8000 10000 12000 12183 677 404 172 105 62 52 43 34 19 197 lgPR score Number of objects Figure 4: Histogram for query: g[Object content contains "aging"] p .00-.01 .01-.02 .02-.03 .03-.04 .04-.05 .05-.06 .06-.07 .07-.08 .08-.09 .09-.10 .10-1.00 0 10 20 30 40 50 60 70 80 81 41 12 4 32 11 8 0 2 1 46 lgPR score Number of objects Figure 5: Histogram for query: g [Object content contains "degenerative joint"] p Finally, we compared the ranking produced by lgPR and PageRank. We apply PageRank to the entire data graph of 28.5 million objects and 19.4 million links described in section 4.1. For the three sample queries (described in the next section), there are no PubMed ID's in common to the Top 25, 50, 100 for each of the queries, except that the top 50 of query with Lash term "allele" have 1 PubMed publications in common, and top 100 of same query have 3 in common. We speculate that the link structure of the RG is distinct compared to the link structure of the data graph; hence applying lgPR to the RG results in dissimilar ranking compared to a priori applying PageRank to the entire data graph. We summarize that the lgPR score can both identify those objects with a very low ranking that may not be of interest to the user. However, it can also be used to discriminate amongst objects in the TOS whose ranking has a much lower variation of scores. Finally, lgPR ranking is not the same as that produced by PageRank applied to the entire data graph. 4.3 User Evaluation In our user evaluation of lgPR, we consider a set of complex queries typical of a scientist searching for gene related PubMed publications, and the Top K results of a word based ranking technique (Iowa) that has been shown to be accurate in answering gene queries [13]. We compare the Iowa Top K publications with the lgPR Top K publications, for some sample gene related queries. We use criteria that reflect both relevance and importance to identify characteristics of lgPR. Researchers are particularly interested in genetic and phe-notypic variations associated with genes; these phenomena are often studied in the context of diseases, in a chromoso-mal region identified by a genomic marker (a unique known sequence) associated with the disease. Genetic and pheno-typic knowledge are described using terms of the Lash controlled vocabulary [7]. We focus on a branch of the Lash vocabulary that relates to phenotypes and population genetics . Terms of interest include linkage disequilibrium, quantitative trait locus and allele. Figure 6 presents a portion of the Lash controlled vocabulary (term hierarchy ). LD is not listed as the synonym to the term linkage disequilibrium, because LD may often refer to another concept . In the following experiment, we did not consider the plural form of some terms, such as alleles to allele, but this can be extended in the future studies. 1. EPIGENETIC ALTERATION 2. GENOMIC SEGMENT LOSS 3. GENOMIC SEGMENT GAIN 4. GENOMIC SEQUENCE ALTERATION 5. PHENOTYPIC ASSOCIATION (synonym: phenotype, trait) (a) locus association (synonym: locus, loci) i. linkage ii. quantitative trait locus (synonym: QTL) (b) allelic association (synonym: allele) i. linkage disequilibrium Figure 6: Branch 5 in Hierarchical controlled vocabulary of genetics terms (Lash Controlled Vocabulary ) The navigational query used in our evaluation experiment can be described in English as follows: "Return all publications in PubMed that are linked to an Entrez Gene entry that is related to the human gene TNF (or its aliases). The entry in PubMed must contain an STS marker and a term from the Lash controlled vocabulary." We used the query term "TNF AND 9606[TAXID]" 2 to sample data from Entrez Gene. We then followed 8 paths to PubMed. Table 3 reports on the number of entries in Entrez Gene as well as the cardinality of the TOS for some sample queries 3 . We briefly describe the word-based ranking method (Iowa) that focuses on ranking documents retrieved by PubMed 2 Note the Taxonomy ID for human is 9606 [4], and term 9606[TAXID] was used to select human genes. 3 We use g["tnf" and aliases in human] to denote g[Object content contains {"tnf" and aliases in human}]; the entries in the first column of Table 3 are similar. 32 Query Cardinality of TOS g["tnf" and aliases in human] 649 g["tnf" and aliases in human] p[STS 2777 marker and "allele"] g["tnf" and aliases in human] p[STS 257 marker and "linkage disequilibrium"] g["tnf" and aliases in human] p[STS 22 marker and "quantitative trait locus"] Table 3: Cardinality of TOS for human gene queries [13], so that relevant documents are ranked higher than non-relevant documents. This method relies on using post-retrieval queries (ranking queries), au-tomatically generated from an external source, viz., Entrez Gene (Locus Link), to rank retrieved documents. The research shows that ranking queries generated from a combination of the Official Gene Symbol, Official Gene Name, Alias Symbols, Entrez Summary, and Protein Products (optional ) were very effective in ranking relevant documents higher in the retrieved list. Documents and ranking queries are represented using the traditional vector-space representation , commonly used in information retrieval. Given a gene, the cosine similarity score between the ranking query vector for the gene and each document vector is computed. Cosine scores are in the [0, 1] range and documents assigned a higher score are ranked higher than documents with a lower score. In the absence of summary and protein product information, ranking queries generated from the gene symbol, name and aliases are used to rank retrieved documents . In this experiment study we are working on the Bio Web documents alone. We use the following criteria to compare the Top K results from Iowa and lgPR, to understand basic characteristics of the two methods. Criteria labeled R appear to judge the relevance of the paper and those labeled I appear to judge importance. Some criteria appear to judge both and are labeled R,I. 1. R: Does the title or abstract of the article contain the term TNF or its aliases in human? Does the article discuss immune response? 2. R,I: Does the article contain any disease related terms? Does the article contain any genomic components (genes, markers, snps, sequences, etc.)? 3. R,I: Does the article discuss biological processes related to the Lash terms? 4. R,I: What is the connectivity of the article to gene entries in Entrez Gene that are related to TNF? Note that as shown in Table 3, there are 649 Entrez Gene entries that are related to human gene TNF. Each PubMed publication was reached by following a result path through the result graph RG that started with one of these Entrez Gene entries. However, some PubMed publications may have been reached along multiple paths in the RG reflecting much greater connectivity . 5. I: What is the category of the article (review, survey, etc.). Does the article address some specific topics or is it a broad brush article? 6. I: Where did the article appear? What is the journal impact factor? Has the article been highly cited? Top 10 Rel. Imp. Criteria PMID (0-5) (0-5) 1. 2. 3. 4. 5. 6. 16271851 4 2 H M H L L L 1946393 4 4 H L H M H H 12217957 4 4 H H H L H H 12545017 4 4 H M H L H H 9757913 3 3 H L H L H H 8882412 4 4 H M H L H H 2674559 4 3 H M H L H L 7495783 4 3 H H H L H L 15976383 5 4 H H H H H L 10698305 3 3 H L H L H H Table 4: Relevance and Importance of Top 10 Puli-cations Reported by Iowa Ranking Method Top 10 Rel. Imp. Criteria PMID (0-5) (0-5) 1. 2. 3. 4. 5. 6. 7560085 5 5 H H H H H H 12938093 5 5 H H H H H H 10998471 3 3 M H H L H L 11290834 5 4 H H H H H L 11501950 4 3 H H H L H L 11587067 5 4 H H H H H L 11845411 2 4 L H H L H H 12133494 5 4 H H H H H L 12594308 4 4 H H H L H H 12619925 5 5 H H H H H H Table 5: Relevance and Importance of Top 10 Puli-cations Reported by lgPR Ranking Method Tables 4 and 5 report Top 10 publications in PubMed that are linked to an Entrez Gene entry that is related to human gene TNF and contain the term linkage disequilibrium. The first column reports the PubMed identifiers (PMIDs) of the Top 10 publications returned by the Iowa and the lgPR ranking methods. The human evaluation results are reported in the fourth to the ninth columns using the the six criteria listed above. An H represents the publication is highly matched to the correspoinding criteria (M and L represents medium and low respectively). An H indicates: 1. The PubMed entry is linked to the human gene TNF with Entrez Gene identifier GeneID:7124. 2. The publication contains both diseases related terms and genomic components. 3. The publication contains multiple Lash terms. 4. The connectivity is high, if there are more than five related gene entries linked to the publication. 5. A research article considered more important than a review or a survey, and a more specific topic is better. 6. The article is published in a journal with the impact factor higher than 10.0, or the article is cited by ten or more publications. We then score the relevance (rel.) and the importance (imp.) in the second and the third columns by combining 33 the number of H and M reported in the six criteria. Criteria 1 weighs twice compared to the other five criteria. We use a number between 0 and 5, in which 5 indicates the corresponding PubMed entry is highly relevant or highly important to the given query. While both rankings appear to identify "good" documents, Iowa appears to favor relevant documents based on their word content. lgPR appears to exploit the link structure of the RG, and have higher in-terconnectivity to TNF related entries in Entrez gene. The publications retrieved by lgPR are more likely to contain diseases related terms or genomic components. The Iowa ranking has a primary focus on the relevance of documents (based on document contents; it is not able to differentiate the importance of these relevant documents. In contrast, lgPR has a primary focus on importance (based on the link structure of the result graph); it is not able to differentiate the relevance of important documents. We conclude that further study is needed to determine how we can exploit the characteristics of both methods. There is no intersection between two sets of Top 10 publications returned by these two ranking methods. The first common PMID is 7935762, which is ranked 24 in the Iowa method and 21 by the lgPR method. CONCLUSIONS We have defined a model for life science sources. The answer to a navigational query are the target objects (TOS) of a layered graph Result Graph (RG). We define two ranking metrics layered graph PageRank (lgPR) and layered graph ObjectRank (lgOR). We also report on the results of experiments on real world data from NCBI/NIH. We show that the ranking distribution of lgPR indeed discriminates among the TOS objects of the RG. The lgPR distribution is not the same as applying PageRank a priori to the data graph. We perform a user experiment on complex queries typical of a scientist searching for gene related PubMed publications, and the Top K results of a word based ranking technique (Iowa) that has been shown to be accurate in answering gene queries the query. Using criteria that judge both relevance and importance, we explore the characteristics of these two rankings. Our preliminary evaluation indicates there may be a benefit or a meta-ranking. We briefly presented layered graph ObjectRank (lgOR) which is an extension to ObjectRank. The challenge of ObjectRank is determining the correct authority weight for each edge. For lgOR, we need to find the weight for the edges that occur in RG. Experiments with users to determine the correct authority weights for lgOR is planned for future work. We expect that IR techniques can be used to determine authority weights. REFERENCES [1] Andrey Balmin, Vagelis Hristidis, and Yannis Papakonstantinou. Objectrank: Authority-based keyword search in databases. In VLDB, pages 564575, 2004. [2] Magdalini Eirinaki, Michalis Vazirgiannis, and Dimitris Kapogiannis. Web path recommendations based on page ranking and markov models. In WIDM '05: Proceedings of the 7th annual ACM international workshop on Web information and data management, pages 29, New York, NY, USA, 2005. ACM Press. [3] Taher H. Haveliwala. Topic-sensitive pagerank. In WWW '02: Proceedings of the 11th international conference on World Wide Web, pages 517526, New York, NY, USA, 2002. ACM Press. [4] Homo sapiens in NCBI Taxonomy Browser. www.ncbi.nih.gov/Taxonomy/Browser/wwwtax.cgi? mode=Info&id=9606. [5] Sepandar D. Kamvar, Taher H. Haveliwala, Christopher D. Manning, and Gene H. Golub. Extrapolation methods for accelerating pagerank computations. In WWW, pages 261270, 2003. [6] Z. Lacroix, L. Raschid, and M.-E. Vidal. Semantic model ot integrate biological resources. In International Workshop on Semantic Web and Databases (SWDB 2006), Atlanta, Georgia, USA, 3-7 April 2006. [7] Alex Lash, Woei-Jyh Lee, and Louiqa Raschid. A methodology to enhance the semantics of links between PubMed publications and markers in the human genome. In Fifth IEEE Symposium on Bioinformatics and Bioengineering (BIBE 2005), pages 185192, Minneapolis, Minnesota, USA, 19-21 October 2005. [8] Carl D. Meyer. Matrix Analysis and Applied Linear Algebra. Society for Industrial and Applied Mathmatics, 2000. [9] G. Mihaila, F. Naumann, L. Raschid, and M. Vidal. A data model and query language to explore enhanced links and paths in life sciences data sources. Proceedings of the Workshop on Web and Databases, WebDB, Maryland, USA, 2005. [10] Rajeev Motwani and Prabhakar Raghavan. Randomized algorithms. Cambridge University Press, New York, NY, USA, 1995. [11] Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking: Bringing order to the web. Technical report, Stanford Digital Library Technologies Project, 1998. [12] Matthew Richardson and Pedro Domingos. Combining link and content information in web search. In Web Dynamics '04: Web Dynamics - Adapting to Change in Content, Size, Topology and Use, pages 179194. Springer, 2004. [13] Aditya Kumar Sehgal and Padmini Srinivasan. Retrieval with gene queries. BMC Bioinformatics, 7:220, 2006. [14] Maria-Esther Vidal, Louiqa Raschid, Natalia M arquez, Marelis C ardenas, and Yao Wu. Query rewriting in the semantic web. In InterDB, 2006. 34

Navigational Query;Link Analysis;PageRank;Ranking

165

Ranking Web Objects from Multiple Communities

Vertical search is a promising direction as it leverages domain-specific knowledge and can provide more precise information for users. In this paper, we study the Web object-ranking problem, one of the key issues in building a vertical search engine. More specifically, we focus on this problem in cases when objects lack relationships between different Web communities , and take high-quality photo search as the test bed for this investigation. We proposed two score fusion methods that can automatically integrate as many Web communities (Web forums) with rating information as possible. The proposed fusion methods leverage the hidden links discovered by a duplicate photo detection algorithm, and aims at minimizing score differences of duplicate photos in different forums . Both intermediate results and user studies show the proposed fusion methods are practical and efficient solutions to Web object ranking in cases we have described. Though the experiments were conducted on high-quality photo ranking , the proposed algorithms are also applicable to other ranking problems, such as movie ranking and music ranking

INTRODUCTION Despite numerous refinements and optimizations, general purpose search engines still fail to find relevant results for many queries. As a new trend, vertical search has shown promise because it can leverage domain-specific knowledge and is more effective in connecting users with the information they want. There are many vertical search engines, including some for paper search (e.g. Libra [21], Citeseer [7] and Google Scholar [4]), product search (e.g. Froogle [5]), movie search [6], image search [1, 8], video search [6], local search [2], as well as news search [3]. We believe the vertical search engine trend will continue to grow. Essentially, building vertical search engines includes data crawling, information extraction, object identification and integration, and object-level Web information retrieval (or Web object ranking) [20], among which ranking is one of the most important factors. This is because it deals with the core problem of how to combine and rank objects coming from multiple communities. Although object-level ranking has been well studied in building vertical search engines, there are still some kinds of vertical domains in which objects cannot be effectively ranked. For example, algorithms that evolved from PageRank [22], PopRank [21] and LinkFusion [27] were proposed to rank objects coming from multiple communities, but can only work on well-defined graphs of heterogeneous data. "Well-defined" means that like objects (e.g. authors in paper search) can be identified in multiple communities (e.g. conferences). This allows heterogeneous objects to be well linked to form a graph through leveraging all the relationships (e.g. cited-by, authored-by and published-by) among the multiple communities. However, this assumption does not always stand for some domains. High-quality photo search, movie search and news search are exceptions. For example, a photograph forum website usually includes three kinds of objects: photos, authors and reviewers. Yet different photo forums seem to lack any relationships, as there are no cited-by relationships. This makes it difficult to judge whether two authors cited are the same author, or two photos are indeed identical photos . Consequently, although each photo has a rating score in a forum, it is non-trivial to rank photos coming from different photo forums. Similar problems also exist in movie search and news search. Although two movie titles can be identified as the same one by title and director in different movie discussion groups, it is non-trivial to combine rating scores from different discussion groups and rank movies effectively. We call such non-trivial object relationship in which identification is difficult, incomplete relationships. Other related work includes rank aggregation for the Web [13, 14], and learning algorithm for rank, such as RankBoost [15], RankSVM [17, 19], and RankNet [12]. We will contrast differences of these methods with the proposed methods after we have described the problem and our methods. We will specifically focus on Web object-ranking problem in cases that lack object relationships or have with incomplete object relationships, and take high-quality photo search as the test bed for this investigation. In the following, we will introduce rationale for building high-quality photo search. 1.1 High-Quality Photo Search In the past ten years, the Internet has grown to become an incredible resource, allowing users to easily access a huge number of images. However, compared to the more than 1 billion images indexed by commercial search engines, actual queries submitted to image search engines are relatively minor , and occupy only 8-10 percent of total image and text queries submitted to commercial search engines [24]. This is partially because user requirements for image search are far less than those for general text search. On the other hand, current commercial search engines still cannot well meet various user requirements, because there is no effective and practical solution to understand image content. To better understand user needs in image search, we conducted a query log analysis based on a commercial search engine. The result shows that more than 20% of image search queries are related to nature and places and daily life categories. Users apparently are interested in enjoying high-quality photos or searching for beautiful images of locations or other kinds. However, such user needs are not well supported by current image search engines because of the difficulty of the quality assessment problem. Ideally, the most critical part of a search engine the ranking function can be simplified as consisting of two key factors: relevance and quality. For the relevance factor , search in current commercial image search engines provide most returned images that are quite relevant to queries, except for some ambiguity. However, as to quality factor, there is still no way to give an optimal rank to an image. Though content-based image quality assessment has been investigated over many years [23, 25, 26], it is still far from ready to provide a realistic quality measure in the immediate future. Seemingly, it really looks pessimistic to build an image search engine that can fulfill the potentially large requirement of enjoying high-quality photos. Various proliferating Web communities, however, notices us that people today have created and shared a lot of high-quality photos on the Web on virtually any topics, which provide a rich source for building a better image search engine. In general, photos from various photo forums are of higher quality than personal photos, and are also much more appealing to public users than personal photos. In addition, photos uploaded to photo forums generally require rich metadata about title, camera setting, category and description to be provide by photographers. These metadata are actually the most precise descriptions for photos and undoubtedly can be indexed to help search engines find relevant results. More important, there are volunteer users in Web communities actively providing valuable ratings for these photos. The rating information is generally of great value in solving the photo quality ranking problem. Motivated by such observations, we have been attempting to build a vertical photo search engine by extracting rich metadata and integrating information form various photo Web forums. In this paper, we specifically focus on how to rank photos from multiple Web forums. Intuitively, the rating scores from different photo forums can be empirically normalized based on the number of photos and the number of users in each forum. However, such a straightforward approach usually requires large manual effort in both tedious parameter tuning and subjective results evaluation, which makes it impractical when there are tens or hundreds of photo forums to combine. To address this problem, we seek to build relationships/links between different photo forums. That is, we first adopt an efficient algorithm to find duplicate photos which can be considered as hidden links connecting multiple forums. We then formulate the ranking challenge as an optimization problem, which eventually results in an optimal ranking function. 1.2 Main Contributions and Organization. The main contributions of this paper are: 1. We have proposed and built a vertical image search engine by leveraging rich metadata from various photo forum Web sites to meet user requirements of searching for and enjoying high-quality photos, which is impossible in traditional image search engines. 2. We have proposed two kinds of Web object-ranking algorithms for photos with incomplete relationships, which can automatically and efficiently integrate as many as possible Web communities with rating information and achieves an equal qualitative result compared with the manually tuned fusion scheme. The rest of this paper is organized as follows. In Section 2, we present in detail the proposed solutions to the ranking problem, including how to find hidden links between different forums, normalize rating scores, obtain the optimal ranking function, and contrast our methods with some other related research. In Section 3, we describe the experimental setting and experiments and user studies conducted to evaluate our algorithm. Our conclusion and a discussion of future work is in Section 4. It is worth noting that although we treat vertical photo search as the test bed in this paper, the proposed ranking algorithm can also be applied to rank other content that includes video clips, poems, short stories, drawings, sculptures , music, and so on. 378 ALGORITHM The difficulty of integrating multiple Web forums is in their different rating systems, where there are generally two kinds of freedom. The first kind of freedom is the rating interval or rating scale including the minimal and maximal ratings for each Web object. For example, some forums use a 5-point rating scale whereas other forums use 3-point or 10-point rating scales. It seems easy to fix this freedom, but detailed analysis of the data and experiments show that it is a non-trivial problem. The second kind of freedom is the varying rating criteria found in different Web forums. That is, the same score does not mean the same quality in different forums. Intuitively, if we can detect same photographers or same photographs, we can build relationships between any two photo forums and therefore can standardize the rating criterion by score normalization and transformation. Fortunately, we find that quite a number of duplicate photographs exist in various Web photo forums. This fact is reasonable when considering that photographers sometimes submit a photo to more than one forum to obtain critiques or in hopes of widespread publicity. In this work, we adopt an efficient duplicate photo detection algorithm [10] to find these photos. The proposed methods below are based on the following considerations. Faced with the need to overcome a ranking problem, a standardized rating criterion rather than a reasonable rating criterion is needed. Therefore, we can take a large scale forum as the reference forum, and align other forums by taking into account duplicate Web objects (duplicate photos in this work). Ideally, the scores of duplicate photos should be equal even though they are in different forums. Yet we can deem that scores in different forums except for the reference forum can vary in a parametric space. This can be determined by minimizing the objective function defined by the sum of squares of the score differences . By formulating the ranking problem as an optimization problem that attempts to make the scores of duplicate photos in non-reference forums as close as possible to those in the reference forum, we can effectively solve the ranking problem. For convenience, the following notations are employed. S ki and S ki denote the total score and mean score of ith Web object (photo) in the kth Web site, respectively. The total score refers to the sum of the various rating scores (e.g., novelty rating and aesthetic rating), and the mean score refers to the mean of the various rating scores. Suppose there are a total of K Web sites. We further use {S kl i |i = 1, ..., I kl ; k, l = 1, ..., K; k = l} to denote the set of scores for Web objects (photos) in kth Web forums that are duplicate with the lth Web forums, where I kl is the total number of duplicate Web objects between these two Web sites. In general, score fusion can be seen as the procedure of finding K transforms k ( S ki ) = e S ki , k = 1, ..., K such that e S ki can be used to rank Web objects from different Web sites. The objective function described in the above Figure 1: Web community integration. Each Web community forms a subgraph, and all communities are linked together by some hidden links (dashed lines). paragraph can then be formulated as min { k |k=2,...,K} K X k=2 I k1 X i=1 w k i "S 1k i k ( S k1 i ) " 2 (1) where we use k = 1 as the reference forum and thus 1 ( S 1i ) = S 1i . w k i ( 0) is the weight coefficient that can be set heuris-tically according to the numbers of voters (reviewers or com-menters ) in both the reference forum and the non-reference forum. The more reviewers, the more popular the photo is and the larger the corresponding weight w k i should be. In this work, we do not inspect the problem of how to choose w k i and simply set them to one. But we believe the proper use of w k i , which leverages more information, can significantly improve the results. Figure 1 illustrates the aforementioned idea. The Web Community 1 is the reference community. The dashed lines are links indicating that the two linked Web objects are actually the same. The proposed algorithm will try to find the best k ( k = 2, ..., K), which has certain parametric forms according to certain models. So as to minimize the cost function defined in Eq. 1, the summation is taken on all the red dashed lines. We will first discuss the score normalization methods in Section 2.2, which serves as the basis for the following work. Before we describe the proposed ranking algorithms, we first introduce a manually tuned method in Section 2.3, which is laborious and even impractical when the number of communities become large. In Section 2.4, we will briefly explain how to precisely find duplicate photos between Web forums. Then we will describe the two proposed methods: Linear fusion and Non-linear fusion, and a performance measure for result evaluation in Section 2.5. Finally, in Section 2.6 we will discuss the relationship of the proposed methods with some other related work. 2.2 Score Normalization Since different Web (photo) forums on the Web usually have different rating criteria, it is necessary to normalize them before applying different kinds of fusion methods. In addition, as there are many kinds of ratings, such as ratings for novelty, ratings for aesthetics etc, it is reasonable to choose a common one -- total score or average score -that can always be extracted in any Web forum or calculated by corresponding ratings. This allows the normaliza-379 tion method on the total score or average score to be viewed as an impartial rating method between different Web forums . It is straightforward to normalize average scores by lin-early transforming them to a fixed interval. We call this kind of score as Scaled Mean Score. The difficulty, however, of using this normalization method is that, if there are only a few users rating an object, say a photo in a photo forum, the average score for the object is likely to be spammed or skewed. Total score can avoid such drawbacks that contain more information such as a Web object's quality and popularity. The problem is thus how to normalize total scores in different Web forums. The simplest way may be normalization by the maximal and minimal scores. The drawback of this normalization method is it is non robust, or in other words, it is sensitive to outliers. To make the normalization insensitive to unusual data, we propose the Mode-90% Percentile normalization method. Here, the mode score represents the total score that has been assigned to more photos than any other total score. And The high percentile score (e.g.,90%) represents the total score for which the high percentile of images have a lower total score. This normalization method utilizes the mode and 90% percentile as two reference points to align two rating systems, which makes the distributions of total scores in different forums more consistent. The underlying assumption, for example in different photo forums, is that even the qualities of top photos in different forums may vary greatly and be less dependent on the forum quality, the distribution of photos of middle-level quality (from mode to 90% percentile) should be almost of the same quality up to the freedom which reflects the rating criterion (strictness) of Web forums. Photos of this middle-level in a Web forum usually occupy more than 70 % of total photos in that forum. We will give more detailed analysis of the scores in Section 3.2. 2.3 Manual Fusion The Web movie forum, IMDB [16], proposed to use a Bayesian-ranking function to normalize rating scores within one community. Motivated by this ranking function, we propose this manual fusion method: For the kth Web site, we use the following formula e S ki = k ,, n k S ki n k + n k + n k S k n k + n k (2) to rank photos, where n k is the number of votes and n k , S k and k are three parameters. This ranking function first takes a balance between the original mean score S ki and a reference score S k to get a weighted mean score which may be more reliable than S ki . Then the weighted mean score is scaled by k to get the final score f S ki . For n Web communities, there are then about 3n parameters in {( k , n k , S k ) |k = 1, ..., n} to tune. Though this method can achieves pretty good results after careful and thorough manual tuning on these parameters, when n becomes increasingly large, say there are tens or hundreds of Web communities crawled and indexed, this method will become more and more laborious and will eventually become impractical. It is therefore desirable to find an effective fusion method whose parameters can be automatically determined . 2.4 Duplicate Photo Detection We use Dedup [10], an efficient and effective duplicate image detection algorithm, to find duplicate photos between any two photo forums. This algorithm uses hash function to map a high dimensional feature to a 32 bits hash code (see below for how to construct the hash code). Its computational complexity to find all the duplicate images among n images is about O(n log n). The low-level visual feature for each photo is extracted on k k regular grids. Based on all features extracted from the image database, a PCA model is built. The visual features are then transformed to a relatively low-dimensional and zero mean PCA space, or 29 dimensions in our system. Then the hash code for each photo is built as follows: each dimension is transformed to one, if the value in this dimension is greater than 0, and 0 otherwise. Photos in the same bucket are deemed potential duplicates and are further filtered by a threshold in terms of Euclidean similarity in the visual feature space. Figure 2 illustrates the hashing procedure, where visual features -- mean gray values -- are extracted on both 6 6 and 7 7 grids. The 85-dimensional features are transformed to a 32-dimensional vector, and the hash code is generated according to the signs. Figure 2: Hashing procedure for duplicate photo dectection 2.5 Score Fusion In this section, we will present two solutions on score fusion based on different parametric form assumptions of k in Eq. 1. 2.5.1 Linear Fusion by Duplicate Photos Intuitively, the most straightforward way to factor out the uncertainties caused by the different criterion is to scale, rel-380 ative to a given center, the total scores of each unreferenced Web photo forum with respect to the reference forum. More strictly, we assume k has the following form k ( S ki ) = k S ki + t k , k = 2, ..., K (3) 1 ( S 1i ) = S 1i (4) which means that the scores of k(= 1)th forum should be scaled by k relative to the center t k 1k as shown in Figure 3. Then, if we substitute above k to Eq. 1, we get the following objective function, min { k ,t k |k=2,...,K} K X k=2 I k1 X i=1 w k i hS 1k i k S k1 i - t k i 2 . (5) By solving the following set of functions, ( f k = = 0 f t k = 0 , k = 1, ..., K where f is the objective function defined in Eq. 5, we get the closed form solution as: ,, k t k = A -1 k L k (6) where A k = ,, P i w i ( S k1 i ) 2 P i w i S k1 i P i w i S k1 i P i w i (7) L k = ,, P i w i S 1k i S k1 i P i w i S 1k i (8) and k = 2, ..., K. This is a linear fusion method. It enjoys simplicity and excellent performance in the following experiments. Figure 3: Linear Fusion method 2.5.2 Nonlinear Fusion by Duplicate Photos Sometimes we want a method which can adjust scores on intervals with two endpoints unchanged. As illustrated in Figure 4, the method can tune scores between [ C 0 , C 1 ] while leaving scores C 0 and C 1 unchanged. This kind of fusion method is then much finer than the linear ones and contains many more parameters to tune and expect to further improve the results. Here, we propose a nonlinear fusion solution to satisfy such constraints. First, we introduce a transform: c 0 ,c 1 , ( x) = ( " x-c 0 c 1 -c 0 " ( c 1 - c 0 ) + c 0 , if x (c 0 , c 1 ] x otherwise where &gt; 0. This transform satisfies that for x [c 0 , c 1 ], c 0 ,c 1 , ( x) [c 0 , c 1 ] with c 0 ,c 1 , ( c 0 ) = c 0 and c 0 ,c 1 , ( c 1 ) = c 1 . Then we can utilize this nonlinear transform to adjust the scores in certain interval, say ( M, T ], k ( S ki ) = M,T, ( S ki ) . (9) Figure 4: Nonlinear Fusion method. We intent to finely adjust the shape of the curves in each segment. Even there is no closed-form solution for the following optimization problem, min { k |k[2,K]} K X k=2 I k1 X i=1 w k i hS 1k i M ,T, ( S ki ) i 2 it is not hard to get the numeric one. Under the same assumptions made in Section 2.2, we can use this method to adjust scores of the middle-level (from the mode point to the 90 % percentile). This more complicated non-linear fusion method is expected to achieve better results than the linear one. However , difficulties in evaluating the rank results block us from tuning these parameters extensively. The current experiments in Section 3.5 do not reveal any advantages over the simple linear model. 2.5.3 Performance Measure of the Fusion Results Since our objective function is to make the scores of the same Web objects (e.g. duplicate photos) between a non-reference forum and the reference forum as close as possible, it is natural to investigate how close they become to each other and how the scores of the same Web objects change between the two non-reference forums before and after score fusion. Taken Figure 1 as an example, the proposed algorithms minimize the score differences of the same Web objects in two Web forums: the reference forum (the Web Community 1) and a non-reference forum, which corresponds to minimizing the objective function on the red dashed (hidden) links. After the optimization, we must ask what happens to the score differences of the same Web objects in two non-reference forums? Or, in other words, whether the scores of two objects linked by the green dashed (hidden) links become more consistent? We therefore define the following performance measure -measure -- to quantify the changes for scores of the same Web objects in different Web forums as kl = Sim(S lk , S kl ) - Sim(S lk , S kl ) (10) 381 where S kl = ( S kl 1 , ..., S kl I kl ) T , S kl = ( e S kl 1 , ..., e S kl I kl ) T and Sim(a , b) = a b ||a||||b|| . kl &gt; 0 means after score fusion, scores on the same Web objects between kth and lth Web forum become more consistent , which is what we expect. On the contrary, if kl &lt; 0, those scores become more inconsistent. Although we cannot rely on this measure to evaluate our final fusion results as ranking photos by their popularity and qualities is such a subjective process that every person can have its own results, it can help us understand the intermediate ranking results and provide insights into the final performances of different ranking methods. 2.6 Contrasts with Other Related Work We have already mentioned the differences of the proposed methods with the traditional methods, such as PageRank [22], PopRank [21], and LinkFusion [27] algorithms in Section 1. Here, we discuss some other related works. The current problem can also be viewed as a rank aggregation one [13, 14] as we deal with the problem of how to combine several rank lists. However, there are fundamental differences between them. First of all, unlike the Web pages, which can be easily and accurately detected as the same pages, detecting the same photos in different Web forums is a non-trivial work, and can only be implemented by some delicate algorithms while with certain precision and recall. Second, the numbers of the duplicate photos from different Web forums are small relative to the whole photo sets (see Table 1). In another words, the top K rank lists of different Web forums are almost disjointed for a given query. Under this condition, both the algorithms proposed in [13] and their measurements -- Kendall tau distance or Spearman footrule distance -- will degenerate to some trivial cases. Another category of rank fusion (aggregation) methods is based on machine learning algorithms, such as RankSVM [17, 19], RankBoost [15], and RankNet [12]. All of these methods entail some labelled datasets to train a model. In current settings, it is difficult or even impossible to get these datasets labelled as to their level of professionalism or popularity , since the photos are too vague and subjective to rank. Instead, the problem here is how to combine several ordered sub lists to form a total order list. EXPERIMENTS In this section, we carry out our research on high-quality photo search. We first briefly introduce the newly proposed vertical image search engine -- EnjoyPhoto in section 3.1. Then we focus on how to rank photos from different Web forums. In order to do so, we first normalize the scores (ratings) for photos from different multiple Web forums in section 3.2. Then we try to find duplicate photos in section 3.3. Some intermediate results are discussed using measure in section 3.4. Finally a set of user studies is carried out carefully to justify our proposed method in section 3.5. 3.1 EnjoyPhoto: high-quality Photo Search Engine In order to meet user requirement of enjoying high-quality photos, we propose and build a high-quality photo search engine -- EnjoyPhoto, which accounts for the following three key issues: 1. how to crawl and index photos, 2. how to determine the qualities of each photo and 3. how to display the search results in order to make the search process enjoyable. For a given text based query, this system ranks the photos based on certain combination of relevance of the photo to this query (Issue 1) and the quality of the photo (Issue 2), and finally displays them in an enjoyable manner (Issue 3). As for Issue 3, we devise the interface of the system de-liberately in order to smooth the users' process of enjoying high-quality photos. Techniques, such as Fisheye and slides show, are utilized in current system. Figure 5 shows the interface. We will not talk more about this issue as it is not an emphasis of this paper. Figure 5: EnjoyPhoto: an enjoyable high-quality photo search engine, where 26,477 records are returned for the query "fall" in about 0.421 seconds As for Issue 1, we extracted from a commercial search engine a subset of photos coming from various photo forums all over the world, and explicitly parsed the Web pages containing these photos. The number of photos in the data collection is about 2.5 million. After the parsing, each photo was associated with its title, category, description, camera setting, EXIF data 1 (when available for digital images), location (when available in some photo forums), and many kinds of ratings. All these metadata are generally precise descriptions or annotations for the image content, which are then indexed by general text-based search technologies [9, 18, 11]. In current system, the ranking function was specifically tuned to emphasize title, categorization, and rating information. Issue 2 is essentially dealt with in the following sections which derive the quality of photos by analyzing ratings provided by various Web photo forums. Here we chose six photo forums to study the ranking problem and denote them as Web-A, Web-B, Web-C, Web-D, Web-E and Web-F. 3.2 Photo Score Normalization Detailed analysis of different score normalization methods are analyzed in this section. In this analysis, the zero 1 Digital cameras save JPEG (.jpg) files with EXIF (Exchangeable Image File) data. Camera settings and scene information are recorded by the camera into the image file. www.digicamhelp.com/what-is-exif/ 382 0 2 4 6 8 10 0 1000 2000 3000 4000 Normalized Score Total Number (a) Web-A 0 2 4 6 8 10 0 0.5 1 1.5 2 2.5 3 x 10 4 Normalized Score Total Number (b) Web-B 0 2 4 6 8 10 0 0.5 1 1.5 2 x 10 5 Normalized Score Total Number (c) Web-C 0 2 4 6 8 10 0 2 4 6 8 10 x 10 4 Normalized Score Total Number (d) Web-D 0 2 4 6 8 10 0 2000 4000 6000 8000 10000 12000 14000 Normalized Score Total Number (e) Web-E 0 2 4 6 8 10 0 1 2 3 4 5 6 x 10 4 Normalized Score Total Number (f) Web-F Figure 6: Distributions of mean scores normalized to [0 , 10] scores that usually occupy about than 30% of the total number of photos for some Web forums are not currently taken into account. How to utilize these photos is left for future explorations. In Figure 6, we list the distributions of the mean score, which is transformed to a fixed interval [0 , 10]. The distributions of the average scores of these Web forums look quite different. Distributions in Figure 6(a), 6(b), and 6(e) looks like Gaussian distributions, while those in Figure 6(d) and 6(f) are dominated by the top score. The reason of these eccentric distributions for Web-D and Web-F lies in their coarse rating systems. In fact, Web-D and Web-F use 2 or 3 point rating scales whereas other Web forums use 7 or 14 point rating scales. Therefore, it will be problematic if we directly use these averaged scores. Furthermore the average score is very likely to be spammed, if there are only a few users rating a photo. Figure 7 shows the total score normalization method by maximal and minimal scores, which is one of our base line system. All the total scores of a given Web forum are normalized to [0 , 100] according to the maximal score and minimal score of corresponding Web forum. We notice that total score distribution of Web-A in Figure 7(a) has two larger tails than all the others. To show the shape of the distributions more clearly, we only show the distributions on [0 , 25] in Figure 7(b),7(c),7(d),7(e), and 7(f). Figure 8 shows the Mode-90% Percentile normalization method, where the modes of the six distributions are normalized to 5 and the 90% percentile to 8. We can see that this normalization method makes the distributions of total scores in different forums more consistent. The two proposed algorithms are all based on these normalization methods. 3.3 Duplicate photo detection Targeting at computational efficiency, the Dedup algorithm may lose some recall rate, but can achieve a high precision rate. We also focus on finding precise hidden links rather than all hidden links. Figure 9 shows some duplicate detection examples. The results are shown in Table 1 and verify that large numbers of duplicate photos exist in any two Web forums even with the strict condition for Dedup where we chose first 29 bits as the hash code. Since there are only a few parameters to estimate in the proposed fusion methods, the numbers of duplicate photos shown Table 1 are 0 20 40 60 80 100 0 100 200 300 400 500 600 Normalized Score Total Number (a) Web-A 0 5 10 15 20 25 0 1 2 3 4 5 x 10 4 Normalized Score Total Number (b) Web-B 0 5 10 15 20 25 0 1 2 3 4 5 x 10 5 Normalized Score Total Number (c) Web-C 0 5 10 15 20 25 0 0.5 1 1.5 2 2.5 x 10 4 Normalized Score Total Number (d) Web-D 0 5 10 15 20 25 0 2000 4000 6000 8000 10000 Normalized Score Total Number (e) Web-E 0 5 10 15 20 25 0 0.5 1 1.5 2 2.5 3 x 10 4 Normalized Score Total Number (f) Web-F Figure 7: Maxmin Normalization 0 5 10 15 0 200 400 600 800 1000 1200 1400 Normalized Score Total Number (a) Web-A 0 5 10 15 0 1 2 3 4 5 x 10 4 Normalized Score Total Number (b) Web-B 0 5 10 15 0 2 4 6 8 10 12 14 x 10 4 Normalized Score Total Number (c) Web-C 0 5 10 15 0 0.5 1 1.5 2 2.5 x 10 4 Normalized Score Total Number (d) Web-D 0 5 10 15 0 2000 4000 6000 8000 10000 12000 Normalized Score Total Number (e) Web-E 0 5 10 15 0 2000 4000 6000 8000 10000 Normalized Score Total Number (f) Web-F Figure 8: Mode-90% Percentile Normalization sufficient to determine these parameters. The last table column lists the total number of photos in the corresponding Web forums. 3.4 Measure The parameters of the proposed linear and nonlinear algorithms are calculated using the duplicate data shown in Table 1, where the Web-C is chosen as the reference Web forum since it shares the most duplicate photos with other forums. Table 2 and 3 show the measure on the linear model and nonlinear model. As kl is symmetric and kk = 0, we only show the upper triangular part. The NaN values in both tables lie in that no duplicate photos have been detected by the Dedup algorithm as reported in Table 1. The linear model guarantees that the measures related Table 1: Number of duplicate photos between each pair of Web forums A B C D E F Scale A 0 316 1,386 178 302 0 130k B 316 0 14,708 909 8,023 348 675k C 1,386 14,708 0 1,508 19,271 1,083 1,003k D 178 909 1,508 0 1,084 21 155k E 302 8,023 19,271 1,084 0 98 448k F 0 348 1,083 21 98 0 122k 383 Figure 9: Some results of duplicate photo detection Table 2: measure on the linear model. Web-B Web-C Web-D Web-E Web-F Web-A 0.0659 0.0911 0.0956 0.0928 NaN Web-B 0.0672 0.0578 0.0791 0.4618 Web-C 0.0105 0.0070 0.2220 Web-D 0.0566 0.0232 Web-E 0.6525 to the reference community should be no less than 0 theo-retically . It is indeed the case (see the underlined numbers in Table 2). But this model can not guarantee that the measures on the non-reference communities can also be no less than 0, as the normalization steps are based on duplicate photos between the reference community and a non-reference community. Results shows that all the numbers in the measure are greater than 0 (see all the non-underlined numbers in Table 2), which indicates that it is probable that this model will give optimal results. On the contrary, the nonlinear model does not guarantee that measures related to the reference community should be no less than 0, as not all duplicate photos between the two Web forums can be used when optimizing this model. In fact, the duplicate photos that lie in different intervals will not be used in this model. It is these specific duplicate photos that make the measure negative. As a result, there are both negative and positive items in Table 3, but overall the number of positive ones are greater than negative ones (9:5), that indicates the model may be better than the "nor-malization only" method (see next subsection) which has an all-zero measure, and worse than the linear model. 3.5 User Study Because it is hard to find an objective criterion to evaluate Table 3: measure on the nonlinear model. Web-B Web-C Web-D Web-E Web-F Web-A 0.0559 0.0054 -0.0185 -0.0054 NaN Web-B -0.0162 -0.0345 -0.0301 0.0466 Web-C 0.0136 0.0071 0.1264 Web-D 0.0032 0.0143 Web-E 0.214 which ranking function is better, we chose to employ user studies for subjective evaluations. Ten subjects were invited to participate in the user study. They were recruited from nearby universities. As search engines of both text search and image search are familiar to university students, there was no prerequisite criterion for choosing students. We conducted user studies using Internet Explorer 6.0 on Windows XP with 17-inch LCD monitors set at 1,280 pixels by 1,024 pixels in 32-bit color. Data was recorded with server logs and paper-based surveys after each task. Figure 10: User study interface We specifically device an interface for user study as shown in Figure 10. For each pair of fusion methods, participants were encouraged to try any query they wished. For those without specific ideas, two combo boxes (category list and query list) were listed on the bottom panel, where the top 1,000 image search queries from a commercial search engine were provided. After a participant submitted a query, the system randomly selected the left or right frame to display each of the two ranking results. The participant were then required to judge which ranking result was better of the two ranking results, or whether the two ranking results were of equal quality, and submit the judgment by choosing the corresponding radio button and clicking the "Submit" button. For example, in Figure 10, query "sunset" is submitted to the system. Then, 79,092 photos were returned and ranked by the Minmax fusion method in the left frame and linear fusion method in the right frame. A participant then compares the two ranking results (without knowing the ranking methods) and submits his/her feedback by choosing answers in the "Your option." Table 4: Results of user study Norm.Only Manually Linear Linear 29:13:10 14:22:15 -Nonlinear 29:15:9 12:27:12 6:4:45 Table 4 shows the experimental results, where "Linear" denotes the linear fusion method, "Nonlinear" denotes the non linear fusion method, "Norm. Only" means Maxmin normalization method, "Manually" means the manually tuned method. The three numbers in each item, say 29:13:10, mean that 29 judgments prefer the linear fusion results, 10 384 judgments prefer the normalization only method, and 13 judgments consider these two methods as equivalent. We conduct the ANOVA analysis, and obtain the following conclusions: 1. Both the linear and nonlinear methods are significantly better than the "Norm. Only" method with respective P-values 0 .00165(&lt; 0.05) and 0.00073(&lt;&lt; 0.05). This result is consistent with the -measure evaluation result . The "Norm. Only" method assumes that the top 10% photos in different forums are of the same quality . However, this assumption does not stand in general . For example, a top 10% photo in a top tier photo forum is generally of higher quality than a top 10% photo in a second-tier photo forum. This is similar to that, those top 10% students in a top-tier university and those in a second-tier university are generally of different quality. Both linear and nonlinear fusion methods acknowledge the existence of such differences and aim at quantizing the differences. Therefore, they perform better than the "Norm. Only" method. 2. The linear fusion method is significantly better than the nonlinear one with P-value 1 .195 10 -10 . This result is rather surprising as this more complicated ranking method is expected to tune the ranking more finely than the linear one. The main reason for this result may be that it is difficult to find the best intervals where the nonlinear tuning should be carried out and yet simply the middle part of the Mode-90% Percentile Normalization method was chosen. The time-consuming and subjective evaluation methods -- user studies -- blocked us extensively tuning these parameters . 3. The proposed linear and nonlinear methods perform almost the same with or slightly better than the manually tuned method. Given that the linear/nonlinear fusion methods are fully automatic approaches, they are considered practical and efficient solutions when more communities (e.g. dozens of communities) need to be integrated. CONCLUSIONS AND FUTURE WORK In this paper, we studied the Web object-ranking problem in the cases of lacking object relationships where traditional ranking algorithms are no longer valid, and took high-quality photo search as the test bed for this investigation . We have built a vertical high-quality photo search engine, and proposed score fusion methods which can automatically integrate as many data sources (Web forums) as possible. The proposed fusion methods leverage the hidden links discovered by duplicate photo detection algorithm, and minimize score differences of duplicate photos in different forums. Both the intermediate results and the user studies show that the proposed fusion methods are a practical and efficient solution to Web object ranking in the aforesaid relationships. Though the experiments were conducted on high-quality photo ranking, the proposed algorithms are also applicable to other kinds of Web objects including video clips, poems, short stories, music, drawings, sculptures, and so on. Current system is far from being perfect. In order to make this system more effective, more delicate analysis for the vertical domain (e.g., Web photo forums) are needed. The following points, for example, may improve the searching results and will be our future work: 1. more subtle analysis and then utilization of different kinds of ratings (e.g., novelty ratings, aesthetic ratings); 2. differentiating various communities who may have different interests and preferences or even distinct culture understandings; 3. incorporating more useful information, including photographers' and reviewers' information, to model the photos in a heterogeneous data space instead of the current homogeneous one. We will further utilize collaborative filtering to recommend relevant high-quality photos to browsers. One open problem is whether we can find an objective and efficient criterion for evaluating the ranking results, instead of employing subjective and inefficient user studies, which blocked us from trying more ranking algorithms and tuning parameters in one algorithm. ACKNOWLEDGMENTS We thank Bin Wang and Zhi Wei Li for providing Dedup codes to detect duplicate photos; Zhen Li for helping us design the interface of EnjoyPhoto; Ming Jing Li, Longbin Chen, Changhu Wang, Yuanhao Chen, and Li Zhuang etc. for useful discussions. Special thanks go to Dwight Daniels for helping us revise the language of this paper. REFERENCES [1] Google image search. http://images.google.com. [2] Google local search. http://local.google.com/. [3] Google news search. http://news.google.com. [4] Google paper search. http://Scholar.google.com. [5] Google product search. http://froogle.google.com. [6] Google video search. http://video.google.com. [7] Scientific literature digital library. http://citeseer.ist.psu.edu. [8] Yahoo image search. http://images.yahoo.com. [9] R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. New York: ACM Press; Harlow, England: Addison-Wesley, 1999. [10] W. Bin, L. Zhiwei, L. Ming Jing, and M. Wei-Ying. Large-scale duplicate detection for web image search. In Proceedings of the International Conference on Multimedia and Expo, page 353, 2006. [11] S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. In Computer Networks, volume 30, pages 107117, 1998. [12] C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, and G. Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd international conference on Machine learning, pages 89 96, 2005. [13] C. Dwork, R. Kumar, M. Naor, and D. Sivakumar. Rank aggregation methods for the web. In Proceedings 10th International Conference on World Wide Web, pages 613 622, Hong-Kong, 2001. [14] R. Fagin, R. Kumar, and D. Sivakumar. Comparing top k lists. SIAM Journal on Discrete Mathematics, 17(1):134 160, 2003. [15] Y. Freund, R. Iyer, R. E. Schapire, and Y. Singer. An efficient boosting algorithm for combining preferences. 385 Journal of Machine Learning Research, 4(1):933969(37), 2004. [16] IMDB. Formula for calculating the top rated 250 titles in imdb. http://www.imdb.com/chart/top. [17] T. Joachims. Optimizing search engines using clickthrough data. In Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 133 142, 2002. [18] J. M. Kleinberg. Authoritative sources in a hyperlinked environment. Journal of the ACM, 46(5):604632, 1999. [19] R. Nallapati. Discriminative models for information retrieval. In Proceedings of the 25th annual international ACM SIGIR conference on Research and development in information retrieval, pages 64 71, 2004. [20] Z. Nie, Y. Ma, J.-R. Wen, and W.-Y. Ma. Object-level web information retrieval. In Technical Report of Microsoft Research, volume MSR-TR-2005-11, 2005. [21] Z. Nie, Y. Zhang, J.-R. Wen, and W.-Y. Ma. Object-level ranking: Bringing order to web objects. In Proceedings of the 14th international conference on World Wide Web, pages 567 574, Chiba, Japan, 2005. [22] L. Page, S. Brin, R. Motwani, and T. Winograd. The pagerank citation ranking: Bringing order to the web. In Technical report, Stanford Digital Libraries, 1998. [23] A. Savakis, S. Etz, and A. Loui. Evaluation of image appeal in consumer photography. In SPIE Human Vision and Electronic Imaging, pages 111120, 2000. [24] D. Sullivan. Hitwise search engine ratings. Search Engine Watch Articles, http://searchenginewatch. com/reports/article.php/3099931, August 23, 2005. [25] S. Susstrunk and S. Winkler. Color image quality on the internet. In IS&T/SPIE Electronic Imaging 2004: Internet Imaging V, volume 5304, pages 118131, 2004. [26] H. Tong, M. Li, Z. H.J., J. He, and Z. C.S. Classification of digital photos taken by photographers or home users. In Pacific-Rim Conference on Multimedia (PCM), pages 198205, 2004. [27] W. Xi, B. Zhang, Z. Chen, Y. Lu, S. Yan, W.-Y. Ma, and E. A. Fox. Link fusion: a unified link analysis framework for multi-type interrelated data objects. In Proceedings of the 13th international conference on World Wide Web, pages 319 327, 2004. 386

image search;ranking;Web objects

166

Real-world Oriented Information Sharing Using Social Networks

While users disseminate various information in the open and widely distributed environment of the Semantic Web, determination of who shares access to particular information is at the center of looming privacy concerns. We propose a real-world -oriented information sharing system that uses social networks. The system automatically obtains users' social relationships by mining various external sources. It also enables users to analyze their social networks to provide awareness of the information dissemination process. Users can determine who has access to particular information based on the social relationships and network analysis.

INTRODUCTION With the current development of tools and sites that enable users to create Web content, users have become able to easily disseminate various information. For example, users create Weblogs, which are diary-like sites that include various public and private information. Furthermore, the past year has witnessed the emergence of social networking sites that allow users to maintain an online network of friends or associates for social or business purposes. Therein, data related to millions of people and their relationships are publicly available on the Web. Although these tools and sites enable users to easily disseminate information on the Web, users sometimes have difficulty in sharing information with the right people and frequently have privacy concerns because it is difficult to determine who has access to particular information on such applications. Some tools and applications provide control over information access. For example, Friendster, a huge social networking site, offers several levels of control from "public information" to "only for friends". However, it provides only limited support for access control. An appropriate information sharing system that enables all users to control the dissemination of their information is needed to use tools and sites such as Weblog, Wiki, and social networking services fully as an infrastructure of disseminating and sharing information. In the absence of such a system, a user would feel unsafe and would therefore be discouraged from disseminating information. How can we realize such an information sharing system on the Web? One clue exists in the information sharing processes of the real world. Information availability is often closely guarded and shared only with the people of one's social relationships. Confidential project documents which have limited distribution within a division of company, might be made accessible to other colleagues who are concerned with the project. Private family photographs might be shared not only with relatives, but also with close friends. A professor might access a private research report of her student. We find that social relationships play an important role in the process of disseminating and receiving information. This paper presents a real-world oriented information sharing system using social networks. It enables users to control the information dissemination process within social networks. The remainder of this paper is organized as follows: section 2 describes the proposed information sharing system using social networks. In section 3, we describe the application of our system. Finally, we conclude this paper in section 4. INFORMATION SHARING USING SOCIAL NETWORKS Figure 1 depicts the architecture of the proposed information sharing system. The system functions as a "plug-in" for applications so that external applications enable users to leverage social networks to manage their information dis-81 Social network analysis Applications for Information Sharing e.g. Weblog, Wiki, CMS, SNS etc Contents Data Access control List Editor Access control Data (XACML) Social networks extraction Web Web pages, SNS, FOAF, etc Email Sensors Social networks Editor Social networks Data (FOAF) Edit social networks Edit access control list Contents data Access to contents Access permission User data Figure 1: Architecture of the proposed information sharing system Birthplace : Kagawa, Japan Workplace : AIST Job : CS researcher Research topics : Web University : Tokyo univ. Interest : Sumo wrestling ... person Properties Birthplace : Los Angels, US Workplace : Washington Univ. Job : CS researcher Research topics : Web University : UC California Interest : Sumo wrestling ... person Properties Birthplace : Kagawa, Japan Workplace : AIST Job : CS researcher Research topics : Web University : Tokyo univ. Interest : Sumo wrestling ... person Properties Birthplace : Los Angels, US Workplace : Washington Univ. Job : CS researcher Research topics : Web University : UC California Interest : Sumo wrestling ... person Properties person person Event Participate Common Event-participation relationship Common property relationship Figure 2: Two kinds of relationships semination. A user can attach an access control list to his content using his social network when creating content on an application. Then, when the application receives a request to access the content, it determines whether to grant the request based on the access control list. Because users determine the access control to information based on the social network, the system requires social network data. The system obtains users' social networks automatically by mining various external sources such as Web, emails, and sensor information; subsequently, it maintains a database of the social network information. Users can adjust the network if necessary. The system enables users to analyze their social network to provide awareness of the information dissemination process within the social network. Using social relationships and the results of social network analyses, users can decide who can access their information. Currently, the proposed system is applied to an academic society because researchers have various social relationships (e.g., from a student to a professor, from a company to a university ) through their activities such as meetings, projects, and conferences. Importantly, they often need to share various information such as papers, ideas, reports, and schedules . Sometimes, such information includes private or confidential information that ought only to be shared with appropriate people. In addition, researchers have an interest in managing the information availability of their social relationships . The information of social relationships of an academic society, in particular computer science, is easily available online to a great degree. Such information is important to obtain social networks automatically. Hereafter, we explain in detail how social networks are modeled, extracted and analyzed. Then we explain how users can decide to control information access using social networks. 2.1 Representation of Social Relationships With the variety of social relationships that exist in the real world, a salient problem has surfaced: integration and consolidation on a semantic basis. The representation of social relationships must be sufficiently fine-grained that we can capture all details from individual sources of information in a way that these can be recombined later and taken as evidence of a certain relationship. Several representations of social relationships exist. For example, social network sites often simplify the relationship as "friend" or "acquaintance". In the Friend of a Friend (FOAF) [1] vocabulary, which is one of the Semantic Web's largest and most popular ontologies for describing people and whom they know, many kinds of relationships between people are deliberately simplified as "knows" relations. A rich ontological consideration of social relationships is needed for characterization and analysis of individual social networks . We define two kinds of social relationship (Fig. 2) [7]. The first basic structure of social relationship is a person's participation in an event. Social relationships come into existence through events involving two or more individuals. Such events might not require personal contact, but they must involve social interaction. From this event, social relationships begin a lifecycle of their own, during which the characteristics of the relationship might change through interaction or the lack thereof. An event is classified as perdu-rant in the DOLCE ontology [6], which is a popular ontology. For example, an event might be a meeting, a conference, a baseball game, a walk, etc. Assume that person A and person B participate in Event X. In that situation, we note that A and B share an event co-participation relationship under event X. A social relationship might have various social roles asso-ciated with it. For example, a student-professor relationship within a university setting includes an individual playing the role of a professor; another individual plays the role of a student . If A and B take the same role to Event X, they are in a same role relationship under event X (e.g., students at a class, colleagues in a workspace). If A cannot take over B's role or vice versa, A and B are in a role-sharing relationship (e.g., a professor and students, a project leader and staff). Another kind of social relationship is called a common property relationship. Sharing the same property value generates a common property relationship between people. For example, person A and person B have a common working place, common interests, and common experiences. Consequently , they are in a common property relationship with regard to those common properties. 2.2 Extraction of Social Networks If two persons are in either an event co-participation relationship or a common property relationship, they often communicate. The communication media can be diverse: 82 Figure 3: Editor for social relationships Figure 4: Editor for analyzing social networks and assigning an access control list to content face-to-face conversation, telephone call, email, chat, online communication on Weblogs, and so on. If we wish to discover the social relationship by observation, we must estimate relationships from superficial communication. The emerging field of social network mining provides methods for discovering social interactions and networks from legacy sources such as web pages, databases, mailing lists, and personal emails. Currently, we use three kinds of information sources to obtain social relationships using mining techniques. From the Web, we extract social networks using a search engine and the co-occurrence of two persons' names on the Web. Consequently , we can determine the following relationships among researchers: Coauthor, Same affiliation, Same project, Same event (participants of the same conference, workshop, etc.) [8]. Coauthor and Same event correspond to an event co-participation relationship. Same affiliation and same project correspond to a common property relationship. We are also using other sources such as email and sensors (we are developing a device that detects users within social spaces such as parties and conferences) to obtain social relationships. Necessarily, the quality of information obtained by mining is expected to be inferior to that of manually authored profiles . We can reuse those data if a user has already declared his relationships in FOAF or profiles of social networking services. Although users might find it difficult and demanding to record social relations, it would be beneficial to ask users to provide information to obtain social relationships. In addition to the relationship type, another factor of the social relationship is tie strength. Tie strength itself is a complex construct of several characteristics of social relations . It is definable as affective, frequency, trust, comple-mentarity , etc. No consensus for defining and measuring them exists, which means that people use different elicita-tion methods when it comes to determining tie strength. For example, Orkut, a huge social networking service, allows description of the strength of friendship relations on a five-point scale from "haven't met" to "best friend", whereas other sites might choose other scales or terms. In our system, we use trust as a parameter of tie strength. Trust has several very specific definitions. In [4], Golbeck describes trust as credibility or reliability in a human sense: "how much credence should I give to what this person speaks about" and "based on what my friends say, how much should I trust this new person?" In the context of information sharing , trust can be regarded as reliability regarding "how a person will handle my information". Users can give trust directly in a numerical value to a person in his relation. Alternatively, trust is obtainable automatically as authori-tativeness of each person using the social network [8]. The obtained social network data are integrated as extended FOAF files and stored in database. Users can adjust networks if needed (Fig. 3). The social relationship and its tie strength become guiding principles when a user determines an access control list to information. 2.3 Social Network Analysis for Information Sharing The system enables users to analyze their social networks to provide awareness of the information dissemination process within the social network. Social network analysis (SNA) is distinguishable from other fields of sociology by its focus on relationships between actors rather than attributes of actors, a network view, and a belief that structure affects substantive outcomes. Because an actor's position in a network affects information dissemination, SNA provides an important implication for information sharing on the social network. For example, occupying a favored position means that the actor will have better access to information, resources, and social support. The SNA models are based on graphs, with graph measures , such as centrality, that are defined using a sociological interpretation of graph structure. Freeman proposes numerous ways to measure centrality [2]. Considering a social network of actors, the simplest measure is to count the number of others with whom an actor maintains relations. The actor with the most connections, the highest degree, is most central. This measure is called degreeness. Another measure is closeness, which calculates the distance from each actor in the network to every other actor based on connections among all network members. Central actors are closer to all others than are other actors. A third measure is betweenness , which examines the extent to which an actor is situated among others in the network, the extent to which 83 Figure 5: Web site for sharing research information information must pass through them to get to others, and consequently, the extent to which they are exposed to information circulation within the network. If the betweenness of an actor is high, it frequently acts as a local bridge that connects the individual to other actors outside a group. In terms of network ties, this kind of bridge is well known as Granovetter's "weak tie" [5], which contrasts with "strong tie" within a densely-closed group. As the weak tie becomes a bridge between different groups, a large community often breaks up to a set of closely knit group of individuals, woven together more loosely according to occasional interaction among groups. Based on this theory, social network analysis offers a number of clustering algorithms for identifying communities based on network data. The system provides users with these network analyses (Fig. 4) so that they can decide who can access their information . For example, if user wants to diffuse her information , she might consider granting access to a person (with certain trust) who has both high degreeness and betweenness . On the other hand, she must be aware of betweenness when the information is private or confidential. Clustering is useful when a user wishes to share information within a certain group. APPLICATION To demonstrate and evaluate our system, we developed a community site (Fig. 5) using communication tools such as Weblogs, Wikis, and Forums. By that system, studies from different organizations and projects can be disseminated and their information thereby shared. Users can share various information such as papers, ideas, reports, and schedules at the site. Our system is integrated into a site that provides access control to that information. Integrating our system takes advantage of the open and information nature of the communication tools. It also maintains the privacy of the content and activities of those applications. Users can manage their social networks (Fig. 3) and attach the access control list to their content (e.g., Blog entries , profiles, and Wiki pages) using extracted social relationships and social network analysis (Fig. 4). Once a user determines the access control list, she can save it as her information access policy for corresponding content. The access policy is described using extended eXtensible Access Control Markup Language (XACML) and is stored in a database. She can reuse and modify the previous policy if she subsequently creates a similar content. One feature of our system is that it is easily adaptable to new applications because of its plug-and-play design. We are planning to integrate it into various Web sites and applications such as social network sites and RSS readers. RELATED WORKS AND CONCLUSIONS Goecks and Mynatt propose a Saori infrastructure that also uses social networks for information sharing [3]. They obtain social networks from users' email messages and provide sharing policies based on the type of information. We obtain social networks from various sources and integrate them into FOAF files. This facilitates the importation and maintenance of social network data. Another feature is that our system enables users to analyze their social networks. Thereby, users can control information dissemination more effectively and flexibly than through the use of pre-defined policies. As users increasingly disseminate their information on the Web, privacy concerns demand that access to particular information be limited. We propose a real-world oriented information sharing system using social networks. It enables users to control the information dissemination process within social networks, just as they are in the real world. Future studies will evaluate the system with regard to how it contributes to wider and safer information sharing than it would otherwise. We will also develop a distributed system that can be used fully on the current Web. REFERENCES [1] D. Brickley and L. Miller. FOAF: the 'friend of a friend' vocabulary. http://xmlns. com/foaf/0.1/, 2004. [2] L. C. Freeman. Centrality in social networks: Conceptual clarification, Social Networks, Vol.1, pp.215239, 1979. [3] J. Goecks and E. D. Mynatt. Leveraging Social Networks for Information Sharing In Proc. of CSCW'04, 2004. [4] J. Golbeck, J. Hendler, and B. Parsia. Trust networks on the semantic web, in Proc. WWW 2003, 2003. [5] M. Granovetter. Strength of weak ties, American Journal of Sociology, Vol.18, pp.13601380, 1973. [6] C. Masolo, S. Borgo, A. Gangemi, N. Guarinno, and A. Oltramari. WonderWeb Deliverable D18, http://wonderweb.semanticweb.org/deliverable/D18.shtml [7] Y. Matsuo, M. Hamasaki, J. Mori, H. Takeda and K. Hasida. Ontological Consideration on Human Relationship Vocabulary for FOAF. In Proc. of the 1st Workshop on Friend of a Friend, Social Networking and Semantic Web, 2004. [8] Y. Matsuo, H. Tomobe, K. Hasida, M. Ishiz uka. Finding Social Network for Trust Calculation. In Proc. of 16th European Conference on Artificial Intelligence, 2004. 84

Social network;Information sharing

167

Remote Access to Large Spatial Databases

Enterprises in the public and private sectors have been making their large spatial data archives available over the Internet . However, interactive work with such large volumes of online spatial data is a challenging task. We propose two efficient approaches to remote access to large spatial data. First, we introduce a client-server architecture where the work is distributed between the server and the individual clients for spatial query evaluation, data visualization, and data management. We enable the minimization of the requirements for system resources on the client side while maximizing system responsiveness as well as the number of connections one server can handle concurrently. Second, for prolonged periods of access to large online data, we introduce APPOINT (an Approach for Peer-to-Peer Offloading the INTernet). This is a centralized peer-to-peer approach that helps Internet users transfer large volumes of online data efficiently. In APPOINT, active clients of the client-server architecture act on the server's behalf and communicate with each other to decrease network latency, improve service bandwidth, and resolve server congestions.

INTRODUCTION In recent years, enterprises in the public and private sectors have provided access to large volumes of spatial data over the Internet. Interactive work with such large volumes of online spatial data is a challenging task. We have been developing an interactive browser for accessing spatial online databases: the SAND (Spatial and Non-spatial Data) Internet Browser. Users of this browser can interactively and visually manipulate spatial data remotely. Unfortunately, interactive remote access to spatial data slows to a crawl without proper data access mechanisms. We developed two separate methods for improving the system performance, together , form a dynamic network infrastructure that is highly scalable and provides a satisfactory user experience for interactions with large volumes of online spatial data. The core functionality responsible for the actual database operations is performed by the server-based SAND system. SAND is a spatial database system developed at the University of Maryland [12]. The client-side SAND Internet Browser provides a graphical user interface to the facilities of SAND over the Internet. Users specify queries by choosing the desired selection conditions from a variety of menus and dialog boxes. SAND Internet Browser is Java-based, which makes it deployable across many platforms. In addition, since Java has often been installed on target computers beforehand, our clients can be deployed on these systems with little or no need for any additional software installation or customiza-tion . The system can start being utilized immediately without any prior setup which can be extremely beneficial in time-sensitive usage scenarios such as emergencies. There are two ways to deploy SAND. First, any standard Web browser can be used to retrieve and run the client piece (SAND Internet Browser) as a Java application or an applet. This way, users across various platforms can continuously access large spatial data on a remote location with little or 1 5 no need for any preceding software installation. The second option is to use a stand-alone SAND Internet Browser along with a locally-installed Internet-enabled database management system (server piece). In this case, the SAND Internet Browser can still be utilized to view data from remote locations . However, frequently accessed data can be downloaded to the local database on demand, and subsequently accessed locally. Power users can also upload large volumes of spatial data back to the remote server using this enhanced client. We focused our efforts in two directions. We first aimed at developing a client-server architecture with efficient caching methods to balance local resources on one side and the significant latency of the network connection on the other. The low bandwidth of this connection is the primary concern in both cases. The outcome of this research primarily addresses the issues of our first type of usage (i.e., as a remote browser application or an applet) for our browser and other similar applications. The second direction aims at helping users that wish to manipulate large volumes of online data for prolonged periods. We have developed a centralized peer-to -peer approach to provide the users with the ability to transfer large volumes of data (i.e., whole data sets to the local database) more efficiently by better utilizing the distributed network resources among active clients of a client-server architecture. We call this architecture APPOINT -Approach for Peer-to-Peer Offloading the INTernet. The results of this research addresses primarily the issues of the second type of usage for our SAND Internet Browser (i.e., as a stand-alone application). The rest of this paper is organized as follows. Section 2 describes our client-server approach in more detail. Section 3 focuses on APPOINT, our peer-to-peer approach. Section 4 discusses our work in relation to existing work. Section 5 outlines a sample SAND Internet Browser scenario for both of our remote access approaches. Section 6 contains concluding remarks as well as future research directions. THE CLIENT-SERVER APPROACH Traditionally, Geographic Information Systems (GIS) such as ArcInfo from ESRI [2] and many spatial databases are designed to be stand-alone products. The spatial database is kept on the same computer or local area network from where it is visualized and queried. This architecture allows for instantaneous transfer of large amounts of data between the spatial database and the visualization module so that it is perfectly reasonable to use large-bandwidth protocols for communication between them. There are however many applications where a more distributed approach is desirable . In these cases, the database is maintained in one location while users need to work with it from possibly distant sites over the network (e.g., the Internet). These connections can be far slower and less reliable than local area networks and thus it is desirable to limit the data flow between the database (server) and the visualization unit (client) in order to get a timely response from the system. Our client-server approach (Figure 1) allows the actual database engine to be run in a central location maintained by spatial database experts, while end users acquire a Java-based client component that provides them with a gateway into the SAND spatial database engine. Our client is more than a simple image viewer. Instead, it operates on vector data allowing the client to execute many operations such as zooming or locational queries locally. In Figure 1: SAND Internet Browser -- Client-Server architecture. essence, a simple spatial database engine is run on the client. This database keeps a copy of a subset of the whole database whose full version is maintained on the server. This is a concept similar to `caching'. In our case, the client acts as a lightweight server in that given data, it evaluates queries and provides the visualization module with objects to be displayed. It initiates communication with the server only in cases where it does not have enough data stored locally. Since the locally run database is only updated when additional or newer data is needed, our architecture allows the system to minimize the network traffic between the client and the server when executing the most common user-side operations such as zooming and panning. In fact, as long as the user explores one region at a time (i.e., he or she is not panning all over the database), no additional data needs to be retrieved after the initial population of the client-side database. This makes the system much more responsive than the Web mapping services. Due to the complexity of evaluating arbitrary queries (i.e., more complex queries than window queries that are needed for database visualization), we do not perform user-specified queries on the client. All user queries are still evaluated on the server side and the results are downloaded onto the client for display. However, assuming that the queries are selective enough (i.e., there are far fewer elements returned from the query than the number of elements in the database), the response delay is usually within reasonable limits. 2.1 Client-Server Communication As mentioned above, the SAND Internet Browser is a client piece of the remotely accessible spatial database server built around the SAND kernel. In order to communicate with the server, whose application programming interface (API) is a Tcl-based scripting language, a servlet specifically designed to interface the SAND Internet Browser with the SAND kernel is required on the server side. This servlet listens on a given port of the server for incoming requests from the client. It translates these requests into the SAND-Tcl language. Next, it transmits these SAND-Tcl commands or scripts to the SAND kernel. After results are provided by the kernel, the servlet fetches and processes them, and then sends those results back to the originating client. Once the Java servlet is launched, it waits for a client to initiate a connection. It handles both requests for the actual client Java code (needed when the client is run as an applet) and the SAND traffic. When the client piece is launched, it connects back to the SAND servlet, the communication is driven by the client piece; the server only responds to the client's queries. The client initiates a transaction by 6 sending a query. The Java servlet parses the query and creates a corresponding SAND-Tcl expression or script in the SAND kernel's native format. It is then sent to the kernel for evaluation or execution. The kernel's response naturally depends on the query and can be a boolean value, a number or a string representing a value (e.g., a default color) or, a whole tuple (e.g., in response to a nearest tuple query). If a script was sent to the kernel (e.g., requesting all the tuples matching some criteria), then an arbitrary amount of data can be returned by the SAND server. In this case, the data is first compressed before it is sent over the network to the client. The data stream gets decompressed at the client before the results are parsed. Notice, that if another spatial database was to be used instead of the SAND kernel, then only a simple modification to the servlet would need to be made in order for the SAND Internet Browser to function properly. In particular , the queries sent by the client would need to be recoded into another query language which is native to this different spatial database. The format of the protocol used for communication between the servlet and the client is unaffected. THE PEER-TO-PEER APPROACH Many users may want to work on a complete spatial data set for a prolonged period of time. In this case, making an initial investment of downloading the whole data set may be needed to guarantee a satisfactory session. Unfortunately, spatial data tends to be large. A few download requests to a large data set from a set of idle clients waiting to be served can slow the server to a crawl. This is due to the fact that the common client-server approach to transferring data between the two ends of a connection assumes a designated role for each one of the ends (i.e, some clients and a server). We built APPOINT as a centralized peer-to-peer system to demonstrate our approach for improving the common client-server systems. A server still exists. There is a central source for the data and a decision mechanism for the service. The environment still functions as a client-server environment under many circumstances. Yet, unlike many common client-server environments, APPOINT maintains more information about the clients. This includes, inventories of what each client downloads, their availabilities, etc. When the client-server service starts to perform poorly or a request for a data item comes from a client with a poor connection to the server, APPOINT can start appointing appropriate active clients of the system to serve on behalf of the server, i.e., clients who have already volunteered their services and can take on the role of peers (hence, moving from a client-server scheme to a peer-to-peer scheme). The directory service for the active clients is still performed by the server but the server no longer serves all of the requests. In this scheme, clients are used mainly for the purpose of sharing their networking resources rather than introducing new content and hence they help offload the server and scale up the service. The existence of a server is simpler in terms of management of dynamic peers in comparison to pure peer-to -peer approaches where a flood of messages to discover who is still active in the system should be used by each peer that needs to make a decision. The server is also the main source of data and under regular circumstances it may not forward the service. Data is assumed to be formed of files. A single file forms the atomic means of communication. APPOINT optimizes requests with respect to these atomic requests. Frequently accessed data sets are replicated as a byproduct of having been requested by a large number of users. This opens up the potential for bypassing the server in future downloads for the data by other users as there are now many new points of access to it. Bypassing the server is useful when the server's bandwidth is limited. Existence of a server assures that unpopular data is also available at all times. The service depends on the availability of the server. The server is now more resilient to congestion as the service is more scalable. Backups and other maintenance activities are already being performed on the server and hence no extra administrative effort is needed for the dynamic peers. If a peer goes down, no extra precautions are taken. In fact, APPOINT does not require any additional resources from an already existing client-server environment but, instead, expands its capability. The peers simply get on to or get off from a table on the server. Uploading data is achieved in a similar manner as downloading data. For uploads, the active clients can again be utilized. Users can upload their data to a set of peers other than the server if the server is busy or resides in a distant location. Eventually the data is propagated to the server. All of the operations are performed in a transparent fashion to the clients. Upon initial connection to the server, they can be queried as to whether or not they want to share their idle networking time and disk space. The rest of the operations follow transparently after the initial contact. APPOINT works on the application layer but not on lower layers . This achieves platform independence and easy deploy-ment of the system. APPOINT is not a replacement but an addition to the current client-server architectures. We developed a library of function calls that when placed in a client-server architecture starts the service. We are developing advanced peer selection schemes that incorporate the location of active clients, bandwidth among active clients, data-size to be transferred, load on active clients, and availability of active clients to form a complete means of selecting the best clients that can become efficient alternatives to the server. With APPOINT we are defining a very simple API that could be used within an existing client-server system easily. Instead of denial of service or a slow connection, this API can be utilized to forward the service appropriately. The API for the server side is: start(serverPortNo) makeFileAvailable(file,location,boolean) callback receivedFile(file,location) callback errorReceivingFile(file,location,error) stop() Similarly the API for the client side is: start(clientPortNo,serverPortNo,serverAddress) makeFileAvailable(file,location,boolean) receiveFile(file,location) sendFile(file,location) stop() The server, after starting the APPOINT service, can make all of the data files available to the clients by using the makeFileAvailable method. This will enable APPOINT to treat the server as one of the peers. The two callback methods of the server are invoked when a file is received from a client, or when an error is encountered while receiving a file from a client. APPOINT guar-7 Figure 2: The localization operation in APPOINT. antees that at least one of the callbacks will be called so that the user (who may not be online anymore) can always be notified (i.e., via email). Clients localizing large data files can make these files available to the public by using the makeFileAvailable method on the client side. For example, in our SAND Internet Browser, we have the localization of spatial data as a function that can be chosen from our menus. This functionality enables users to download data sets completely to their local disks before starting their queries or analysis. In our implementation, we have calls to the APPOINT service both on the client and the server sides as mentioned above. Hence, when a localization request comes to the SAND Internet Browser, the browser leaves the decisions to optimally find and localize a data set to the APPOINT service. Our server also makes its data files available over APPOINT. The mechanism for the localization operation is shown with more details from the APPOINT protocols in Figure 2. The upload operation is performed in a similar fashion. RELATED WORK There has been a substantial amount of research on remote access to spatial data. One specific approach has been adopted by numerous Web-based mapping services (MapQuest [5], MapsOnUs [6], etc.). The goal in this approach is to enable remote users, typically only equipped with standard Web browsers, to access the company's spatial database server and retrieve information in the form of pictorial maps from them. The solution presented by most of these vendors is based on performing all the calculations on the server side and transferring only bitmaps that represent results of user queries and commands. Although the advantage of this solution is the minimization of both hardware and software resources on the client site, the resulting product has severe limitations in terms of available functionality and response time (each user action results in a new bitmap being transferred to the client). Work described in [9] examines a client-server architecture for viewing large images that operates over a low-bandwidth network connection. It presents a technique based on wavelet transformations that allows the minimization of the amount of data needed to be transferred over the network between the server and the client. In this case, while the server holds the full representation of the large image , only a limited amount of data needs to be transferred to the client to enable it to display a currently requested view into the image. On the client side, the image is reconstructed into a pyramid representation to speed up zooming and panning operations. Both the client and the server keep a common mask that indicates what parts of the image are available on the client and what needs to be requested. This also allows dropping unnecessary parts of the image from the main memory on the server. Other related work has been reported in [16] where a client-server architecture is described that is designed to provide end users with access to a server. It is assumed that this data server manages vast databases that are impractical to be stored on individual clients. This work blends raster data management (stored in pyramids [22]) with vector data stored in quadtrees [19, 20]. For our peer-to-peer transfer approach (APPOINT), Nap-ster is the forefather where a directory service is centralized on a server and users exchange music files that they have stored on their local disks. Our application domain, where the data is already freely available to the public, forms a prime candidate for such a peer-to-peer approach. Gnutella is a pure (decentralized) peer-to-peer file exchange system. Unfortunately, it suffers from scalability issues, i.e., floods of messages between peers in order to map connectivity in the system are required. Other systems followed these popular systems, each addressing a different flavor of sharing over the Internet. Many peer-to-peer storage systems have also recently emerged. PAST [18], Eternity Service [7], CFS [10], and OceanStore [15] are some peer-to-peer storage systems. Some of these systems have focused on anonymity while others have focused on persistence of storage. Also, other approaches , like SETI@Home [21], made other resources, such as idle CPUs, work together over the Internet to solve large scale computational problems. Our goal is different than these approaches. With APPOINT, we want to improve existing client-server systems in terms of performance by using idle networking resources among active clients. Hence, other issues like anonymity, decentralization, and persistence of storage were less important in our decisions. Confirming the authenticity of the indirectly delivered data sets is not yet addressed with APPOINT. We want to expand our research , in the future, to address this issue. From our perspective, although APPOINT employs some of the techniques used in peer-to-peer systems, it is also closely related to current Web caching architectures. Squirrel [13] forms the middle ground. It creates a pure peer-to-peer collaborative Web cache among the Web browser caches of the machines in a local-area network. Except for this recent peer-to-peer approach, Web caching is mostly a well-studied topic in the realm of server/proxy level caching [8, 11, 14, 17]. Collaborative Web caching systems, the most relevant of these for our research, focus on creating either a hierarchical, hash-based, central directory-based, or multicast-based caching schemes. We do not compete with these approaches. In fact, APPOINT can work in tandem with collaborative Web caching if they are deployed together. We try to address the situation where a request arrives at a server, meaning all the caches report a miss. Hence, the point where the server is reached can be used to take a central decision but then the actual service request can be forwarded to a set of active clients, i.e., the down-8 load and upload operations. Cache misses are especially common in the type of large data-based services on which we are working. Most of the Web caching schemes that are in use today employ a replacement policy that gives a priority to replacing the largest sized items over smaller-sized ones. Hence, these policies would lead to the immediate replacement of our relatively large data files even though they may be used frequently. In addition, in our case, the user community that accesses a certain data file may also be very dispersed from a network point of view and thus cannot take advantage of any of the caching schemes. Finally, none of the Web caching methods address the symmetric issue of large data uploads. A SAMPLE APPLICATION FedStats [1] is an online source that enables ordinary citizens access to official statistics of numerous federal agencies without knowing in advance which agency produced them. We are using a FedStats data set as a testbed for our work. Our goal is to provide more power to the users of FedStats by utilizing the SAND Internet Browser. As an example, we looked at two data files corresponding to Environmen-tal Protection Agency (EPA)-regulated facilities that have chlorine and arsenic, respectively. For each file, we had the following information available: EPA-ID, name, street, city, state, zip code, latitude, longitude, followed by flags to indicate if that facility is in the following EPA programs: Hazardous Waste, Wastewater Discharge, Air Emissions, Abandoned Toxic Waste Dump, and Active Toxic Release. We put this data into a SAND relation where the spatial attribute `location' corresponds to the latitude and longitude . Some queries that can be handled with our system on this data include: 1. Find all EPA-regulated facilities that have arsenic and participate in the Air Emissions program, and: (a) Lie in Georgia to Illinois, alphabetically. (b) Lie within Arkansas or 30 miles within its border. (c) Lie within 30 miles of the border of Arkansas (i.e., both sides of the border). 2. For each EPA-regulated facility that has arsenic, find all EPA-regulated facilities that have chlorine and: (a) That are closer to it than to any other EPA-regulated facility that has arsenic. (b) That participate in the Air Emissions program and are closer to it than to any other EPA-regulated facility which has arsenic. In order to avoid reporting a particular facility more than once, we use our `group by EPA-ID' mechanism. Figure 3 illustrates the output of an example query that finds all arsenic sites within a given distance of the border of Arkansas. The sites are obtained in an incremental manner with respect to a given point. This ordering is shown by using different color shades. With this example data, it is possible to work with the SAND Internet Browser online as an applet (connecting to a remote server) or after localizing the data and then opening it locally. In the first case, for each action taken, the client-server architecture will decide what to ask for from the server. In the latter case, the browser will use the peer-to -peer APPOINT architecture for first localizing the data. CONCLUDING REMARKS An overview of our efforts in providing remote access to large spatial data has been given. We have outlined our approaches and introduced their individual elements. Our client-server approach improves the system performance by using efficient caching methods when a remote server is accessed from thin-clients. APPOINT forms an alternative approach that improves performance under an existing client-server system by using idle client resources when individual users want work on a data set for longer periods of time using their client computers. For the future, we envision development of new efficient algorithms that will support large online data transfers within our peer-to-peer approach using multiple peers simultane-ously . We assume that a peer (client) can become unavail-able at any anytime and hence provisions need to be in place to handle such a situation. To address this, we will augment our methods to include efficient dynamic updates. Upon completion of this step of our work, we also plan to run comprehensive performance studies on our methods. Another issue is how to access data from different sources in different formats. In order to access multiple data sources in real time, it is desirable to look for a mechanism that would support data exchange by design. The XML protocol [3] has emerged to become virtually a standard for describing and communicating arbitrary data. GML [4] is an XML variant that is becoming increasingly popular for exchange of geographical data. We are currently working on making SAND XML-compatible so that the user can instantly retrieve spatial data provided by various agencies in the GML format via their Web services and then explore, query, or process this data further within the SAND framework . This will turn the SAND system into a universal tool for accessing any spatial data set as it will be deployable on most platforms, work efficiently given large amounts of data, be able to tap any GML-enabled data source, and provide an easy to use graphical user interface. This will also convert the SAND system from a research-oriented prototype into a product that could be used by end users for accessing , viewing, and analyzing their data efficiently and with minimum effort. REFERENCES [1] Fedstats: The gateway to statistics from over 100 U.S. federal agencies. http://www.fedstats.gov/, 2001. [2] Arcinfo: Scalable system of software for geographic data creation, management, integration, analysis, and dissemination. http://www.esri.com/software/ arcgis/arcinfo/index.html, 2002. [3] Extensible markup language (xml). http://www.w3.org/XML/, 2002. [4] Geography markup language (gml) 2.0. http://opengis.net/gml/01-029/GML2.html, 2002. [5] Mapquest: Consumer-focused interactive mapping site on the web. http://www.mapquest.com, 2002. [6] Mapsonus: Suite of online geographic services. http://www.mapsonus.com, 2002. [7] R. Anderson. The Eternity Service. In Proceedings of the PRAGOCRYPT'96, pages 242252, Prague, Czech Republic, September 1996. [8] L. Breslau, P. Cao, L. Fan, G. Phillips, and S. Shenker. Web caching and Zipf-like distributions: 9 Figure 3: Sample output from the SAND Internet Browser -- Large dark dots indicate the result of a query that looks for all arsenic sites within a given distance from Arkansas. Different color shades are used to indicate ranking order by the distance from a given point. Evidence and implications. In Proceedings of the IEEE Infocom'99, pages 126134, New York, NY, March 1999. [9] E. Chang, C. Yap, and T. Yen. Realtime visualization of large images over a thinwire. In R. Yagel and H. Hagen, editors, Proceedings IEEE Visualization'97 (Late Breaking Hot Topics), pages 4548, Phoenix, AZ, October 1997. [10] F. Dabek, M. F. Kaashoek, D. Karger, R. Morris, and I. Stoica. Wide-area cooperative storage with CFS. In Proceedings of the ACM SOSP'01, pages 202215, Banff, AL, October 2001. [11] A. Dingle and T. Partl. Web cache coherence. Computer Networks and ISDN Systems, 28(7-11):907920, May 1996. [12] C. Esperanca and H. Samet. Experience with SAND/Tcl: a scripting tool for spatial databases. Journal of Visual Languages and Computing, 13(2):229255, April 2002. [13] S. Iyer, A. Rowstron, and P. Druschel. Squirrel: A decentralized peer-to-peer Web cache. Rice University/Microsoft Research, submitted for publication, 2002. [14] D. Karger, A. Sherman, A. Berkheimer, B. Bogstad, R. Dhanidina, K. Iwamoto, B. Kim, L. Matkins, and Y. Yerushalmi. Web caching with consistent hashing. Computer Networks, 31(11-16):12031213, May 1999. [15] J. Kubiatowicz, D. Bindel, Y. Chen, S. Czerwinski, P. Eaton, D. Geels, R. Gummadi, S. Rhea, H. Weatherspoon, W. Weimer, C. Wells, and B. Zhao. OceanStore: An architecture for global-scale persistent store. In Proceedings of the ACM ASPLOS'00, pages 190201, Cambridge, MA, November 2000. [16] M. Potmesil. Maps alive: viewing geospatial information on the WWW. Computer Networks and ISDN Systems, 29(813):13271342, September 1997. Also Hyper Proceedings of the 6th International World Wide Web Conference, Santa Clara, CA, April 1997. [17] M. Rabinovich, J. Chase, and S. Gadde. Not all hits are created equal: Cooperative proxy caching over a wide-area network. Computer Networks and ISDN Systems, 30(22-23):22532259, November 1998. [18] A. Rowstron and P. Druschel. Storage management and caching in PAST, a large-scale, persistent peer-to-peer storage utility. In Proceedings of the ACM SOSP'01, pages 160173, Banff, AL, October 2001. [19] H. Samet. Applications of Spatial Data Structures: Computer Graphics, Image Processing, and GIS. Addison-Wesley, Reading, MA, 1990. [20] H. Samet. The Design and Analysis of Spatial Data Structures. Addison-Wesley, Reading, MA, 1990. [21] SETI@Home. http://setiathome.ssl.berkeley.edu/, 2001. [22] L. J. Williams. Pyramidal parametrics. Computer Graphics, 17(3):111, July 1983. Also Proceedings of the SIGGRAPH'83 Conference, Detroit, July 1983. 10

GIS;Client/server;Peer-to-peer;Internet

168

ResearchExplorer: Gaining Insights through Exploration in Multimedia Scientific Data

An increasing amount of heterogeneous information about scientific research is becoming available on-line. This potentially allows users to explore the information from multiple perspectives and derive insights and not just raw data about a topic of interest. However, most current scientific information search systems lag behind this trend; being text-based, they are fundamentally incapable of dealing with multimedia data. An even more important limitation is that their information environments are information-centric and therefore are not suitable if insights are desired. Towards this goal, in this paper, we describe the design of a system, called ResearchExplorer, which facilitates exploring multimedia scientific data to gain insights. This is accomplished by providing an interaction environment for insights where users can explore multimedia scientific information sources. The multimedia information is united around the notion of research event and can be accessed in a unified way. Experiments are conducted to show how ResearchExplorer works and how it cardinally differs from other search systems.

INTRODUCTION Current web search engines and bibliography systems are information-centric. Before searching for information, users need to construct a query typically, by using some keywords to represent the information they want. After the query is issued, the system retrieves all information relevant to the query. The results from such queries are usually presented to users by listing all relevant hits. Thus, with these information-centric systems, users can find information such as a person's homepage, a paper, a research project's web page, and so on. However, when users want to know the following types of things, they are unable to find answers easily with current search systems: 1) Evolution of a field 2) People working in the field 3) A person's contribution to the field 4) Classical papers (or readings) in the field 5) Conferences/journals in the field 6) How the research of a person or an organization (group, dept, university, etc) has evolved. The reasons why current information-centric search systems have difficulty to help users to find answers to questions above are due to the limitations of their information environments. First, some issues result from their data modeling. For example, to answer the question of "evolution of a field", the most important information components, which are time and location, need to be captured and appropriately presented or utilized. However, in typical bibliography systems such information is rigidly utilized (if at all available) in the time-stamping sense. Second, many important issues arise due to the presentation methods utilized by such systems. For example, even though users can find all papers of a person with some systems, it is not easy for users to observe the trend if the results are just listed sequentially. As an alternative, presenting results in a visual form can make trend easier to identify. Third, some of the questions listed above can not be answered directly by the system because the answers depend on individual person. For example, different users will have different judgments on a researcher's contribution to a field. To form their own thoughts, users may need to investigate and compare several factors many times. In this case, it is too tedious if each query is a new query. Thus, it is necessary that the system can maintain query and user states and allow users to refine queries dynamically. In other words, the user can not only query but also explore information. For this study, we propose a bibliography system with novel interaction environment that aids not just in syntactic query retrieval but also aids in developing insights. The goal of this Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MIR'04, October 1516, 2004, New York, New York, USA. Copyright 2004 ACM 1-58113-940-3/04/0010...$5.00. 7 system is to provide users an interaction environment where information is modeled, accessed, and presented in such a way that users can gain insights easily through exploration. Specifically, in the interaction environment, scientific information is modeled around the notion of a research event, which brings together all semantically related information regardless of the media (text, image, or video), through which it is expressed. Thus, when users explore the information space, they can view research in multiple media formats. Further, the interaction environment presents information using multidimensional views, which include temporal and spatial views. At the same time, the interaction environment shows information of other attributes of research, like category and people information. In summary, the contribution of this work is to propose a novel interaction environment for insights. Although the system is focused on scientific information, we believe the techniques developed in this work are applicable to other applications and can work as a framework guiding design of interaction environments for insights. The paper is structured as follows. We begin with an introduction of interaction environment for insights. Section 3 describes the system architecture. Section 4 explains data modeling of the interaction environment. Section 5 presents how the interaction environment is implemented. Section 6 discusses experiments and results. Section 7 gives a review of related work. Section 8 concludes. INTERACTION ENVIRONMENT FOR INSIGHTS Our goal in designing the system is to provide an interaction environment for users to explore multimedia scientific data to gain insights into research. Insight is commonly understood as follows. Insight: the clear (and often sudden) understanding of a complex situation [21]. From the definition, we can see insight is different from information. If insight is gained, people should be able to understand the inner nature of things. To illustrate their difference, we refer the reader to Figure 1. In the figure, left part shows two columns of numbers. What these numbers convey to people is just information. It is very difficult for people to understand the relationship between numbers in these two columns by looking at numbers only. But if we show these numbers by a chart as in the right hand, people can easily tell and understand that the two columns have linear relationship. That is the insight. In this case, people gain insight by understanding relationship, which is visualized by a certain technique. In the context of research, insights should include clear understanding of different situations. Examples of these situations are a research field, a person, an organization, and a specific research event which will be defined later. 2.2 Key Characteristics of Interaction Environment for Insights An interaction environment for insights is an environment that helps users to gain insights through exploration. It consists of a database to store data, and user interface to explore data. Such an environment has the following key characteristics. 1) Database to store information. As described in section 1, spatio-temporal characteristics of information are critical to present the evolution of a situation. Clear understanding a situation often requires understanding how the situation evolves. Therefore, spatio-temporal aspects of information should be captured in data modeling. In addition, the data modeling should be able to unify multimedia information. Multimedia enables users to observe things using different senses. Some media can help people to understand quickly. Figure 2 shows such an example. By looking at the text at the top, people may not be able to understand what the paper "Content Based Image Synthesis" talks about. But with the help of the images below, people can get an idea quickly what the paper is about. Further, multimedia provides users opportunity to view things from different perspectives. This is especially important when clear understanding a situation requires examine the situation from multiple angles. 2) User interface. As people gain insights by exploration, they may need to check into a situation repeatedly and from different viewpoints. Thus, interaction between a user and the environment becomes very important. The design of user interface should take this into consideration. We believe the key features of UI are as follows. a. The UI should support exploration of the spatial and temporal characteristics of information. b. The UI should support direct interactions between the users and the information. This requires the UI to have two characteristics: First, the UI should have the same query and presentation space. In other words, a window in the UI can not only be used to show information but also be used to specify queries. For example, time and location windows can show temporal and spatial information. At the same time, users can issue temporal and spatial queries in time and location windows. To Figure 1. Information vs. insight 51 255 153 459 408 204 102 306 357 1 5 3 9 8 4 2 6 7 0 50 100 150 200 250 300 350 400 450 500 1 2 3 4 5 6 7 8 9 8 specify a query, the operation should be simple and direct. The other characteristic is the reflective nature of the UI. This means once information in a window is changed, all other windows will be updated automatically. This helps users to interact with the environment directly and effectively. c. The UI enables users to issue dynamic query. In some current interaction environments, users are constrained in forming queries. For example, users can only generate temporal query with one time interval. In an interaction environment for insights, users should be able to form a query with multiple choices. This provides users more flexibility to look into a situation of interest. d. The UI maintains the query state. It should know which query whose results are used in the search condition of another query and which query is based on another query's results. This helps users not only to be aware to contexts but also to form complex queries. e. The UI should have zoom-in/zoom-out functionality that allows examining the information at different resolutions. When large volume of data is retrieved, there is readability issue. To address this issue, zoom-in/zoom -out functionality is needed. f. Different visualization techniques need to be used. As shown in figure 1, visualization techniques help users to understand relationships and gain insights. However, different relationships need different visualization techniques. For example, social relationships are of network structure. Temporal relationships are two dimensional. To visualize these two types of relationships effectively, it requires different techniques. These characteristics will guide the interaction environment design of the system we are discussing. SYSTEM ARCHITECTURE Figure 3 shows the high level architecture of ResearchExplorer. There are three main components: Event Collector, Event Database and Interaction Environment. One of the functions of Event Collector is to gather data from different sources. Then it parses and assimilates information around the notion of research event. Finally, it sends these data to Event Database. Event Database is a database of events. It stores all information around events. ResearchExplorer uses a natural XML database for Event Database. The reasons will be explained in next section. In this database, all information about a research event is stored as an XML file. The schema will be defined in section 4.2. Interaction Environment consists of User Interface and Searcher. Through the UI, users form a query. The query is then converted into XPath format by the Searcher and sent to the Event Database. After the results are retrieved from the Event Database, the Searcher gets them back to the User Interface to present to users. In this paper, our focus is not on how to collect data or unify multimedia information by event. Interested readers can refer to [5][6][14][15][18] for information gathering and [17] for multimedia information assimilation. What we focus is on the design of Interaction Environment based on research event. EVENT DATABASE As described above, the interaction environment for insights needs data modeling which can capture temporal and spatial characteristics, and unify multimedia. A recent work in [17] has proposed a unified multimedia data model that is capable of describing spatial-temporal characteristics of information. This model is built on the notion of events. Its efficacy has been demonstrated in different domains including modeling of multimedia information related to meetings [17] and personal Research Events Event Collector Searcher User Interface Data Event Database Figure 3. ResearchExplorer architecture. query Interaction Environment Figure 2. Multimedia helps understanding. Paper Title: Content Based Image Synthesis 9 multimedia information management [16]. We base our current research on these ideas and extend them further to the specific domain of scientific information management. In [17], an event is defined as follows. Event: An event is an observed physical reality parameterized by space and the time. The observations describing the event are defined by the nature or physics of the observable, the observation model, and the observer. This definition was given to events in general. In order to be concrete in research domain, a specific event definition to research is necessary. Therefore, based on their definition, we are defining an event in research domain, which is called research event, as follows. Research event: A research event is a set of semantically correlated events within research domain, parameterized by time, location, participant, and content. Note that semantics is contextual. It depends on many factors like time, location, people etc. Thus, a research event is flexible. For instance, it can be a research paper, a thesis, a technical report, a book, a patent, a presentation, an image, a video, or a project combining part or all of aforementioned. Semantics also depends on domain level. It is generated differently at different domain levels even though from the same event. That is because different aspects of the event are emphasized at different domain levels. Thus, a research event could be part of another one. For example, someone is talking is an event by itself. At the same time, it is part of a seminar event as well. The definition of a research event provides us with the central characteristics to meet the requirements of our application. By the definition, a research event is parameterized by time and location. It can capture the dynamics of itself. Thus, users can easily observe how events evolve, which is helpful to insight generation. Relationships between events can be shown in terms of attributes of an event. This will enable users to observe events in a big context and get deeper understanding. Further, all multimedia data is unified around the notion of a research event. Thus, a research event becomes an access point to multimedia data. 4.2 Semi-Structured Data Multimedia data about scientific research does not follow a rigid structure. For example, research papers have reference while images do not have. Even for references, the number of citations varied over different papers. At the same time, these data do have some common information components such as time and location information. This semi-structured characteristic makes methods like relational database for storing structured data unsuitable. Techniques for storing semi-structured data are appropriate instead. XML is one of the solutions to model semi-structured data. It has become very popular for introducing semantics in text. And it has rapidly replaced automatic approaches to deduce semantics from the data in text files. This approach to explicitly introduce tags to help processes compute semantics has been very successful so far [13]. Based on this, we choose XML to store research event information. Figure 4 shows the schema of XML files for research events. 4.3 Description of the Data Model Based on the definition of research event, four fundamental information components are needed to describe a research event. These components are: when the research event happens, where the research event occurs, who participates in the research event, and what the research event is about. Thus, a data model as follows is proposed to represent a research event. As shown in Figure 5, a research event is characterized by the following attributes: Name, Time, Participant, Category, Mediasource, Subevents, and Free Attributes. Here Name refers to the name of a research event, Time refers to the times when the research is done, Participant refers to people who do the research and their affiliations, Category refers to the ACM Classification of the research, which can belong to several categories, Mediasource contains media type and source (URL) of the media covering the research event, Subevent refers to part of the research event and has the same structure of a research event, Free Attributes are used to capture media specific characteristics when needed, for example, it refers to reference for a paper. As described above, the data model encapsulates all information components of a research event by one or more attributes. When component is captured by Time, where is captured by Participant Affiliation, who is captured by Participant, and what is captured by Name and Categories. Multimedia supporting the research event is brought in by Mediasource attribute. 4.4 XML Database In ResearchExplorer, Berkeley DB XML [3] is chosen for Event Database. Berkeley DB XML is an application-specific native XML data manager. It is supplied as a library that links directly into the application's address space. Berkeley DB XML provides storage and retrieval for native XML data and semi-structured data. So it can meet the requirements of Event Database. Figure 4. Schema of research event. 10 USER INTERFACE In ResearchExplorer, a unified presentation-browsing-querying interface is used. Research events are shown in multidimensional views. As multimedia data is organized around research event, the data is presented by fundamental components of research event, i.e., When, Where, Who, and What. Figure 7(a) shows a screenshot of the user interface we developed. There are totally five windows plus a text box. In the upper are timeline and map windows showing time and location information of research events. In the lower right, there are two windows showing people and category information. The window in the lower left is different from those windows aforementioned. It is used to show multimedia data of research events. Once a research event is selected, multimedia data like papers, images, and videos are presented in this window and they are presented according to the event-subevent structure. Clicking on a specific media instance label will lead users to the original source of the media and trigger appropriate application for that particular kind of media. So users can view original media as they want. The text box is designed for keyword-based searching. It enables users to search information in traditional way. 5.2 Research Representation Time and location are the primary parameters based on which dynamics is captured. Therefore, they are depicted as the primary exploration dimensions. The way to represent research events in these windows is critical. In ResearchExplorer, two different representation methods are used. We borrowed idea of representing research events from [16] in the timeline window where research events are represented by rectangles. A rectangle spans the duration of a research event. Within each rectangle, there may be smaller rectangles. These smaller ones represent subevents of the research event. All rectangles for one research event are nested according to the event-subevent structure. The media, presenting the research event, are represented by icons in the rectangles. Icons are chosen intuitively for users to recognize easily. They are specific to each media. Icons belong to the same research event are grouped together in chronological order. The fidelity of such a representation is maintained during temporal zoom-in/zoom-out operations as described later. The recursive nature of the representation is used to capture aggregate relationships where a research event may comprise of other events. The primary purpose of such a representation is to provide users with a structural and temporal view of research events. In the map window, research events are represented by "dot maps" [19]. Each dot in the map shows a research event at the location of the dot. By means of dot maps, the precision of location information is high, and the variable density of dots conveys information about the amount of research events at a location. 5.3 What-You-See-Is-What-You-Get (WYSIWYG) In the system, WYSIWYG search is employed the query and presentation spaces are the same. As described above, windows serve as a way to display information and relationships of research events. These windows, except details window, serve another function in specifying queries as well. Contrary to many search interfaces where users specify several properties and then press a button to issue a query, users can issue a query by a simple operation in this user interface. For example, users can launch a query by specifying a time interval, a location region, a person's name, or a research category. Figure 6 shows examples of these methods. In ResearchExplorer, exploration is based on sessions. Each session consists of one or more queries. A query is either a new session query or a refine query. A new session query is the first query of each session. All other queries in a session are refine queries. For new session query, the system retrieves results from the database. If it is a refine query, the query will not be sent to the database. It will be executed based on the results set of the new session query of that session. With this method, users can choose a broad set of results first, and then observe any subset of the results of interest. This is very important because knowledge is accumulated as users manipulate the results by choosing different perspectives. Once a refine query is posed, results of the query will be highlighted in all windows. Figure 5. Graphical representation of a research event. RE: Research Event N: Name T: Time PS: Participants P: Participant PN: Participant's Name FN: First Name MS: Media Source MT: Media Type S: Source SS: Subevents SE: Subevent FA: Free Attribute N T RE C SS PS AN LA LO FN PN PA LN P CS FA ... ... MT S LN: Last Name PA: Participant Affiliation AN: Affiliation Name LA: Latitude LO: Longitude CS: Categories C: Category SE ... MS Figure 6. Different query methods. (a) shows a query by time. (b) shows a query by a person. (c) shows query by specifying a region of locations. (d) shows query by category. 11 5.4 Reflective UI In designing the user interface, multiple window coordination strategy is used. By means of this strategy, components of the user interface are coupled tightly. The windows respond to user activity in a unified manner such that user interaction in one window is reflected instantly in other windows. For example, when the user selects a research event in timeline window, this research event will be highlighted in the map and other windows. This cooperative visualization is effective in information exploration as it maintains the context as the user interacts with the data. Figure 7(a) shows an example where a research event is selected and its information in other windows is highlighted. 5.5 Interactions with Time and Location Information In ResearchExplorer, both timeline and map provide zoom-in/zoom -out function. This makes it possible for users to look at how a research event evolves in details. The timeline has year as the highest level of temporal representation. So it's likely that two subevents of a research event are overlapped when they are shown in year level. With zoom-in functionality, users can zoom into finer level to see the temporal relationship between these two subevents. The location window in ResearchExplorer has been implemented using an open source JavaBeansTM based package called OpenMapTM [2]. Research events are presented as dots on the map. Due to the size limitation, it is hard for users to differentiate events when they are close to each other. Similarly as in timeline, users can zoom into that area and see the events at finer resolution. Further, panning of the entire map is also supported. EXPERIMENTS In this section, we conduct some experiments as case studies. These case studies will show how users can look into details of research events and observe relationships between research events. 6.1 Experiment I: Exploring Information Exploring information with context is one of important features of ResearchExplorer. With this function, users can refine retrieved results to check into different aspects of a situation. In this experiment, we are interested in how users can refine results as they explore information. Assume following information is of interest: Show all research events on AI during the time period from 1989 through 2004? Out of the results above, show the part done in CA For each person, show all research events he/she participated in. To find answers to the first query, users can select the time interval from 1989 to 2004 and then choose AI category only. Figure 7(a) shows all results. Note that the results consist of all research events in AI, but they do not include all in the world. As shown in the figure, timeline window shows the temporal information and temporal relationships between research events, map window shows the distribution of location, category window shows what all categories these research events belong to, participant window shows all people being involved in these research events. The details window shows all multimedia data about research events. In this window, a research event named "UMN MegaScout" is placed at the top after the rectangle representing this event is clicked in the timeline window. If we click the image thumbnails, the original images will be shown. Figures 7(b) shows the three images and four videos. When we look at these images and video frames, we can have a better understanding about this research. In other words, direct observation of the multimedia data of a research event helps users to gain insights about the research. In order to answer the second query, we specify a region on the map which encloses all dots on CA. The research events are then highlighted in timeline windows as shown in Figure 7(c). To check research events a person participated in, we only need to move the mouse cursor to be over the person's name. Similarly, all research events he/she participated in will be highlighted. 6.2 Experiment II: Comparisons with Other Systems In this section, we compare ResearchExplorer with other systems in terms of functionalities. Without loss of generality, we choose Google [9], CiteSeer [7], and ACM Digital Library [1] as examples. First, we compare the presentation methods. Figure 8 shows the screen shots of these systems after a query of "artificial intelligence" is issued. Compared with ResearchExplorer, these systems are unable to show how AI research evolves. Users thus can not get a whole picture of AI research, which otherwise is important to understand this area and conduct research in AI. Another comparison is done on query and explore functions. The results are shown in Table 1. RELATED WORK There are many systems which can search for scientific information. These systems can be classified into two categories. The first are bibliographical systems developed especially for scientific information searching. ACM Digital Library, IEEE Xplore [10], and INSPEC [11] are good examples of this class. These systems store information about publications, which are from some pre-selected sources, in their repositories. They organize data by using structural information of publications like title, author, etc. CiteSeer is another well-known system of this kind. Compared with the aforementioned, it collects publications from the web and performs citation indexing in addition to full-text indexing [8]. Another class is web search engine for general information. The most well-known of this type is Google. Systems of this class index the text contained in documents, allowing users to find information using keyword search. Our work differs significantly from these systems. These systems are designed to concentrating in information providing. Our work is focused on providing an interaction environment for insights into research. Other related work comes from research on multimedia experience. Boll and Westermann [4] presented the Medither multimedia event space, a decentralized peer-to-peer infrastructure that allows to publish, to find and to be notified about multimedia events of interest. Our focus was not to create a multimedia event space but rather to develop an interaction environment for users to experience multimedia events. The Informedia group at CMU has also worked on multimedia 12 experience [20]. However there are important differences here their main goal was to capture and integrate personal multimedia Image1 Image3 Image2 Video4 Video1 Video2 Video3 (a) (c) (b) Figure 7. ResearchExplorer UI. (a) shows screen shot of the UI. (b) shows the images and videos of a research event. (c) shows the highlighted results when a spatial refinement is made. Figure 8. Screen shots of other search systems. 13 experiences not create an environment for experiencing multimedia personally. In [12], Jain envisioned the essence of an experiential environment. The main goal of experiential environment is for insights. Following this work, there are some other work on experiential environment [13][16]. However, in these work, there is little discussion on experiential environments. Our work developed further some ideas in [12] and concretized the design framework of an interaction environment for insights. CONCLUSION We have described a novel system which helps users to gain insights through exploring multimedia scientific data. Although framework for designing an interaction environment for insights is identified, the implementation is a first step towards a mature system for insights. In future work, we will build on the methods described here. Also, we will investigate more on relationships between research events and methodologies to present these relationships. We believe some of the more interesting research problems will be identified when new relationships between research events are used to help users to gain insights. ACKNOWLEDGMENTS We would thank Punit Gupta and Rachel L. Knickmeyer for their help on timeline and map components of ResearchExplorer. REFERENCES [1] ACM Digital Library, http://portal.acm.org/portal.cfm. [2] BBN Technologies (1999), OpenMapTM Open Systems Mapping Technoloy, http://openmap.bbn.com/. [3] Berkeley DB XML, http://www.sleepycat.com/products/xml.shtml. [4] Boll, S., and Westermann, U. Medither -- an Event Space for Context-Aware Multimedia Experiences. Proc. of the 2003 ACM SIGMM Workshop on Experiential Telepresence (ETP '03), 21-30. [5] Brin, S., and Page, L. The Anatomy of a Large-Scale Hypertextual Web Search Engine. Proc. of 7 th International World Wide Web Conference (WWW '98), 107-117. [6] Cho, J., Garcia-Monila, H., and Page, L. Efficient Crawling through URL Ordering. Proc. of 7 th WWW Conference (1998), 161-172. [7] CiteSeer, http://citeseer.ist.psu.edu/cis. [8] Giles, C. L., Bollacker, K. D., and Lawrence, S. CiteSeer: An Automatic Citation Indexing System. The Third ACM Conference on Digital Libraries (1998), 89-98. [9] Google, http://www.google.com. [10] IEEE Xplore, http://ieeexplore.ieee.org/Xplore/DynWel.jsp. [11] INSPEC, http://www.iee.org/Publish/INSPEC/. [12] Jain, R. Experiential Computing. Communications of the ACM, 46, 7 (July 2003), 48-54. [13] Jain, R., Kim, P., and Li, Z. Experiential Meeting System. Proc. of the 2003 ACM SIGMM Workshop on Experiential Telepresence (ETP '03), 1-12. [14] Rowe, N. C. Marie-4: A High-Recall, Self-Improving Web Crawler that Finds Images using Captions. IEEE Intelligent Systems, 17, 4 (2002), 8-14. [15] Shkapenyuk, V., and Suel, T. Design and Implementation of a High-Performance Distributed Web Crawler. Proc. of the Intl. Conf. on Data Engineering (ICDE '02). [16] Singh, R., Knickmeyer, R. L., Gupta, P., and Jain, R. Designing Experiential Environments for Management of Personal Multimedia. ACM Multimedia 2004. To Appear. [17] Singh, R., Li, Z., Kim, P., and Jain, R. Event-Based Modeling and Processing of Digital Media. 1 st ACM SIGMOD Workshop on Computer Vision Meets Databases, Paris, France, 2004. [18] Teng, S-H., Lu, Q., and Eichstaedt, M. Collaborative Web Crawling: Information Gathering/Processing over Internet. Proc. Of the 32 nd Hawaii Intl. Conf. on System Sciences (1999). [19] Toyama, K., Logan, R., Roseway, A., and Anandan, P. Geographic Location Tags on Digital Images. ACM Multimedia (2003), 156-166. [20] Wactlar, H. D., Christel, M. G., Hauptmann A. G., and Gong, Y. Informedia Experience-on-Demand: Capturing, Integrating and Communicating Experiences across People, Time and Space. ACM Computing Surveys 31(1999). [21] WordNet 2.0 http://www.cogsci.princeton.edu/cgi-bin/webwn . Table 1. Comparisons of ResearchExplorer and other systems Functions ResearchExplorer Google CiteSeer ACM Digital Library Show spatio-temporal relationships Yes No No Can only list results by date order. Same query and presentation space? Yes No No No Dynamic query Yes No No No Maintain query state Yes No No No Zoom-in/zoom-out Yes No No No Visualization techniques Multiple Listing only Listing only Listing only 14

Event;Research Event;Multimedia Data;Spatio-Temporal Data;Exploration;Interaction Environment;Insight

169

Robustness Analysis of Cognitive Information Complexity Measure using Weyuker Properties

Cognitive information complexity measure is based on cognitive informatics, which helps in comprehending the software characteristics. For any complexity measure to be robust, Weyuker properties must be satisfied to qualify as good and comprehensive one. In this paper, an attempt has also been made to evaluate cognitive information complexity measure in terms of nine Weyuker properties, through examples. It has been found that all the nine properties have been satisfied by cognitive information complexity measure and hence establishes cognitive information complexity measure based on information contained in the software as a robust and well-structured one.

Introduction Many well known software complexity measures have been proposed such as McCabe's cyclomatic number [8], Halstead programming effort[5], Oviedo's data flow complexity measures[9], Basili's measure[3][4], Wang's cognitive complexity measure[11] and others[7]. All the reported complexity measures are supposed to cover the correctness, effectiveness and clarity of software and also to provide good estimate of these parameters. Out of the numerous proposed measures, selecting a particular complexity measure is again a problem, as every measure has its own advantages and disadvantages. There is an ongoing effort to find such a comprehensive complexity measure, which addresses most of the parameters of software. Weyuker[14] has suggested nine properties, which are used to determine the effectiveness of various software complexity measures. A good complexity measure should satisfy most of the Weyuker's properties. A new complexity measure based on weighted information count of a software and cognitive weights has been developed by Kushwaha and Misra[2]. In this paper an effort has been made to estimate this cognitive information complexity measure as robust and comprehensive one by evaluating this against the nine Weyuker's properties. Cognitive Weights of a Software Basic control structures [BCS] such as sequence, branch and iteration [10][13] are the basic logic building blocks of any software and the cognitive weights (W c ) of a software [11] is the extent of difficulty or relative time and effort for comprehending a given software modeled by a number of BCS's. These cognitive weights for BCS's measure the complexity of logical structures of the software. Either all the BCS's are in a linear layout or some BCS's are embedded in others. For the former case, we sum the weights of all the BCS's and for the latter, cognitive weights of inner BCS's are multiplied with the weight of external BCS's. Cognitive Information Complexity Measure (CICM) Since software represents computational information and is a mathematical entity, the amount of information contained in the software is a function of the identifiers that hold the information and the operators that perform the operations on the information i.e. Information = f (Identifiers, Operators) Identifiers are variable names, defined constants and other labels in a software. Therefore information contained in one line of code is the number of all operators and operands in that line of code. Thus in k th line of code the Information contained is: I k = (Identifiers + Operands) k = (ID k + OP k ) IU Where ID = Total number of identifiers in the k th LOC of software, OP = Total number of operators in the k th LOC of software, IU is the Information Unit representing that at least any identifier or operator has one unit information in them. Total Information contained in a software (ICS) is sum of information contained in each line of code i.e. LOCS ICS = (I k ) k=1 Where I k = Information contained in k th line of code, LOCS = Total lines of code in the software. Thus, it is the information contained in the identifiers and the necessary operations carried out by the operators in achieving the desired goal of the software, which makes software difficult to understand. Once we have established that software can be comprehended as information defined in information units (IU's) [2], the weighted information count is defined as : The Weighted Information Count of a line of code (WICL) of a software is a function of identifiers, operands and LOC and is defined as : ACM SIGSOFT Software Engineering Notes Page 1 January 2006 Volume 31 Number 1 ACM SIGSOFT Software Engineering Notes Page 1 January 2006 Volume 31 Number 1 WICL k = ICS k / [LOCS k] Where WIC k = Weighted Information Count for the k th line, ICS k = Information contained in a software for the k th line. The Weighted Information Count of the Software (WICS) is defined as : LOCS WICS = WICL k k = 1 In order to be a complete and robust measure, the measure of complexity should also consider the internal control structure of the software. These basic control structures have also been considered as the Newton's law in software engineering [10, 11]. These are a set of fundamental and essential flow control mechanisms that are used for building the logical architectures of software. Using the above definitions, Cognitive Information Complexity Measure (CICM) is defined as the product of weighted information count of software(WICS) and the cognitive weight (W c ) of the BCS in the software i.e. CICM = WICS * W c This complexity measure encompasses all the major parameters that have a bearing on the difficulty in comprehending software or the cognitive complexity of the software. It clearly establishes a relationship between difficulty in understanding software and its cognitive complexity. It introduces a method to measure the amount of information contained in the software thus enabling us to calculate the coding efficiency (E I ) as ICS / LOCS [2]. Evaluation of Cognitive Information Complexity Measure Weyuker[14] proposed the nine properties to evaluate any software complexity measure. These properties also evaluate the weakness of a measure in a concrete way. With the help of these properties one can determine the most suitable measure among the different available complexity measures. In the following paragraphs, the cognitive information complexity measure has been evaluated against the nine Weyuker properties for establishing itself as a comprehensive measure. Property 1: ( P)( Q)(|P| |Q|) Where P and Q are program body. This property states that a measure should not rank all programs as equally complex. Now consider the following two examples given in Fig. 1 and Fig. 2. For the program given in Fig. 1 in Appendix I, there are two control structures: a sequential and a iteration. Thus cognitive weight of these two BCS's is 1 + 3 = 4. Weighted information count for the above program is as under: WICS = 3/6 + 1/4 + 6/3+ 4/2 = 4.75 Hence Cognitive information complexity measure (CICM) is: CICM = WICS * W c = 4.75 * 4 = 19.0 For the program given in Fig. 2 in Appendix I there is only one sequential structure and hence the cognitive weight W c is 1. WICS for the above program is 2.27. Hence CICM for the above program is 2.27 * 1 = 2.27. From the complexity measure of the above two programs, it can be seen that the CICM is different for the two programs and hence satisfies this property. Property 2: Let c be a non-negative number. Then there are only finitely many programs of complexity c. Calculation of WICS depends on the number of identifiers and operators in a given program statement as well as on the number of statements remaining that very statement in a given program. Also all the programming languages consist of only finite number of BCS's. Therefore CICM cannot rank complexity of all programs as c. Hence CICM holds for this property. Property 3: There are distinct programs P and Q such that |P| = |Q|. For the program given in Fig. 3 in Appendix I, the CICM for the program is 19, which is same as that of program in Fig. 1. Therefore this property holds for CICM. Property 4: ( P)( Q) (PQ & |P| |Q|) Referring to program illustrated in Fig.1, we have replaced the while loop by the formula "sum = (b+1)*b/2" and have illustrated the same in Fig.2. Both the programs are used to calculate the sum of first n integer. The CICM for both the programs is different, thus establishing this property for CICM. Property 5: ( P)( Q)(|P| |P;Q| and |Q| |P;Q|). Consider the program body given in Fig.4 in Appendix I: The program body for finding out the factorial of a number consists of one sequential and one branch BCS's. Therefore W c = 3. For the program body for finding out the prime number, there are one sequential, one iteration and two branch BCS's. Therefore W c = 1 + 2*3*2 = 13. For the main program body for finding out the prime and factorial of the number, there are one sequential, two call and one branch BCS's. Therefore W c = 1+5+15+2 = 23. WICS for the program is 5.1. Therefore the Cognitive Information Complexity Measure for the above program = 5.1 * 23 = 117.3. Now consider the program given in Fig.5 in Appendix I to check for prime. There is one sequential, one iteration and three branch BCS's. Therefore W c = 1 + 2*3*2 + 2 = 15. WICS = 1.85. So CICM = 1.85 * 15 = 27.79. For the program given in Fig.6 in Appendix I, there is one sequential, one iteration and one branch BCS's . Wc for this program is 7 and WICS is .5.11. Hence CICM = WICS * W c = 5.11 * 7 = 35.77. It is clear from the above example that if we take the two-program body, one for calculating the factorial and another for checking for ACM SIGSOFT Software Engineering Notes Page 2 January 2006 Volume 31 Number 1 prime whose CICM are 27.79 and 35.77 that are less than 117.3. So property 5 also holds for CICM. Property 6(a) : ( P)( Q)(R)(|P| = |Q|) & (|P;R| |Q;R|) Let P be the program illustrated in Fig.1 and Q is the program illustrated in Fig.3. The CICM of both the programs is 19. Let R be the program illustrated in Fig.6. Appending R to P we have the program illustrated in Fig.7 in Appendix I. Cognitive weight for the above program is 9 and WICS is 8.3. Therefore CICM = 8.3*9=74.7. Similarly appending R to Q we have W c = 9 and WICS = 8.925. Therefore CICM = 8.925*9 = 80.325 and 74.7 80.325. This proves that Property 6(a) holds for CICM. Property 6(b): ( P)( Q)(R)(|P| = |Q|) & (|R;P| |R:Q|) To illustrate the above property let us arbitrarily append three program statements in the programs given in Fig.1, we have the program given in Fig.8 in Appendix I. There is only one sequential and one iteration BCS. Hence cognitive weight is 1 + 3 = 4. There is only one sequential and one iteration BCS. Hence cognitive weight is 1 + 3 = 4 and WICS = 5.58. So CICM = 5.58 * 4 = 22.32. Similarly appending the same three statements to program in Fig.3 we again have cognitive weights = 4 and WICS = 5.29. Therefore CICM = 21.16 22.32. Hence this property also holds for CICM. Property 7: There are program bodies P and Q such that Q is formed by permuting the order of the statement of P and (|P| |Q|). Since WICS is dependent on the number of operators and operands in a given program statement and the number of statements remaining after this very program statement, hence permuting the order of statement in any program will change the value of WICS. Also cognitive weights of BCS's depend on the sequence of the statement[1]. Hence CICM will be different for the two programs. Thus CICM holds for this property also. Property 8 : If P is renaming of Q, then |P| = |Q|. CICM is measured in numeric and naming or renaming of any program has no impact on CICM. Hence CICM holds for this property also. Property 9: ( P)( Q)(|P| + |Q|) &lt; (|P;Q|) OR ( P)( Q)(R)(|P| + |Q| + |R|) &lt; (|P;Q;R|) For the program illustrated in Fig.4, if we separate the main program body P by segregating Q (prime check) and R (factorial), we have the program illustrated in Fig.9 as shown in Appendix I. The above program has one sequential and one branch BCS. Thus cognitive weight is 4 and WICS is 1.475. Therefore CICM = 4.425. Hence 4.425 + 27.79 + 35.77 &lt; 117.3. This proves that CICM also holds for this property. Comparative Study of Cognitive Information Complexity Measure and Other Measures in Terms of Weyuker Properties In this section cognitive information complexity measure has been compared with other complexity measures in terms of all nine Weyuker's properties. P.N.- Property Number, S.C.- Statement Count, C N. - Cyclomatic Number, E.M.- Effort Measure, D.C.-Dataflow Complexity, C.C.M. - Cognitive Complexity Measure, CICM Cognitive Information Complexity Measure, Y- Yes, N NO P N SC CN EM DC CCM CICM. 1 Y Y Y Y Y Y 2 Y N Y N Y Y 3 Y Y Y Y Y Y 4 Y Y Y Y Y Y 5 Y Y N N Y Y 6 N N Y Y N Y 7 N N N Y Y Y 8 Y Y Y Y Y Y 9 N N Y Y Y Y Table 1: Comparison of complexity measures with Weyuker properties. It may be observed from the table 1 that complexity of a program using effort measure, data flow measure and Cognitive Information Complexity Measure depend directly on the placement of statement and therefore all these measures hold for property 6 also. All the complexity measure intend to rank all the programs differently Conclusion Software complexity measures serves both as an analyzer and a predicator in quantitative software engineering. Software quality is defined as the completeness, correctness, consistency, no misinterpretation, and no ambiguity, feasible verifiable in both specification and implementation. For a good complexity measure it is very necessary that the particular complexity measure not only satisfies the above-mentioned property of software quality but also satisfies the nine Weyuker properties. The software complexity in terms of cognitive information complexity measure thus has been established as a well- structured complexity measure . References [1] Misra,S and Misra,A.K.(2005): Evaluating Cognitive Complexity measure with Weyuker Properties, Proceeding of the3 rd IEEE International Conference on Cognitive Informatics(ICCI'04). ACM SIGSOFT Software Engineering Notes Page 3 January 2006 Volume 31 Number 1 [2] Kushwaha,D.S and Misra,A.K.(2005): A Modified Cognitive Information Complexity Measure of Software, Proceeding of the 7 th International Conference on Cognitive Systems(ICCS'05) (accepted for presentation) [3] Basili,V.R. and Phillips(1983): T.Y,Metric analysis and data validation across fortran projection .IEEE Trans.software Eng.,SE 9(6):652-663,1983. [4] Basili,V.R.(1980): Qualitative software complexity model: A summary in tutorial on models and method for software management and engineering .IEEE Computer Society Press ,Los Alamitos,CA,1980 . [5] Halstead,M.(1977): Elements of software science,Elsevier North Holland,New York.1997. [6] Klemola, T. and Rilling, J.(2003): A Cognitive Complexity Metric Based on Category Learning, , Proceeding of the 2 nd IEEE International Conference on Cognitive Informatics(ICCI'03). [7] Kearney, J.K., Sedlmeyer, R. L., Thompson, W.B., Gray, M. A. and Adler, M. A.(1986): Software complexity measurement. ACM Press, Newyork, 28:1044-1050,1986 [8] McCabe, T.A.(1976): Complexity measure. IEEE trans.software engg.,(se-2,6):308-320,1976. [9] Oviedo, E.(1980): Control flow, data and program complexity .in Proc. IEEE COMPSAC, Chicago, lL, pages 146-152,November 1980. [10] Wang. and Shao,J.(2002): On cognitive informatics, keynote lecture, proceeding of the .1 st IEEE International Conference on Cognitive Informatics, pages 34-42,August 2002. [11] Wang ,Y .and Shao,J.(2004): Measurement of the Cognitive Functional Complexity of Software, Proceeding of the 3 rd IEEE International Conference on Cognitive Informatics(ICCI'04) [12] Wang,Y .and Shao,J.(2003): On cognitive informatics, Proceeding of the 2 nd IEEE International Conference on Cognitive Informatics(ICCI'03),London,England,IEEE CS Press, pages 67-71 ,August 2003 . [13] Wang, Y.(2002): The real time process algebra (rtpa). Annals of Software Engineering, an international journal, and 14:235-247 ,2002. [14] Weyuker, E.(1988): Evaluating software complexity measure. IEEE Transaction on Software Complexity Measure, 14(9): 1357-1365 ,september1988. ACM SIGSOFT Software Engineering Notes Page 4 January 2006 Volume 31 Number 1 Appendix I /*Calculate the sum of first n integer*/ main() { int i, n, sum=0; printf(&quot;enter the number&quot;); //BCS1 scanf(&quot;%d&quot; , &n); for (i=1;i&lt;=n;i++) //BCS2 sum=sum+i; printf(&quot;the sum is %d&quot; ,sumssss); getch();} Fig. 1 : Source code of the sum of first n integers. main() { int b; int sum = 0; Printf("Enter the Number"); Scanf("%d", &n); Sum = (b+1)*b/2; Printf("The sum is %d",sum); getch(); } Fig. 2 : Source code to calculate sum of first n integers. # define N 10 main( ) { int count float, sum,average,number; sum = count =0; while (count &lt; N ) { scanf (" %f",& number); sum = sum+ number; count = count+1; } average = sum / N; printf ("Average =%f",average); } Fig. 3 : Source code to calculate the average of a set of N numbers. #include&lt; stdio.h &gt; #include&lt; stdlib.h &gt; int main() { long fact(int n); int isprime(int n); int n; long int temp; clrscr(); printf(&quot;\n input the number&quot;); //BCS11 scanf(&quot;%d&quot;,&n); temp=fact(n); //BCS12 {printf(&quot;\n is prime&quot;);} int flag1=isprime(n); //BCS13 if (flag1= =1) //BCS14 else {printf(&quot;\n is not prime&quot;)}; printf(&quot;\n factorial(n)=%d&quot;,temp); getch(); long fact(int n) { long int facto=1; //BCS21 if (n= =0) //BCS22 facto=1;else facto=n*fact(n-1); return(facto); } int isprime(int n) { int flag; //BCS31 if (n= =2) flag=1; //BCS32 else for (int i=2;i&lt;n;i++) //BCS33 { if (n%i= =0) //BCS34 { flag=0; Therefore Wc = 3 break; } else { flag=1 ;}} return (flag);}} Fig. 4: Source code to check prime number and to calculate factorial of the number #include&lt; stdio.h &gt; #include&lt; stdlib.h &gt; #include&lt; conio.h &gt; int main() { //BCS1 int flag = 1,n; clrscr(); printf(&quot;\ n enter the number&quot;); scanf(&quot;%d&quot;,&n); if (n= =2) flag=1; //BCS21 else {for (int i=2;i&lt;n;i++) //BCS22 if (n%i= =0) //BCS23 { flag=0; break;} else{ flag=1; continue;} } if(flag) //BCS3 printf(&quot;the number is prime&quot;); else printf(&quot;the number is not prime&quot;); grtch();} Fig.5 : Source code for checking prime number #include&lt; stdio.h &gt; #include&lt; stdlib.h &gt; #include&lt; conio.h &gt; int main () { long int fact=1; int n; clrscr(); ACM SIGSOFT Software Engineering Notes Page 5 January 2006 Volume 31 Number 1 printf(&quot;\ input the number&quot;); //BCS1 scanf(&quot;%d&quot;,&n); if (n==0) //BCS21 else for(int i=n;i&gt;1;i--) //BCS22 fact=fact*i; printf(&quot;\n factorial(n)=%1d&quot;,fact); getch();} Fig.6 : Source code for calculating factorial of a number Int main() { long fact(int n); int i, n, sum=0; printf(&quot;enter the number&quot;); scanf(&quot;%d&quot; , &n); temp = fact(n); for (i=1;i&lt;=n;i++) sum=sum+i; printf(&quot;the sum is %d&quot; ,sum); getch(); long fact(int n){ long int facto = 1; if (n == 0) facto = 1 else facto = n*fact(n-1); return(facto);}} Fig.7: Source code of sum of first n integer and factorial of n. main() { int a,b,result; result = a/b; printf(the result is %d",result); int i, n, sum=0; printf(&quot;enter the number&quot;); scanf(&quot;%d&quot; , &n); for (i=1;i&lt;=n;i++) sum=sum+i; printf(&quot;the sum is %d&quot; ,sum); getch();} Fig. 8 : Source code of division and the sum of first n integers. int main(){ int n; long int temp; clrscr(); printf("\n input the number"); scanf("%d",&n); temp = fact(n); {printf("\ is prime");} int flag1 = isprime(n); if (flag1 == 1) else {printf("\n is not prime)}; printf("\n factorial(n) = %d",temp); getch();} Fig.9 : Source code of main program body of program in Fig.4 ACM SIGSOFT Software Engineering Notes Page 6 January 2006 Volume 31 Number 1

cognitive weight;cognitive information complexity measure;basic control structures;cognitive information complexity unit;Weighted information count

17

A Pseudo Random Coordinated Scheduling Algorithm for Bluetooth Scatternets

The emergence of Bluetooth as a default radio interface allows handheld devices to be rapidly interconnected into ad hoc networks. Bluetooth allows large numbers of piconets to form a scatternet using designated nodes that participate in multiple piconets. A unit that participates in multiple piconets can serve as a bridge and forwards traffic between neighbouring piconets. Since a Bluetooth unit can transmit or receive in only one piconet at a time, a bridging unit has to share its time among the different piconets. To schedule communication with bridging nodes one must take into account their availability in the different piconets, which represents a difficult , scatternet wide coordination problem and can be an important performance bottleneck in building scatternets. In this paper we propose the Pseudo-Random Coordinated Scatternet Scheduling (PCSS) algorithm to perform the scheduling of both intra and inter-piconet communication. In this algorithm Bluetooth nodes assign meeting points with their peers such that the sequence of meeting points follows a pseudo random process that is different for each pair of nodes. The uniqueness of the pseudo random sequence guarantees that the meeting points with different peers of the node will collide only occasionally. This removes the need for explicit information exchange between peer devices, which is a major advantage of the algorithm. The lack of explicit signaling between Bluetooth nodes makes it easy to deploy the PCSS algorithm in Bluetooth devices, while conformance to the current Bluetooth specification is also maintained. To assess the performance of the algorithm we define two reference case schedulers and perform simulations in a number of scenarios where we compare the performance of PCSS to the performance of the reference schedulers.

INTRODUCTION Short range radio technologies enable users to rapidly interconnect handheld electronic devices such as cellular phones, palm devices or notebook computers. The emergence of Bluetooth [1] as default radio interface in these devices provides an opportunity to turn them from stand-alone tools into networked equipment. Building Bluetooth ad hoc networks also represents, however, a number of new challenges, partly stemming from the fact that Bluetooth was originally developed for single hop wireless connections. In this paper we study the scheduling problems of inter-piconet communication and propose a lightweight scheduling algorithm that Bluetooth nodes can employ to perform the scheduling of both intra and inter-piconet communication. Bluetooth is a short range radio technology operating in the unlicensed ISM (Industrial-Scientific-Medical) band using a frequency hopping scheme. Bluetooth (BT) units are organized into piconets. There is one Bluetooth device in each piconet that acts as the master , which can have any number of slaves out of which up to seven can be active simultaneously. The communication within a piconet is organized by the master which polls each slave according to some polling scheme. A slave is only allowed to transmit in a slave-to -master slot if it has been polled by the master in the previous master-to-slave slot. In Section 3 we present a brief overview of the Bluetooth technology. A Bluetooth unit can participate in more than one piconet at any time but it can be a master in only one piconet. A unit that participates in multiple piconets can serve as a bridge thus allowing the piconets to form a larger network. We define bridging degree as the number of piconets a bridging node is member of. A set of piconets that are all interconnected by such bridging units is referred to as a scatternet network (Figure 1). Since a Bluetooth unit can transmit or receive in only one piconet at a time, bridging units must switch between piconets on a time division basis. Due to the fact that different piconets are not synchronized in time a bridging unit necessarily loses some time while switching from one piconet to the other. Furthermore, the temporal unavailability of bridging nodes in the different piconets makes it difficult to coordinate the communication with them, which impacts throughput and can be an important performance constraint in building scatternets. There are two important phenomena that can reduce the efficiency of the polling based communication in Bluetooth scatternets: slaves that have no data to transmit may be unnecessarily polled, while other slaves with data to transmit may have to wait to be polled; and at the time of an expected poll one of the nodes of a master-slave node pair may not be present in the piconet (the slave that is being polled is not listening or the master that is expected to poll is not polling). The first problem applies to polling based schemes in general, while the second one is specific to the Bluetooth environment. In order to improve the efficiency of inter-piconet communication the scheduling algorithm has to coordinate the presence of bridging nodes in the different piconets such that the effect of the second phenomenon be minimized. However, the scheduling of inter-piconet communication expands to a scatternet wide coordination problem. Each node that has more than one Bluetooth links have to schedule the order in which it communicates with its respective neighbours. A node with multiple Bluetooth links can be either a piconet master or a bridging node or both. The scheduling order of two nodes will mutually depend on each other if they have a direct Bluetooth link in which case they have to schedule the communication on their common link for the same time slots. This necessitates some coordination between the respective schedulers. For instance in Figure 1 the scheduling order of node A and the scheduling order of its bridging neighbours, B, C, D and E mutually depend on each other, while nodes D and E further effects nodes F, G and H as well. Furthermore, the possible loops in a scatternet (e.g., A-E-G-H-F-D) makes it even more complicated to resolve scheduling conflicts. In case of bursty traffic in the scatternet the scheduling problem is further augmented by the need to adjust scheduling order in response to dynamic variation of traffic intensity. In a bursty traffic environment it is desirable that a node spends most of its time on those links that have a backlogged burst of data. One way to address the coordination problem of inter-piconet scheduling is to explicitly allocate, in advance, time slots for communication in each pair of nodes. Such a hard coordination approach eliminates ambiguity with regards to a node's presence in piconets, but it implies a complex, scatternet wide coordination problem and requires explicit signaling between nodes of a scatternet . In the case of bursty traffic, hard coordination schemes generate a significant computation and signaling overhead as the communication slots have to be reallocated in response to changes in traffic intensity and each time when a new connection is established or released. In this paper we propose the Pseudo-Random Coordinated Scatternet Scheduling algorithm which falls in the category of soft coordination schemes. In soft coordination schemes nodes decide their presence in piconets based on local information. By nature, soft coordination schemes cannot guarantee conflict-free participation of bridging nodes in the different piconets, however, they have a significantly reduced complexity. In the PCSS algorithm coordination is achieved by implicit rules in the communication without the need of exchanging explicit control information. The low complexity of the algorithm and its conformance to the current Bluetooth specification allow easy implementation and deployment. The first key component of the algorithm is the notion of checkpoints which are defined in relation to each pair of nodes that are connected by a Bluetooth link and which represent predictable points in time when packet transmission can be initiated on the particular link. In other words, checkpoints serve as regular meeting points for neighboring nodes when they can exchange packets. In order to avoid systematic collision of checkpoints on different links of a node the position of checkpoints follows a pseudo random sequence that is specific to the particular link the checkpoints belong to. The second key component of the algorithm is the dynamic adjustment of checking intensity, which is necessary in order to effec-tively support bursty data traffic. Bandwidth can be allocated and deallocated to a particular link by increasing and decreasing checkpoint intensity, respectively. To assess the performance of the algorithm we define two reference schedulers and relate the performance of the PCSS scheme to these reference algorithms in a number of simulation scenarios. The remainder of the paper is structured as follows. In Section 2 we give an overview of related work focusing on Bluetooth scheduling related studies available in the literature. Section 3 gives a brief overview of the Bluetooth technology. In Section 4 and 5 we introduce the proposed algorithm. In Section 6 we define the reference schedulers. Finally, in Section 7 we present simulation results. RELATED WORK A number of researchers have addressed the issue of scheduling in Bluetooth. Most of these studies have been restricted, however, to the single piconet environment, where the fundamental question is the polling discipline used by the piconet master to poll its slaves. These algorithms are often referred to as intra-piconet scheduling schemes. In [7] the authors assume a simple round robin polling scheme and investigate queueing delays in master and slave units depending on the length of the Bluetooth packets used. In [5] Johansson et al. analyze and compare the behavior of three different polling algorithms. They conclude that the simple round robin scheme may perform poorly in Bluetooth systems and they propose a scheme called Fair Exhaustive Polling. The authors demonstrate the strength of this scheme and argue in favor of using multi-slot packets. Similar conclusions are drawn by Kalia et al. who argue that the traditional round robin scheme may result in waste and un-fairness [8]. The authors propose two new scheduling disciplines that utilize information about the status of master and slave queues. In [9, 10] the authors concentrate on scheduling policies designed with the aim of low power consumption. A number of scheduling policies are proposed which exploit either the park or sniff low power modes of Bluetooth. 194 Although the above studies have revealed a number of important performance aspects of scheduling in Bluetooth piconets, the algorithms developed therein are not applicable for inter-piconet communication . In [6] the authors have shown that constructing an optimal link schedule that maximizes total throughput in a Bluetooth scatternet is an NP hard problem even if scheduling is performed by a central entity. The authors also propose a scheduling algorithm referred to as Distributed Scatternet Scheduling Algorithm (DSSA), which falls in the category of distributed, hard coordination schemes. Although the DSSA algorithm provides a solution for scheduling communication in a scatternet, some of its idealized properties (e.g., nodes are aware of the traffic requirements of their neighbours) and its relatively high complexity make it difficult to apply it in a real life environment. There is an ongoing work in the Personal Area Networking (PAN) working group of the Bluetooth Special Interest Group (SIG) [2] to define an appropriate scheduling algorithm for Bluetooth scatternets BLUETOOTH BACKGROUND Bluetooth is a short range radio technology that uses frequency hopping scheme, where hopping is performed on 79 RF channels spaced 1 MHz apart. Communication in Bluetooth is always between master and slave nodes. Being a master or a slave is only a logical state: any Bluetooth unit can be a master or a slave. The Bluetooth system provides full-duplex transmission based on slotted Time Division Duplex (TDD) scheme, where each slot is 0.625 ms long. Master-to-slave transmission always starts in an even-numbered time slot, while slave-to-master transmission always starts in an odd-numbered time slot. A pair of master-to-slave and slave-to-master slots are often referred to as a frame. The communication within a piconet is organized by the master which polls each slave according to some polling scheme. A slave is only allowed to transmit in a slave-to-master slot if it has been polled by the master in the previous master-to-slave slot. The master may or may not include data in the packet used to poll a slave. Bluetooth packets can carry synchronous data (e.g., real-time traffic) on Synchronous Connection Oriented (SCO) links or asynchronous data (e.g., elastic data traffic, which is the case in our study) on Asynchronous Connectionless (ACL) links. Bluetooth packets on an ACL link can be 1, 3 or 5 slot long and they can carry different amount of user data depending on whether the payload is FEC coded or not. Accordingly, the Bluetooth packet types DH1, DH3 and DH5 denote 1, 3 and 5 slot packets, respectively, where the payload is not FEC encoded, while in case of packet types DM1, DM3 and DM5 the payload is protected with FEC encoding. There are two other types of packets, the POLL and NULL packets that do not carry user data. The POLL packet is used by the master when it has no user data to the slave but it still wants to poll it. Similarly, the NULL packet is used by the slave to respond to the master if it has no user data. For further information regarding the Bluetooth technology the reader is referred to [1, 3]. OVERVIEW OF THE PCSS ALGORITHM Coordination in the PCSS algorithm is achieved by the unique pseudo random sequence of checkpoints that is specific to each master-slave node pair and by implicit information exchange between peer devices. A checkpoint is a designated Bluetooth frame. The activity of being present at a checkpoint is referred to as to check. A master node actively checks its slave by sending a packet to the slave at the corresponding checkpoint and waiting for a response from the slave. The slave node passively checks its master by listening to the master at the checkpoint and sending a response packet in case of being addressed. The expected behaviour of nodes is that they show up at each checkpoint on all of their links and check their peers for available user data. The exchange of user data packets started at a checkpoint can be continued in the slots following the checkpoint. A node remains active on the current link until there is user data in either the master-to-slave or slave-to-master directions or until it has to leave for a next checkpoint on one of its other links. In the PCSS scheme we exploit the concept of randomness in assigning the position of checkpoints, which excludes the possibility that checkpoints on different links of a node will collide systematically, thus giving the node an equal chance to visit all of its checkpoints. The pseudo random procedure is similar to the one used to derive the pseudo random frequency hopping sequence. In particular, the PCSS scheme assigns the positions of checkpoints on a given link following a pseudo random sequence that is generated based on the Bluetooth clock of the master and the MAC address of the slave. This scheme guarantees that the same pseudo random sequence will be generated by both nodes of a master-slave pair, while the sequences belonging to different node pairs will be different. Figure 2 shows an example for the pseudo random arrangement of checkpoints in case of a node pair A and B. The length of the current base checking interval is denoted by T (i) check and the current checking intensity is defined accordingly as 1 T (i) check . There is one checkpoint within each base checking interval and the position of the checkpoint within this window is changing from one time window to the other in a pseudo random manner. checkpoints of A toward B checkpoints of B toward A 1 frame T (i) check Figure 2: Pseudo-random positioning of checkpoints Since the pseudo random sequence is different from one link to another , checkpoints on different links of a node will collide only occasionally . In case of collision the node can attend only one of the colliding checkpoints, which implies that the corresponding neighbours have to be prepared for a non-present peer. That is, the master might not poll and the slave might not listen at a checkpoint. We note that a collision occurs either if there are more than one checkpoints scheduled for the same time slot or if the checkpoints are so close to each other that a packet transmission started at the first checkpoint necessarily overlaps the second one. Furthermore, if the colliding checkpoints belong to links in different piconets, the necessary time to perform the switch must be also taken into account. During the communication there is the possibility to increase or decrease the intensity of checkpoints depending on the amount of user data to be transmitted and on the available capacity of the node. According to the PCSS algorithm a node performs certain traffic measurements at the checkpoints and increases or decreases the current checking intensity based on these measurements. Since 195 nodes decide independently about the current checking intensity without explicit coordination, two nodes on a given link may select different base checking periods. In order to ensure that two nodes with different checking intensities on the same link can still communicate we require the pseudo random generation of checkpoints to be such that the set of checkpoint positions at a lower checking intensity is a subset of checkpoint positions at any higher checking intensities. In the Appendix we are going to present a pseudo random scheme for generating the position of checkpoints, which has the desired properties. OPERATION OF PCSS In what follows, we describe the procedures of the PCSS algorithm. We start by the initialization process which ensures that two nodes can start communication as soon as a new link has been established or the connection has been reset. Next, we describe the rules that define how nodes calculate their checkpoints, decide upon their presence at checkpoints and exchange packets. Finally, we present the way neighboring nodes can dynamically increase and decrease of checkpoint intensity. 5.1 Initialization In the PCSS algorithm there is no need for a separate initialization procedure to start communication, since the pseudo random generation of checkpoints is defined such that once a master slave node pair share the same master's clock and slave's MAC address information , it is guaranteed that the same pseudo random sequence will be produced at each node. That is, it is guaranteed that two nodes starting checkpoint generation at different time instants with different checking intensities will be able to communicate. It is the own decision of the nodes to select an appropriate initial checking intensity , which may depend for example on the free capacities of the node or on the amount of data to transmit. Once the communication is established the increase and decrease procedures will adjust the possibly different initial checking intensities to a common value. 5.2 Communication A pair of nodes can start exchanging user data packets at a checkpoint , which can expand through the slots following the checkpoint. The nodes remain active on the current link following a checkpoint until there is user data to be transmitted or one of them has to leave in order to attend a checkpoint on one of its other links. After a POLL/NULL packet pair has been exchanged indicating that there is no more user data left the nodes switch off their transmit-ters/receivers and remain idle until a next checkpoint comes on one of their links. However, during the communication any of the nodes can leave in order to attend a coming checkpoint on one of its other links. After one of the nodes has left the remaining peer will realize the absence of the node and will go idle until the time of its next checkpoint. If the master has left earlier the slave will realize the absence of the master at the next master-to-slave slot by not receiving the expected poll. In the worst case the master has left before receiving the last packet response from the slave, which can be a 5 slot packet in which case the slave wastes 5+1 slots before realizing the absence of the master. Similarly, if the master does not get a response from the slave it assumes that the slave has already left the checkpoint and goes idle until its next checkpoint. Note that the master may also waste 5+1 slots in the worst case before realizing the absence of the slave. A node stores the current length of the base checking interval and the time of the next checkpoint for each of its Bluetooth links separately . For its i th link a node maintains the variable T (i) check to store the length of the current base checking period in number of frames and the variable t (i) check , which stores the Bluetooth clock of the master at the next checkpoint. After passing a checkpoint the variable t (i) check is updated to the next checkpoint by running the pseudo random generator ( P seudoChkGen) with the current value of the master's clock t (i) and the length of the base checking period T (i) check and with the MAC address of the slave A (i) slave as input parameters; t (i) check = P seudoChkGen(T (i) check , A (i) slave , t (i) ). The procedure P seudoChkGen is described in the Appendix. There is a maximum and minimum checking interval T max = 2 f max and T min = 2 f min , respectively. The length of the checking period must be a power of 2 number of frames and it must take a value from the interval [2 f min , 2 f max ]. 5.3 Increasing and Decreasing Checking Intensity The increase and decrease procedures are used to adjust the checking intensity of a node according to the traffic intensity and to the availability of the peer device. Each node decides independently about the current checking intensity based on traffic measurements at checkpoints. Since the time spent by a node on a link is proportional to the ratio of the number of checkpoints on that link and the number of checkpoints on all links of the node, the bandwidth allocated to a link can be controlled by the intensity of checkpoints on that link. This can be shown by the following simple calculation. Let us assume that the node has L number of links and assume further that for the base checking periods on all links of the node it holds that T min T (i) check T max , i = 1, . . . , L. Then the average number of checkpoints within an interval of length T max is N = L i =1 T max T (i) check , and the average time between two consecutive checkpoints is t = T max N = 1 L i =1 1 T (i) check , provided that the pseudo random generator produces a uniformly distributed sequence of checkpoints. Then, the share of link j from the total capacity of the node is r j = 1/T (j) check L i =1 1 T (i) check . A node has to measure the utilization of checkpoints on each of its links separately in order to provide input to the checking intensity increase and decrease procedures. According to the algorithm a given checkpoint is considered to be utilized if both nodes have shown up at the checkpoint and at least one Bluetooth packet carrying user data has been transmitted or received. If there has not been a successful poll at the checkpoint due to the unavailability of any of the nodes or if there has been only a POLL/NULL packet pair exchange but no user data has been transmitted, the checkpoint is considered to be unutilized. We note that due to packet losses the utilization of a given checkpoint might be interpreted differently by the nodes. However, this does not impact correct operation of the algorithm. 196 To measure the utilization of checkpoints (i) on the i th link of the node we employ the moving average method as follows. The utilization of a checkpoint equals to 1 if it has been utilized, otherwise it equals to 0. If the checkpoint has been utilized the variable (i) is updated as, (i) = q uti (i) + (1 - q uti ) 1; if the checkpoint has not been utilized it is updated as, (i) = q uti (i) + (1 - q uti ) 0, where 0 q uti &lt; 1 is the time scale parameter of the moving average method. A further parameter of the utilization measurement is the minimum number of samples that have to be observed before the measured utilization value is considered to be confident and can be used as input to decide about increase and decrease of checking intensity. This minimum number of samples is a denoted by N sample,min . Finally, a node also has to measure its total utilization, which is defined as the fraction of time slots where the node has been active (transmitted or received) over the total number of time slots. To measure the total utilization of a node we employ the moving average method again. Each node measures its own utilization (node) and updates the (node) variable after each N uti,win number of slots as follows: (node) = q (node) uti (node) + (1 - q (node) uti ) (win) , where (win) is the fraction of time slots in the past time window of length N uti,win where the node has been active over the total number of time slots N uti,win . If the utilization of checkpoints on link i falls below the lower threshold lower , the current base checking period T (i) check will be doubled. Having a low checkpoint utilization can be either because one or both of the nodes have not shown up at all of the checkpoints or because there is not enough user data to be transmitted. In either cases the intensity of checkpoints has to be decreased. Whenever a decrease or increase is performed on link i the measured utilization (i) must be reset. Since the parameter T (i) check is one of the inputs to the pseudo random checkpoint generation process, P seudoChkGen the checkpoints after the decrease will be generated according to the new period. Furthermore, due to the special characteristic of the checkpoint generation scheme the remaining checkpoints after the decrease will be a subset of the original checkpoints, which guarantees that the two nodes can sustain communication independent of local changes in checking intensities. An example for the checking intensity decrease in case of a node pair A and B is shown in Figure 3. First, node A decreases checking intensity by doubling its current base checking period in response to the measured low utilization. As a consequence node B will find node A on average only at every second checkpoint and its measured utilization will decrease rapidly. When the measured utilization at node B falls below the threshold lower , B realizes that its peer has a lower checking intensity and follows the decrease by doubling its current base checking period. Although we have not explicitly indicated in the Figure, it is assumed that there has been user data exchanged at each checkpoint where both nodes were present. =0.35&lt; lower =0.36&lt; lower node A reduces the checking intensity, by doubling its base period checkpoints of B toward A checkpoints of A toward B doubles its base period node B realizes the decrease and =0.6=0.5 =0.5 =0.2 =0.7 =0.48 =0.56 =0.46 =0.5 =0.58 =0.35 =0.35 =0.65 =0.2 Figure 3: Checking intensity decrease Recall from the utilization measurement procedure that there is a minimum number of checkpoints N sample,min that has to be sam-pled before the measured utilization is considered to be confident and can be used to decide about checking intensity decrease. The parameter N sample,min together with the parameter of the moving average method q uti determine the time scale over which the utilization of checkpoints has to be above the threshold lower , otherwise the node decreases checking intensity. It might be also reasonable to allow that the parameter N sample,min and the moving average parameter q uti can be changed after each decrease or increase taking into account for example the current checking intensity , the available resources of the node or the amount of user data to be transmitted, etc. However, in the current implementation we apply fixed parameter values. After a checkpoint where user data has been exchanged (not only a POLL/NULL packet pair) checking intensity can be increased provided that the measured utilization of checkpoints exceeds the upper threshold upper and the node has available capacity. Formally a checking intensity increase is performed on link i if the following two conditions are satisfied: (i) &gt; upper and (node) &lt; (node) upper , where (node) upper is the upper threshold of the total utilization of the node. This last condition ensures that the intensity of checkpoints will not increase unbounded. The intensity of checkpoints is doubled at each increase by dividing the current length of the base checking period T (i) check by 2. For typical values of upper we recommend 0.8 upper 0.9 in which case the respective lower value should be lower 0.4 in order to avoid oscillation of increases and decreases. Figure 4 shows an example where node A and B communicate and after exchanging user data at the second checkpoint both nodes double the checking intensity. In the Figure we have explicitly indicated whether there has been user data exchanged at a checkpoint or not. user data checkpoints of B toward A checkpoints of A toward B =0.8&gt; upper =0.8&gt; upper =0.7 =0.2 =0.55 =0.4 =0.55 =0.7 user data user data user data =0.4 =0.2 checking intensity both node A and B double =0.3 =0.3 Figure 4: Checking intensity increase 197 REFERENCE ALGORITHMS In this section we define the Ideal Coordinated Scatternet Scheduler (ICSS) and the Uncoordinated Greedy Scatternet Scheduler (UGSS) reference algorithms. The ICSS algorithm represents the "ideal" case where nodes exploit all extra information when scheduling packet transmissions, which would not be available in a realistic scenario. The UGSS algorithm represents the greedy case where nodes continuously switch among their Bluetooth links in a random order. 6.1 The ICSS Algorithm The ICSS algorithm is a hypothetical, ideal scheduling algorithm that we use as a reference case in the evaluation of the PCSS scheme. In the ICSS algorithm a node has the following extra information about its neighbours, which represents the idealized property of the algorithm: a node is aware of the already pre-scheduled transmissions of its neighbours; and a node is aware of the content of the transmission buffers of its neighbours. According to the ICSS algorithm each node maintains a scheduling list, which contains the already pre-scheduled tasks of the node. A task always corresponds to one packet pair exchange with a given peer of the node. Knowing the scheduling list of the neighbours allows the node to schedule communication with its neighbours without overlapping their other communication, such that the capacity of the nodes is utilized as much as possible. Furthermore being aware of the content of the transmission buffers of neighbours eliminates the inefficiencies of the polling based scheme, since there will be no unnecessary polls and the system will be work-conserving. In the scheduling list of a node there is at most one packet pair exchange scheduled in relation to each of its peers, provided that there is a Bluetooth packet carrying user data either in the transmission buffer of the node or in the transmission buffer of the peer or in both. After completing a packet exchange on a given link the two nodes schedule the next packet exchange, provided that there is user data to be transmitted in at least one of the directions. If there is user data in only one of the directions, a POLL or NULL packet is assumed for the reverse direction depending on whether it is the master-to-slave or slave-to-master direction, respectively. The new task is fitted into the scheduling lists of the nodes using a first fit strategy. According to this strategy the task is fitted into the first time interval that is available in both of the scheduling lists and that is long enough to accommodate the new task. Note that the algorithm strives for maximal utilization of node capacity by trying to fill in the unused gaps in the scheduling lists. If there is no more user data to be transmitted on a previously busy link, the link goes to idle in which case no tasks corresponding to the given link will be scheduled until there is user data again in at least one of the directions. An example for the scheduling lists of a node pair A and B is shown in Figure 5. The tasks are labeled with the name of the corresponding peers the different tasks belong to. Each node has as many pre-scheduled tasks in its scheduling list as the number of its active Bluetooth links. A link is considered to be active if there is schedule the next packet pair exchange for node A and B scheduling list of node A current time scheduling list of node B t t peer B peer A peer A peer B peer E peer E peer C peer D peer C Figure 5: Example for the scheduling lists of a node pair in case of the ICSS algorithm user data packet in at least one of the directions. Node A has active peers B and C, while node B has active peers A, D and E. After node A and B have finished the transmission of a packet pair they schedule the next task for the nearest time slots that are available in both of their scheduling lists and the number of consecutive free time slots is greater than or equal to the length of the task. 6.2 The UGSS Algorithm In the UGSS algorithm Bluetooth nodes do not attempt to coordinate their meeting points, instead each node visits its neighbours in a random order. Nodes switch continuously among their Bluetooth links in a greedy manner. If the node has n number of links it chooses each of them with a probability of 1/n. The greedy nature of the algorithm results in high power consumption of Bluetooth devices. If the node is the master on the visited link it polls the slave by sending a packet on the given link. The type of Bluetooth packet sent can be a 1, 3 or 5 slot packet carrying useful data or an empty POLL packet depending on whether there is user data to be transmitted or not. After the packet has been sent the master remains active on the link in order to receive any response from the slave. If the slave has not been active on the given link at the time when the master has sent the packet it could not have received the packet and consequently it will not send a response to the master. After the master has received the response of the slave or if it has sensed the link to be idle indicating that no response from the salve can be expected, it selects the next link to visit randomly. Similar procedure is followed when the node is the slave on the visited link. The slave tunes its receiver to the master and listens for a packet transmission from the master in the current master-to -slave slot. If the slave has not been addressed by the master in the actual master-to-slave slot it immediately goes to the next link. However, if the slave has been addressed it remains active on the current link and receives the packet. After having received the packet of the master the slave responds with its own packet in the following slave-to-master slot. After the slave has sent its response it selects the next link to visit randomly. SIMULATION RESULTS First, we evaluate the algorithm in a realistic usage scenario, which is the Network Access Point (NAP) scenario. Next we investigate theoretical configurations and obtain asymptotical results that reveals the scaling properties of the algorithm. For instance we investigate the carried traffic in function of the number of forwarding 198 hops along the path and in function of bridging degree. Both in the realistic and theoretical configurations we relate the performance of the PCSS scheme to the performance of the ICSS and UGSS reference algorithms. Before presenting the scenarios and simulation results we shortly describe the simulation environment and define the performance metrics that are going to be measured during the simulations. 7.1 Simulation Environment We have developed a Bluetooth packet level simulator, which is based on the Plasma simulation environment [4]. The simulator has a detailed model of all the packet transmission, reception procedures in the Bluetooth Baseband including packet buffering, upper layer packet segmentation/reassemble, the ARQ mechanism, etc. The simulator supports all Bluetooth packet types and follows the same master-slave slot structure as in Bluetooth. For the physical layer we employ a simplified analytical model that captures the frequency collision effect of interfering piconets. In the current simulations the connection establishment procedures, e.g., the inquiry and page procedures are not simulated in detail and we do not consider dynamic scatternet formation either. Instead we perform simulations in static scatternet configurations where the scatternet topology is kept constant during one particular run of simulation. In the current simulations we run IP directly on top of the Bluetooth link layer and we apply AODV as the routing protocol in the IP layer. The simulator also includes various implementations of the TCP protocol (we employed RenoPlus) and supports different TCP/IP applications, from which we used TCP bulk data transfer in the current simulations. One of the most important user perceived performance measures is the achieved throughput. We are going to investigate the throughput in case of bulk TCP data transfer and in case of Constant Bit Rate (CBR) sources. In order to take into account the power consumption of nodes we define activity ratio of a node, r act as the fraction of time when the node has been active over the total elapsed time; and power efficiency, p ef f as the fraction of the number of user bytes success-fully communicated (transmitted and received) over the total time the node has been active. The power efficiency shows the number of user bytes that can be communicated by the node during an active period of length 1 sec. Power efficiency can be measured in [kbit/sec], or assuming that being active for 1 sec consumes 1 unit of energy we can get a more straightforward dimension of [kbit/energy unit], which is interpreted as the number of bits that can be transmitted while consuming one unit of energy. 7.2 Network Access Point Scenario In this scenario we have a NAP that is assumed to be connected to a wired network infrastructure and it provides network access via its Bluetooth radio interface. The NAP acts as a master and up to 7 laptops, all acting as slaves, can connect to the NAP. Furthermore we assume that each laptop has a Bluetooth enabled mouse and each laptop connects to its mouse by forming a new piconet as it is shown in Figure 6. We simulate a bulk TCP data transfer from the NAP towards each laptop separately. Regarding the traffic generated by the mouse we assume that the mouse produces a 16 byte long packet each 50 ms, NAP laptop max 7 laptop mouse mouse Figure 6: Network Access Point Scenario periodically. In the NAP-laptop communication we are interested in the achieved throughput while in the laptop-mouse communication we are concerned with the delay perceived by the mouse. In the current scenario we switched off the dynamic checkperiod adjustment capability of the PCSS algorithm and we set the base checking period to 32 frames (40 ms), which is in accordance with the delay requirement of a mouse. Note that this same base checking period is applied also on the NAP-laptop links, although, there is no delay requirement for the TCP traffic. However, the current implementation in the simulator does not yet support the setting of the base checking periods for each link separately. The dynamic checking period adjustment would definitely improve the throughput of NAP-laptop communication as we are going to see later in case of other configurations. The simulation results are shown in Figure 7. In plot (a) the averaged throughput of NAP-laptop communications are shown in the function of number of laptops for the different algorithms, respectively . Graph (b) plots the sum of the throughputs between the NAP and all laptops. As we expect, the individual laptop throughput decreases as the number of laptops increases. However, it is important to notice that the sum of laptop throughputs do not decrease with increasing number of laptops in case of the PCSS and ICSS algorithms. As the number of laptops increases the efficient coordination becomes more important and the total carried traffic will decrease with the uncoordinated UGSS scheme. The increase of the total throughput in case of the PCSS algorithm is the consequence of the fixed checking intensities, which allocates one half of a laptop capacity to the mouse and the other half to the NAP. In case of small number of laptops this prevents the laptops to fully utilize the NAP capacity, which improves as the number of laptops increases. The 99% percentile of the delay seen by mouse packets is shown in plot (c). The delay that can be provided with the PCSS algorithm is determined by the base checking period that we use. Recall, that in the current setup the base checking period of the PCSS scheme was set to 32 frames, which implies that the delay has to be in the order of 32 frames, as shown in the figure. The low delay with the UGSS algorithm is due to the continuous switching among the links of a node, which ensures high polling intensity within a piconet and frequent switching between piconets. The UGSS algorithm provides an unnecessarily low delay, which is less than the delay requirement at the expense of higher power consumption. Plots (d) and (e) show the averaged activity ratio over all laptops and mice, respectively. The considerably higher throughput achieved for small number of laptops by the ICSS scheme explains its higher activity ratio. On graph (f) the averaged power efficiency of laptops is shown, which relates the number of bytes transmitted to the total time of activity. The power efficiency of the PCSS 199 scheme decreases with increasing number of laptops, which is a consequence of the fixed checking intensities. Since the NAP has to share its capacity among the laptops, with an increasing number of laptops there will be an increasing number of checkpoints where the NAP cannot show up. In such cases the dynamic checking intensity adjustment procedure could help by decreasing checking intensity on the NAP-laptop links. Recall that in the current scenario we employed fixed checking intensities in order to satisfy the mouse delay requirement. It is also important to notice that with the uncoordinated UGSS scheme the activity ratio of a mouse is relatively high, which is an important drawback considering the low power capabilities of such devices. 7.3 Impact of Number of Forwarding Hops In what follows, we investigate the performance impact of the number of forwarding hops along the communication path in the scatternet configuration shown in Figure 8. The configuration consists of a chain of S/M forwarding nodes ( F i ) and a certain number of additional node pairs connected to each forwarding node in order to generate background traffic. The number of S/M forwarding nodes is denoted by N F . There are N B number of background node pairs connected to each forwarding node as masters. The background traffic flows from each source node B (S) ij to its destination pair B (D) ij through the corresponding forwarding node F i . The traffic that we are interested in is a bulk TCP data transfer between node S and D. The background traffic is a CBR source, which generates 512 byte long IP packets with a period of length 0.05 sec. D B (D) 1i B (D) 2i B (D) NF i B (S) NF i B (S) 2i B (S) 1i F 1 F 2 F NF S Figure 8: Impact of number of forwarding nodes During the simulations we vary the number of forwarding hops N F and the number of background node pairs N B connected to each forwarding node. As one would expect, with increasing number of forwarding hops and background node pairs the coordinated algorithms will perform significantly better than the one without any coordination (UGSS). The throughput of the S-D traffic as a function of the number of forwarding nodes ( N F ) without background traffic ( N B = 0) and with two pairs of background nodes ( N B = 2) are shown in Figure 9 (a) and (b), respectively. The throughput in case of no cross traffic drops roughly by half when we introduce the first forwarding node. Adding additional forwarding hops continuously reduces the throughput, however, the decrease at each step is less drasti-cal . We note that in case of the ICSS scheme one would expect that for N F &gt; 1 the throughput should not decrease by adding additional forwarding hops. However, there are a number of other effects besides the number of forwarding hops that decrease the throughput. For instance, with an increasing number of forwarding hops the number of piconets in the same area increases, which, in turn, causes an increasing number of lost packets over the radio interface due to frequency collisions. Furthermore with increasing number of hops the end-to-end delay suffered by the TCP flow increases , which makes the TCP connection less reactive to recover from packet losses. In the no background traffic case the PCSS scheme performs close to the UGSS algorithm in terms of throughput. However, as we introduce two pairs of background nodes the UGSS algorithm fails completely, while the PCSS scheme still achieves approximately 20 kbit/sec throughput. Furthermore, the power efficiency of the PCSS scheme is an order of magnitude higher than that of the UGSS algorithm in both cases, which indicates that the PCSS algorithm consumes significantly less power to transmit the same amount of data than the UGSS scheme. 7.4 Impact of Bridging Degree Next we investigate the performance of scheduling algorithms as the number of piconets that a bridging node participates in is increased . The scatternet setup that we consider is shown in Figure 10, where we are interested in the performance of the bridging node C. Node C is an all slave bridging node and it is connected to master nodes P i , where the number of these master nodes is denoted by N P . To each master node P i we connect N L number of leaf nodes as slaves in order to generate additional background load in the piconets. We introduce bulk TCP data transfer from node C towards each of its master node P i and CBR background traffic on each L ij - P i link. The packet generation interval for background sources was set to 0.25 sec, which corresponds to a 16 kbit/sec stream. During the simulation we vary the number of piconets N P participated by node C and investigate the performance of the PCSS algorithm with and without dynamic checkpoint intensity changes. The number of background nodes N L connected to each master node P i was set to N L = 3 and it was kept constant in the simulations. C P 1 L 1i L 1N L L N P 1 L N P N L P N P Figure 10: Impact of number of participated piconets The throughputs of TCP flows between node C and each P i are averaged and it is shown in Figure 10 (a). The sum of TCP throughputs are plotted in graph (b) and the power efficiency of the central node is shown in graph (c). The PCSS algorithm has been tested both with fixed base checking periods equal to 32 frames ("PCSS-32" ) and with dynamic checking intensity changes as well ("PCSS-dyn" ). The parameter settings of the dynamic case is shown in Table 1. q uti = 0.7 N sample,min = 4 lower = 0.3 upper = 0.7 q (node) uti = 0.7 N uti,win = 10 (node) max = 0.8 T min = 8 T max = 256 Table 1: Parameter setting of the dynamic PCSS scheme 200 0 50 100 150 200 250 300 350 400 450 500 1 2 3 4 5 6 7 Throughput [kbit/s] Number of laptops TCP throughput per laptop PCSS UGSS ICSS 0 50 100 150 200 250 300 350 400 450 500 1 2 3 4 5 6 7 Throughput [kbit/s] Number of laptops Sum TCP throughput of laptops PCSS UGSS ICSS 0 0.01 0.02 0.03 0.04 0.05 0.06 1 2 3 4 5 6 7 Delay [sec] Number of laptops 0.99 percentile of mouse dealy PCSS UGSS ICSS (a) (b) (c) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 2 3 4 5 6 7 Activity ratio Number of laptops Activity Ratio of laptops PCSS UGSS ICSS 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 2 3 4 5 6 7 Activity ratio Number of laptops Activity Ratio of mice PCSS UGSS ICSS 0 100 200 300 400 500 600 1 2 3 4 5 6 7 kbit/Energy unit Number of laptops Power efficiency of laptops PCSS UGSS ICSS (d) (e) (f) Figure 7: Throughput, delay and power measures in the function of number of laptops connected to the NAP 0 50 100 150 200 250 300 350 400 450 500 0 1 2 3 4 5 6 7 8 TCP throughput [kbit/s] Number of forwarding nodes (N_F) TCP throughput without background nodes (N_B=0) PCSS UGSS ICSS 0 50 100 150 200 250 300 350 400 450 500 0 1 2 3 4 5 6 7 8 TCP throughput [kbit/s] Number of forwarding nodes (N_F) TCP throughput with 2 pairs of background nodes (N_B=2) PCSS UGSS ICSS 0 100 200 300 400 500 600 0 1 2 3 4 5 6 7 8 Power Efficiency [kbit/Energy unit] Number of forwarding nodes (N_F) Power efficiency of forwarding nodes (N_B=2) PCSS UGSS ICSS (a) (b) (c) Figure 9: Throughput and power efficiency in function of number of forwarding hops It is important to notice that the per flow TCP throughputs in case of the dynamic PCSS scheme matches quite closely the throughput achieved by the ICSS algorithm and it significantly exceeds the throughput that has been achieved by the fixed PCSS. This large difference is due to the relatively low background traffic in the neighbouring piconets of node C, in which case the dynamic PCSS automatically reduces checkpoint intensity on the lightly loaded links and allocates more bandwidth to the highly loaded ones by increasing checking intensity. CONCLUSIONS We have presented Pseudo Random Coordinated Scatternet Scheduling, an algorithm that can efficiently control communication in Bluetooth scatternets without exchange of control information between Bluetooth devices. The algorithm relies on two key components, namely the use of pseudo random sequences of meeting points, that eliminate systematic collisions, and a set of rules that govern the increase and decrease of meeting point intensity without explicit coordination. We have evaluated the performance of PCSS in a number of simulation scenarios, where we have compared throughput and power measures achieved by PCSS to those achieved by two reference schedulers. The first reference scheduler is an uncoordinated greedy algorithm, while the other is a hypothetical "ideal" scheduler . In all the scenarios investigated we have found that PCSS achieves higher throughput than the uncoordinated reference algorithm. Moreover, with the traffic dependent meeting point intensity adjustments the throughput and power measures of PCSS quite closely match the results of the "ideal" reference algorithm. At the same time PCSS consumes approximately the same amount of power as the ideal scheduler to achieve the same throughput, which is significantly less than the power consumption of the uncoordinated reference scheduler. REFERENCES [1] Bluetooth Special Interest Group. Bluetooth Baseband Specification Version 1.0 B. http://www.bluetooth.com/. 201 0 50 100 150 200 250 300 350 400 450 1 2 3 4 5 6 Throughput [kbit/s] Number of piconets participated by the central node (N_P) Averaged TCP throughput between central node and master nodes PCSS-32 PCSS-dyn UGSS ICSS 0 50 100 150 200 250 300 350 400 450 1 2 3 4 5 6 Throughput [kbit/s] Number of piconets participated by the central node (N_P) Sum of TCP throughputs at the central node PCSS-32 PCSS-dyn UGSS ICSS 0 100 200 300 400 500 600 1 2 3 4 5 6 Power efficiency [kbit/Energy unit] Number of piconets participated by the central node (N_P) Effective power of central node PCSS-32 PCSS-dyn UGSS ICSS (a) (b) (c) Figure 11: Throughput and power efficiency in function of the bridging degree of node C [2] Bluetooth Special Interest Group. http://www.bluetooth.com/. [3] J. Haartsen. BLUETOOTH- the universal radio interface for ad-hoc, wireless connectivity. Ericsson Review, (3), 1998. [4] Z. Haraszti, I. Dahlquist, A. Farago, and T. Henk. Plasma an integrated tool for ATM network operation. In Proc. International Switching Symposium, 1995. [5] N. Johansson, U. Korner, and P. Johansson. Performance evaluation of scheduling algorithms for Bluetooth. In IFIP TC6 WG6.2 Fifth International Conference on Broadband Communications (BC'99), Hong Kong, November 1999. [6] N. Johansson, U. Korner, and L. Tassiulas. A distributed scheduling algorithm for a Bluetooth scatternet. In Proc. of The Seventeenth International Teletraffic Congress, ITC'17, Salvador da Bahia, Brazil, September 2001. [7] P. Johansson, N. Johansson, U. Korner, J. Elgg, and G. Svennarp. Short range radio based ad hoc networking: Performance and properties. In Proc. of ICC'99, Vancouver, 1999. [8] M. Kalia, D. Bansal, and R. Shorey. MAC scheduling and SAR policies for Bluetooth: A master driven TDD pico-cellular wireless system. In IEEE Mobile Multimedia Communications Conference MOMUC'99, San Diego, November 1999. [9] M. Kalia, D. Bansal, and R. Shorey. MAC scheduling policies for power optimization in Bluetooth: A master driven TDD wireless system. In IEEE Vehicular Technology Conference 2000, Tokyo, 2000. [10] M. Kalia, S. Garg, and R. Shorey. Efficient policies for increasing capacity in Bluetooth: An indoor pico-cellular wireless system. In IEEE Vehicular Technology Conference 2000, Tokyo, 2000. APPENDIX Here, we present the procedure for generating the pseudo random sequence of checkpoints, where we reuse the elements of the pseudo random frequency hop generation procedure available in Bluetooth. The inputs to the checkpoint generation procedure P seudoChkGen are the current checking period T (i) check , the Bluetooth MAC address of the slave A slave and the current value of the master's clock t (i) . A node can perform checkpoint generation using the P seudoChkGen procedure at any point in time, it is always guaranteed that the position of checkpoint generated by the two nodes will be the same, as it has been pointed out in Section 5.1. Nevertheless the typical case will be that whenever a node arrives to a checkpoint it generates the position of the next checkpoint on the given link. The variable t (i) check always stores the master's clock at the next checkpoint, thus it needs to be updated every time a checkpoint is passed. Here we note that the Bluetooth clock of a device is a 28 bit counter, where the LSB changes at every half slot. Let us assume that the base period of checkpoints on the i th link of the node is T (i) check = 2 j-2 , j &gt; 2 number of frames, which means that there is one pseudo randomly positioned checkpoint in each consecutive time interval of length T (i) check and the j th bit of the Bluetooth clock changes at every T (i) check . Upon arrival to a checkpoint the variable t (i) check equals to the current value of the master's clock on that link. After the checkpoint generation procedure has been executed the variable t (i) check will store the master's clock at the time of the next checkpoint on that link. Before starting the procedure the variable t (i) check is set to the current value of the master's clock t (i) in order to cover the general case when at the time of generating the next checkpoint the value of t (i) check does not necessarily equals to the current value of the master's clock t (i) . The position of the next checkpoint is obtained such that the node first adds the current value of T (i) check to the variable t (i) check , clears the bits [j - 1, . . . , 0] of t (i) check and then generates the bits [j - 1, . . . , 2] one by one using the procedure P seudoBitGen(X, W ctrl ). When generating the k th bit ( j -1 k 2) the clock bits X = t (i) check [k+1, . . . , k+5] are fed as inputs to the P seudoBitGen procedure, while the control word W ctrl is derived from t (i) check including the bits already generated and from the MAC address of the slave A slave . The schematic view of generating the clock bits of the next checkpoint is illustrated in Figure 12. W 27. X PseudoBitGen ctrl k. k+5. k+1. 28. 0. 1. 2. Figure 12: Generating the clock bits of the next checkpoint 202 The P seudoBitGen procedure is based on the pseudo random scheme used for frequency hop selection in Bluetooth. However , before presenting the P seudoBitGen procedure we give the pseudo-code of the P seudoChkGen procedure. PseudoChkGen procedure: t (i) : the current value of the master's clock; T (i) check = 2 j-2 , j &gt; 2: current length of the base checkperiod in terms of number of frames. t (i) check = t (i) ; t (i) check [j - 1, . . . , 0] = 0; t (i) check = t (i) check + T (i) check ; k = j - 1; while ( k 2) X[0, . . . , 4] = t (i) check [k + 1, . . . , k + 5]; t (i) check [k] = P seudoBitGen(X, W ctrl ); k=k-1; end Finally, we discuss the P seudoBitGen procedure, which is illustrated in Figure 13. 5 A Y 5 X 5 O 1 5 B Z 5 PERM5 5 C 9 D V 5 V[k mod 5] bit selector X O R Add mod 32 Figure 13: The PseudoBitGen procedure The control words of the P seudoBitGen procedure W ctrl = {A, B, C, D} are the same as the control words of the frequency hop selection scheme in Bluetooth and they are shown in Table 2. However, the input X and the additional bit selection operator at the end are different. As it has been discussed above the input X is changing depending on which bit of the checkpoint is going to be generated. When generating the k th clock bit of the next checkpoint the clock bits X = t (i) check [k + 1, . . . , k + 5] are fed as inputs and the bit selection operator at the end selects the (k mod 5) th bit of the 5 bits long output V . A A slave [27 - 23] t (i) check [25 - 21] B B[0 - 3] = A slave [22 - 19], B[4] = 0 C A slave [8, 6, 4, 2, 0] t (i) check [20 - 16] D A slave [18 - 10] t (i) check [15 - 7] Table 2: Control words The operation PERM5 is a butterfly permutation, which is the same as in the frequency hop selection scheme of Bluetooth and it is described in Figure 14. Each bit of the control word P is associated with a given bit exchange in the input word. If the given bit of the control word equals to 1 the corresponding bit exchange is performed otherwise skipped. The control word P is obtained from C and D, such that P [i] = D[i], i = 0 . . . 8 and P [j + 9] = C[j], j = 0 . . . 4. Z[1] Z[2] Z[3] Z[4] Z[0] P[11,10] P[13,12] P[9,8] P[7,6] P[5,4] P[3,2] P[1,0] Figure 14: Butterfly permutation 203

checkpoint;total utilization;piconets;threshold;scatternet;PCSS algorithm;Bluetooth;slaves;inter-piconet communication;scheduling;intensity;Network Access Point;bridging unit

170

Run-Time Dynamic Linking for Reprogramming Wireless Sensor Networks

From experience with wireless sensor networks it has become apparent that dynamic reprogramming of the sensor nodes is a useful feature. The resource constraints in terms of energy, memory, and processing power make sensor network reprogramming a challenging task. Many different mechanisms for reprogramming sensor nodes have been developed ranging from full image replacement to virtual machines. We have implemented an in-situ run-time dynamic linker and loader that use the standard ELF object file format. We show that run-time dynamic linking is an effective method for reprogramming even resource constrained wireless sensor nodes. To evaluate our dynamic linking mechanism we have implemented an application-specific virtual machine and a Java virtual machine and compare the energy cost of the different linking and execution models. We measure the energy consumption and execution time overhead on real hardware to quantify the energy costs for dynamic linking. Our results suggest that while in general the overhead of a virtual machine is high, a combination of native code and virtual machine code provide good energy efficiency. Dynamic run-time linking can be used to update the native code, even in heterogeneous networks.

Introduction Wireless sensor networks consist of a collection of programmable radio-equipped embedded systems. The behavior of a wireless sensor network is encoded in software running on the wireless sensor network nodes. The software in deployed wireless sensor network systems often needs to be changed, both to update the system with new functionality and to correct software bugs. For this reason dynamically reprogramming of wireless sensor network is an important feature. Furthermore, when developing software for wireless sensor networks, being able to update the software of a running sensor network greatly helps to shorten the development time. The limitations of communication bandwidth, the limited energy of the sensor nodes, the limited sensor node memory which typically is on the order of a few thousand bytes large, the absence of memory mapping hardware, and the limited processing power make reprogramming of sensor network nodes challenging. Many different methods for reprogramming sensor nodes have been developed, including full system image replacement [14, 16], approaches based on binary differences [15, 17, 31], virtual machines [18, 19, 20], and loadable native code modules in the first versions of Contiki [5] and SOS [12]. These methods are either inefficient in terms of energy or require non-standard data formats and tools. The primary contribution of this paper is that we investigate the use of standard mechanisms and file formats for reprogramming sensor network nodes. We show that in-situ dynamic run-time linking and loading of native code using the ELF file format, which is a standard feature on many operating systems for PC computers and workstations, is feasible even for resource-constrained sensor nodes. Our secondary contribution is that we measure and quantify the energy costs of dynamic linking and execution of native code and compare it to the energy cost of transmission and execution of code for two virtual machines: an application-specific virtual machine and the Java virtual machine. We have implemented a dynamic linker in the Contiki operating system that can link, relocate, and load standard ELF object code files. Our mechanism is independent of the particular microprocessor architecture on the sensor nodes and we have ported the linker to two different sensor node platforms with only minor modifications to the architecture dependent module of the code. To evaluate the energy costs of the dynamic linker we implement an application specific virtual machine for Contiki together with a compiler for a subset of Java. We also adapt the Java virtual machine from the lejOS system [8] to run under Contiki. We measure the energy cost of reprogramming and executing a set of program using dynamic linking of native code and the two virtual machines. Using the measurements and a simple energy consumption model we calculate break-even points for the energy consumption of the different mechanisms. Our results suggest that while the execution time overhead of a virtual machine is high, a combination of native code and virtual machine code may give good energy efficiency. The remainder of this paper is structured as follows. In Section 2 we discuss different scenarios in which reprogramming is useful. Section 3 presents a set of mechanisms for executing code inside a sensor node and in Section 4 we discuss loadable modules and the process of linking, relocating , and loading native code. Section 5 describes our implementation of dynamic linking and our virtual machines. Our experiments and the results are presented in Section 6 and discuss the results in Section 7. Related work is reviewed in Section 8. Finally, we conclude the paper in Section 9. Scenarios for Software Updates Software updates for sensor networks are necessary for a variety of reasons ranging from implementation and testing of new features of an existing program to complete reprogramming of sensor nodes when installing new applications. In this section we review a set of typical reprogramming scenarios and compare their qualitative properties. 2.1 Software Development Software development is an iterative process where code is written, installed, tested, and debugged in a cyclic fashion . Being able to dynamically reprogram parts of the sensor network system helps shorten the time of the development cycle. During the development cycle developers typically change only one part of the system, possibly only a single algorithm or a function. A sensor network used for software development may therefore see large amounts of small changes to its code. 2.2 Sensor Network Testbeds Sensor network testbeds are an important tool for development and experimentation with sensor network applications . New applications can be tested in a realistic setting and important measurements can be obtained [36]. When a new application is to be tested in a testbed the application typically is installed in the entire network. The application is then run for a specified time, while measurements are collected both from the sensors on the sensor nodes, and from network traffic. For testbeds that are powered from a continuous energy source, the energy consumption of software updates is only of secondary importance. Instead, qualitative properties such as ease of use and flexibility of the software update mechanism are more important. Since the time required to make an update is important, the throughput of a network-wide software update is of importance. As the size of the transmitted binaries impact the throughput, the binary size still can be Update Update Update Program Scenario frequency fraction level longevity Development Often Small All Short Testbeds Seldom Large All Long Bug fixes Seldom Small All Long Reconfig. Seldom Small App Long Dynamic Application Often Small App Long Table 1. Qualitative comparison between different reprogramming scenarios. used as an evaluation metric for systems where throughput is more important than energy consumption. 2.3 Correction of Software Bugs The need for correcting software bugs in sensor networks was early identified [7]. Even after careful testing, new bugs can occur in deployed sensor networks caused by, for example , an unexpected combination of inputs or variable link connectivity that stimulate untested control paths in the communication software [30]. Software bugs can occur at any level of the system. To correct bugs it must therefore be possible to reprogram all parts of the system. 2.4 Application Reconfiguration In an already installed sensor network, the application may need to be reconfigured. This includes change of parameters , or small changes in the application such as changing from absolute temperature readings to notification when thresholds are exceeded [26]. Even though reconfiguration not necessarily include software updates [25], application reconfiguration can be done by reprogramming the application software. Hence software updates can be used in an application reconfiguration scenario. 2.5 Dynamic Applications There are many situations where it is useful to replace the application software of an already deployed sensor network. One example is the forest fire detection scenario presented by Fok et al. [9] where a sensor network is used to detect a fire. When the fire detection application has detected a fire, the fire fighters might want to run a search and rescue application as well as a fire tracking application. While it may possible to host these particular applications on each node despite the limited memory of the sensor nodes, this approach is not scalable [9]. In this scenario, replacing the application on the sensor nodes leads to a more scalable system. 2.6 Summary Table 1 compares the different scenarios and their properties . Update fraction refers to what amount of the system that needs to be updated for every update, update level to at what levels of the system updates are likely to occur, and program longevity to how long an installed program will be expected to reside on the sensor node. Code Execution Models and Reprogramming Many different execution models and environments have been developed or adapted to run on wireless sensor nodes. 16 Some with the notion of facilitating programming [1], others motivated by the potential of saving energy costs for reprogramming enabled by the compact code representation of virtual machines [19]. The choice of the execution model directly impacts the data format and size of the data that needs to be transported to a node. In this section we discuss three different mechanisms for executing program code inside each sensor node: script languages, virtual machines, and native code. 3.1 Script Languages There are many examples of script languages for embedded systems, including BASIC variants, Python interpreters [22], and TCL machines [1]. However, most script interpreters target platforms with much more resources than our target platforms and we have therefore not included them in our comparison. 3.2 Virtual Machines Virtual machines are a common approach to reduce the cost of transmitting program code in situations where the cost of distributing a program is high. Typically, program code for a virtual machine can be made more compact than the program code for the physical machine. For this reason virtual machines are often used for programming sensor networks [18, 19, 20, 23]. While many virtual machines such as the Java virtual machine are generic enough to perform well for a variety of different types of programs, most virtual machines for sensor networks are designed to be highly configurable in order to allow the virtual machine to be tailored for specific applications . In effect, this means that parts of the application code is implemented as virtual machine code running on the virtual machine, and other parts of the application code is implemented in native code that can be used from the programs running on the virtual machine. 3.3 Native Code The most straightforward way to execute code on sensor nodes is by running native code that is executed directly by the microcontroller of the sensor node. Installing new native code on a sensor node is more complex than installing code for a virtual machine because the native code uses physical addresses which typically need to be updated before the program can be executed. In this section we discuss two widely used mechanisms for reprogramming sensor nodes that execute native code: full image replacement and approaches based on binary differences. 3.3.1 Full Image Replacement The most common way to update software in embedded systems and sensor networks is to compile a complete new binary image of the software together with the operating system and overwrite the existing system image of the sensor node. This is the default method used by the XNP and Deluge network reprogramming software in TinyOS [13]. The full image replacement does not require any additional processing of the loaded system image before it is loaded into the system, since the loaded image resides at the same, known, physical memory address as the previous system image. For some systems, such as the Scatterweb system code [33], the system contains both an operating system image and a small set of functions that provide functionality for loading new operating system images. A new operating system image can overwrite the existing image without overwriting the loading functions. The addresses of the loading functions are hard-coded in the operating system image. 3.3.2 Diff-based Approaches Often a small update in the code of the system, such as a bugfix, will cause only minor differences between in the new and old system image. Instead of distributing a new full system image the binary differences, deltas, between the modified and original binary can be distributed. This reduces the amount of data that needs to be transferred. Several types of diff-based approaches have been developed [15, 17, 31] and it has been shown that the size of the deltas produced by the diff-based approaches is very small compared to the full binary image. Loadable Modules A less common alternative to full image replacement and diff-based approaches is to use loadable modules to perform reprogramming. With loadable modules, only parts of the system need to be modified when a single program is changed. Typically, loadable modules require support from the operating system. Contiki and SOS are examples of systems that support loadable modules and TinyOS is an example of an operating system without loadable module support. A loadable module contains the native machine code of the program that is to be loaded into the system. The machine code in the module usually contains references to functions or variables in the system. These references must be resolved to the physical address of the functions or variables before the machine code can be executed. The process of resolving those references is called linking. Linking can be done either when the module is compiled or when the module is loaded. We call the former approach pre-linking and the latter dynamic linking. A pre-linked module contains the absolute physical addresses of the referenced functions or variables whereas a dynamically linked module contains the symbolic names of all system core functions or variables that are referenced in the module. This information increases the size of the dynamically linked module compared to the pre-linked module. The difference is shown in Figure 1. Dynamic linking has not previously been considered for wireless sensor networks because of the perceived run-time overhead , both in terms of execution time, energy consumption, and memory requirements. The machine code in the module usually contains references not only to functions or variables in the system, but also to functions or variables within the module itself. The physical address of those functions will change depending on the memory address at which the module is loaded in the system. The addresses of the references must therefore be updated to the physical address that the function or variable will have when the module is loaded. The process of updating these references is known as relocation. Like linking, relocation can be done either at compile-time or at run-time. When a module has been linked and relocated the program loader loads the module into the system by copying the 17 memcpy /* ... */ } void radio_send() { /* ... */ } 0x0237 0x1720 Core memcpy(); radio_send(); call 0x1720 call 0x0237 Module with dynamic linking information Pre-linked module memcpy(); radio_send(); call 0x0000 call 0x0000 call instruction call instruction radio_send int memcpy() { Figure 1. The difference between a pre-linked module and a module with dynamic linking information: the pre-linked module contains physical addresses whereas the dynamically linked module contains symbolic names. linked and relocated native code into a place in memory from where the program can be executed. 4.1 Pre-linked Modules The machine code of a pre-linked module contains absolute addresses of all functions and variables in the system code that are referenced by the module. Linking of the module is done at compile time and only relocation is performed at run-time. To link a pre-linked module, information about the physical addresses of all functions and variables in the system into which the module is to be loaded must be available at compile time. There are two benefits of pre-linked modules over dynamically linked modules. First, pre-linked modules are smaller than dynamically linked modules which results in less information to be transmitted. Second, the process of loading a pre-linked module into the system is less complex than the process of linking a dynamically linked module. However, the fact that all physical addresses of the system core are hard-coded in the pre-linked module is a severe drawback as a pre-linked module can only be loaded into a system with the exact same physical addresses as the system that was to generate the list of addresses that was used for linking the module. In the original Contiki system [5] we used pre-linked binary modules for dynamic loading. When compiling the Contiki system core, the compiler generated a map file containing the mapping between all globally visible functions and variables in the system core and their addresses. This list of addresses was used to pre-link Contiki modules. We quickly noticed that while pre-linked binary modules worked well for small projects with a homogeneous set of sensor nodes, the system quickly became unmanageable when the number of sensor nodes grew. Even a small change to the system core of one of the sensor nodes would make it impossible to load binary a module into the system bedcase the addresses of variables and functions in the core were different from when the program was linked. We used version numbers to guard against this situation. Version numbers did help against system crashes, but did not solve the general problem: new modules could not be loaded into the system. 4.2 Dynamic Linking With dynamic linking, the object files do not only contain code and data, but also names of functions are variables of the system core that are referenced by the module. The code in the object file cannot be executed before the physical addresses of the referenced variables and functions have been filled in. This process is done at run time by a dynamic linker. In the Contiki dynamic linker we use two file formats for the dynamically linked modules, ELF and Compact ELF. 4.2.1 ELF - Executable and Linkable Format One of the most common object code format for dynamic linking is the Executable and Linkable Format (ELF) [3]. It is a standard format for object files and executables that is used for most modern Unix-like systems. An ELF object file include both program code and data and additional information such as a symbol table, the names of all external unresolved symbols, and relocation tables. The relocation tables are used to locate the program code and data at other places in memory than for which the object code originally was assembled. Additionally, ELF files can hold debugging information such as the line numbers corresponding to specific machine code instructions, and file names of the source files used when producing the ELF object. ELF is also the default object file format produced by the GCC utilities and for this reason there are a number of standard software utilities for manipulating ELF files available. Examples include debuggers, linkers, converters, and programs for calculating program code and data memory sizes. These utilities exist for a wide variety of platforms, including MS Windows, Linux, Solaris, and FreeBSD. This is a clear advantage over other solutions such as FlexCup [27], which require specialized utilities and tools. Our dynamic linker in Contiki understands the ELF format and is able to perform dynamic linking, relocation, and loading of ELF object code files. The debugging features of the ELF format are not used. 4.2.2 CELF - Compact ELF One problem with the ELF format is the overhead in terms of bytes to be transmitted across the network, compared to pre-linked modules. There are a number of reasons for the extra overhead. First, ELF, as any dynamically relocatable file format, includes the symbolic names of all referenced functions or variables that need to be linked at run-time. Second , and more important, the ELF format is designed to work on 32-bit and 64-bit architectures. This causes all ELF data structures to be defined with 32-bit data types. For 8-bit or 16-bit targets the high 16 bits of these fields are unused. To quantify the overhead of the ELF format we devise an alternative to the ELF object code format that we call CELF - Compact ELF. A CELF file contains the same information as an ELF file, but represented with 8 and 16-bit datatypes. 18 CELF files typically are half the size of the corresponding ELF file. The Contiki dynamic loader is able to load CELF files and a utility program is used to convert ELF files to CELF files. It is possible to further compress CELF files using lossless data compression. However, we leave the investigation of the energy-efficiency of this approach to future work. The drawback of the CELF format is that it requires a special compressor utility is for creating the CELF files. This makes the CELF format less attractive for use in many real-world situations. 4.3 Position Independent Code To avoid performing the relocation step when loading a module, it is in some cases possible to compile the module into position independent code. Position independent code is a type of machine code which does not contain any absolute addresses to itself, but only relative references. This is the approach taken by the SOS system. To generate position independent code compiler support is needed. Furthermore, not all CPU architectures support position independent code and even when supported, programs compiled to position independent code typically are subject to size restrictions. For example, the AVR microcontroller supports position independent code but restricts the size of programs to 4 kilobytes. For the MSP430 no compiler is known to fully support position independent code. Implementation We have implemented run-time dynamic linking of ELF and CELF files in the Contiki operating system [5]. To evaluate dynamic linking we have implemented an application specific virtual machine for Contiki together with a compiler for a subset of Java, and have ported a Java virtual machine to Contiki. 5.1 The Contiki Operating System The Contiki operating system was the first operating system for memory-constrained sensor nodes to support dynamic run-time loading of native code modules. Contiki is built around an event-driven kernel and has very low memory requirements. Contiki applications run as extremely lightweight protothreads [6] that provide blocking operations on top of the event-driven kernel at a very small memory cost. Contiki is designed to be highly portable and has been ported to over ten different platforms with different CPU architectures and using different C compilers. Loaded program 00000000000 11111111111 00000000000 00000000000 11111111111 11111111111 00000000000 00000000000 00000000000 11111111111 11111111111 11111111111 00000000000 00000000000 00000000000 11111111111 11111111111 11111111111 RAM Core Loaded program Core ROM Device drivers Contiki kernel Contiki kernel Dynamic linker Symbol table Language run-time Device drivers Figure 2. Partitioning in Contiki: the core and loadable programs in RAM and ROM. A Contiki system is divided into two parts: the core and the loadable programs as shown in Figure 2. The core consists of the Contiki kernel, device drivers, a set of standard applications, parts of the C language library, and a symbol table. Loadable programs are loaded on top of the core and do not modify the core. The core has no information about the loadable programs, except for information that the loadable programs explicitly register with the core. Loadable programs, on the other hand, have full knowledge of the core and may freely call functions and access variables that reside in the core. Loadable programs can call each other by going through the kernel. The kernel dispatches calls from one loaded program to another by looking up the target program in an in-kernel list of active processes. This one-way dependency makes it possible to load and unload programs at run-time without needing to patch the core and without the need for a reboot when a module has been loaded or unloaded. While it is possible to replace the core at run-time by running a special loadable program that overwrites the current core and reboots the system, experience has shown that this feature is not often used in practice. 5.2 The Symbol Table The Contiki core contains a table of the symbolic names of all externally visible variable and function names in the Contiki core and their corresponding addresses. The table includes not only the Contiki system, but also the C language run-time library. The symbol table is used by the dynamic linker when linking loaded programs. The symbol table is created when the Contiki core binary image is compiled. Since the core must contain a correct symbol table, and a correct symbol table cannot be created before the core exists, a three-step process is required to compile a core with a correct symbol table. First, an intermediary core image with an empty symbol table is compiled. From the intermediary core image an intermediary symbol table is created. The intermediary symbol table contains the correct symbols of the final core image, but the addresses of the symbols are incorrect. Second, a second intermediary core image that includes the intermediary symbol table is created. This core image now contains a symbol table of the same size as the one in the final core image so the addresses of all symbols in the core are now as they will be in the final core image. The final symbol table is then created from the second intermediary core image. This symbol table contains both the correct symbols and their correct addresses . Third, the final core image with the correct symbol table is compiled. The process of creating a core image is automated through a simple make script. The symbol table is created using a combination of standard ELF tools. For a typical Contiki system the symbol table contains around 300 entries which amounts to approximately 4 kilobytes of data stored in flash ROM. 5.3 The Dynamic Linker We implemented a dynamic linker for Contiki that is designed to link, relocate, and load either standard ELF files [3] and CELF, Compact ELF, files. The dynamic linker reads 19 ELF/CELF files through the Contiki virtual filesystem interface , CFS, which makes the dynamic linker unaware of the physical location of the ELF/CELF file. Thus the linker can operate on files stored either in RAM, on-chip flash ROM, external EEPROM, or external ROM without modification. Since all file access to the ELF/CELF file is made through the CFS, the dynamic linker does not need to concern itself with low-level filesystem details such as wear-leveling or fragmentation [4] as this is better handled by the CFS. The dynamic linker performs four steps to link, relocate and load an ELF/CELF file. The dynamic linker first parses the ELF/CELF file and extracts relevant information about where in the ELF/CELF file the code, data, symbol table, and relocation entries are stored. Second, memory for the code and data is allocated from flash ROM and RAM, respectively . Third, the code and data segments are linked and relocated to their respective memory locations, and fourth, the code is written to flash ROM and the data to RAM. Currently, memory allocation for the loaded program is done using a simple block allocation scheme. More sophisticated allocation schemes will be investigated in the future. 5.3.1 Linking and Relocating The relocation information in an ELF/CELF file consists of a list of relocation entries. Each relocation entry corresponds to an instruction or address in the code or data in the module that needs to be updated with a new address. A relocation entry contains a pointer to a symbol, such as a variable name or a function name, a pointer to a place in the code or data contained in the ELF/CELF file that needs to be updated with the address of the symbol, and a relocation type which specifies how the data or code should be updated. The relocation types are different depending on the CPU architecture . For the MSP430 there is only one single relocation type, whereas the AVR has 19 different relocation types. The dynamic linker processes a relocation entry at a time. For each relocation entry, its symbol is looked up in the symbol table in the core. If the symbol is found in the core's symbol table, the address of the symbol is used to patch the code or data to which the relocation entry points. The code or data is patched in different ways depending on the relocation type and on the CPU architecture. If the symbol in the relocation entry was not found in the symbol table of the core, the symbol table of the ELF/CELF file itself is searched. If the symbol is found, the address that the symbol will have when the program has been loaded is calculated, and the code or data is patched in the same way as if the symbol was found in the core symbol table. Relocation entries may also be relative to the data, BSS, or code segment in the ELF/CELF file. In that case no symbol is associated with the relocation entry. For such entries the dynamic linker calculates the address that the segment will have when the program has been loaded, and uses that address to patch the code or data. 5.3.2 Loading When the linking and relocating is completed, the text and data have been relocated to their final memory position. The text segment is then written to flash ROM, at the location that was previously allocated. The memory allocated for the data and BSS segments are used as an intermediate storage for transferring text segment data from the ELF/CELF file before it is written to flash ROM. Finally, the memory allocated for the BSS segment is cleared, and the contents of the data segment is copied from the ELF/CELF file. 5.3.3 Executing the Loaded Program When the dynamic linker has successfully loaded the code and data segments, Contiki starts executing the program. The loaded program may replace an already running Contiki service. If the service that is to be replaced needs to pass state to the newly loaded service, Contiki supports the allocation of an external memory buffer for this purpose. However , experience has shown that this mechanism has been very scarcely used in practice and the mechanism is likely to be removed in future versions of Contiki. 5.3.4 Portability Since the ELF/CELF format is the same across different platforms, we designed the Contiki dynamic linker to be easily portable to new platforms. The loader is split into one architecture specific part and one generic part. The generic part parses the ELF/CELF file, finds the relevant sections of the file, looks up symbols from the symbol table, and performs the generic relocation logic. The architecture specific part does only three things: allocates ROM and RAM, writes the linked and relocated binary to flash ROM, and understands the relocation types in order to modify machine code instructions that need adjustment because of relocation. 5.3.5 Alternative Designs The Contiki core symbol table contains all externally visible symbols in the Contiki core. Many of the symbols may never need to be accessed by loadable programs, thus causing ROM overhead. An alternative design would be to let the symbol table include only a handful of symbols, entry points, that define the only ways for an application program to interact with the core. This would lead to a smaller symbol table, but would also require a detailed specification of which entry points that should be included in the symbol table. The main reason why we did not chose this design, however, is that we wish to be able to replace modules at any level of the system . For this reason, we chose to provide the same amount of symbols to an application program as it would have, would it have been compiled directly into the core. However, we are continuing to investigate this alternative design for future versions of the system. 5.4 The Java Virtual Machine We ported the Java virtual machine (JVM) from lejOS [8], a small operating system originally developed for the Lego Mindstorms. The Lego Mindstorms are equipped with an Hitachi H8 microcontroller with 32 kilobytes of RAM available for user programs such as the JVM. The lejOS JVM works within this constrained memory while featuring pre-emptive threads, recursion, synchronization and exceptions. The Contiki port required changes to the RAM-only model of the lejOS JVM. To be able to run Java programs within the 2 kilobytes of RAM available on our hardware platform, Java classes needs to be stored in flash ROM rather than in RAM. The Contiki port stores the class descriptions including bytecode in flash ROM memory. Static class data and class flags that denote if classes have been initialized are stored in RAM 20 as well as object instances and execution stacks. The RAM requirements for the Java part of typical sensor applications are a few hundred bytes. Java programs can call native code methods by declaring native Java methods. The Java virtual machine dispatches calls to native methods to native code. Any native function in Contiki may be called, including services that are part of a loaded Contiki program. 5.5 CVM - the Contiki Virtual Machine We designed the Contiki Virtual Machine, CVS, to be a compromise between an application-specific and a generic virtual machine. CVM can be configured for the application running on top of the machine by allowing functions to be either implemented as native code or as CVM code. To be able to run the same programs for the Java VM and for CVM, we developed a compiler that compiles a subset of the Java language to CVM bytecode. The design of CVM is intentionally similar to other virtual machines, including Mate [19], VM [18], and the Java virtual machine. CVM is a stack-based machine with sepa-rated code and data areas. The CVM instruction set contains integer arithmetic, unconditional and conditional branches, and method invocation instructions. Method invocation can be done in two ways, either by invocation of CVM bytecode functions, or by invocation of functions implemented in native code. Invocation of native functions is done through a special instruction for calling native code. This instruction takes one parameter, which identifies the native function that is to be called. The native function identifiers are defined at compile time by the user that compiles a list of native functions that the CVM program should be able to call. With the native function interface, it is possible for a CVM program to call any native functions provided by the underlying system, including services provided by loadable programs. Native functions in a CVM program are invoked like any other function. The CVM compiler uses the list of native functions to translate calls to such functions into the special instruction for calling native code. Parameters are passed to native functions through the CVM stack. Evaluation To evaluate dynamic linking of native code we compare the energy costs of transferring, linking, relocating, loading, and executing a native code module in ELF format using dynamic linking with the energy costs of transferring, loading, and executing the same program compiled for the CVM and the Java virtual machine. We devise a simple model of the energy consumption of the reprogramming process. Thereafter we experimentally quantify the energy and memory consumption as well as the execution overhead for the reprogramming , the execution methods and the applications. We use the results of the measurements as input into the model which enables us to perform a quantitative comparison of the energy-efficiency of the reprogramming methods. We use the ESB board [33] and the Telos Sky board [29] as our experimental platforms. The ESB is equipped with an MSP430 microcontroller with 2 kilobytes of RAM and 60 kilobytes of flash ROM, an external 64 kilobyte EEPROM, as well as a set of sensors and a TR1001 radio transceiver. PROCESS_THREAD(test_blink, ev, data) { static struct etimer t; PROCESS_BEGIN(); etimer_set(&t, CLOCK_SECOND); while(1) { leds_on(LEDS_GREEN); PROCESS_WAIT_UNTIL(etimer_expired(&t)); etimer_reset(&t); leds_off(LEDS_GREEN); PROCESS_WAIT_UNTIL(etimer_expired(&t)); etimer_reset(&t); } PROCESS_END(); } Figure 3. Example Contiki program that toggles the LEDs every second. The Telos Sky is equipped with an MSP430 microcontroller with 10 kilobytes of RAM and 48 kilobytes of flash ROM together with a CC2420 radio transceiver. We use the ESB to measure the energy of receiving, storing, linking, relocating, loading and executing loadable modules and the Telos Sky to measure the energy of receiving loadable modules. We use three Contiki programs to measure the energy efficiency and execution overhead of our different approaches. Blinker, the first of the two programs, is shown in Figure 3. It is a simple program that toggles the LEDs every second. The second program, Object Tracker, is an object tracking application based on abstract regions [35]. To allow running the programs both as native code, as CVM code, and as Java code we have implemented these programs both in C and Java. A schematic illustration of the C implementation is in Figure 4. To support the object tracker program, we implemented a subset of the abstract regions mechanism in Contiki. The Java and CVM versions of the program call native code versions of the abstract regions functions. The third program is a simple 8 by 8 vector convolution calculation. 6.1 Energy Consumption We model the energy consumption E of the reprogramming process with E = E p + E s + E l + E f where E p is the energy spent in transferring the object over the network, E s the energy cost of storing the object on the device, E l the energy consumed by linking and relocating the object, and E f the required energy for of storing the linked program in flash ROM. We use a simplified model of the network propagation energy where we assume a propagation protocol where the energy consumption E p is proportional to the size of the object to be transferred. Formally, E p = P p s o where s o is the size of the object file to be transfered and P p is a constant scale factor that depends on the network protocol used to transfer the object. We use similar equations for E s (energy for storing the binary) and E l (energy for linking and relocating). The equation for E f (the energy for load-21 PROCESS_THREAD(use_regions_process, ev, data) { PROCESS_BEGIN(); while(1) { value = pir_sensor.value(); region_put(reading_key, value); region_put(reg_x_key, value * loc_x()); region_put(reg_y_key, value * loc_y()); if(value &gt; threshold) { max = region_max(reading_key); if(max == value) { sum = region_sum(reading_key); sum_x = region_sum(reg_x_key); sum_y = region_sum(reg_y_key); centroid_x = sum_x / sum; centroid_y = sum_y / sum; send(centroid_x, centroid_y); } } etimer_set(&t, PERIODIC_DELAY); PROCESS_WAIT_UNTIL(etimer_expired(&t)); } PROCESS_END(); } Figure 4. Schematic implementation of an object tracker based on abstract regions. ing the binary to ROM) contains the size of the compiled code size of the program instead of the size of the object file. This model is intentionally simple and we consider it good enough for our purpose of comparing the energy-efficiency of different reprogramming schemes. 6.1.1 Lower Bounds on Radio Reception Energy We measured the energy consumption of receiving data over the radio for two different radio transceivers: the TR1001 [32], that is used on the ESB board, and the CC2420 [2], that conforms to the IEEE 802.15.4 standard [11] and is used on the Telos Sky board. The TR1001 provides a very low-level interface to the radio medium. The transceiver decodes data at the bit level and transmits the bits in real-time to the CPU. Start bit detection, framing, MAC layer, checksums, and all protocol processing must be done in software running on the CPU. In contrast, the interface provided by the CC2420 is at a higher level. Start bits, framing, and parts of the MAC protocol are handled by the transceiver. The software driver handles incoming and outgoing data on the packet level. Since the TR1001 operates at the bit-level, the communication speed of the TR1001 is determined by the CPU. We use a data rate of 9600 bits per second. The CC2420 has a data rate of 250 kilobits per second, but also incurs some protocol overhead as it provides a more high-level interface. Figure 5 shows the current draw from receiving 1000 bytes of data with the TR1001 and CC2420 radio transceivers. These measurements constitute a lower bound on the energy consumption for receiving data over the radio, as they do not include any control overhead caused by a code propagation protocol. Nor do they include any packet headers . An actual propagation protocol would incur overhead Time Energy Time per Energy per Transceiver (s) (mJ) byte (s) byte (mJ) TR1001 0.83 21 0.0008 0.021 CC2420 0.060 4.8 0.00006 0.0048 Table 2. Lower bounds on the time and energy consumption for receiving 1000 bytes with the TR1001 and CC2420 transceivers. All values are rounded to two significant digits. because of both packet headers and control traffic. For example , the Deluge protocol has a control packet overhead of approximately 20% [14]. This overhead is derived from the total number of control packets and the total number of data packets in a sensor network. The average overhead in terms of number of excessive data packets received is 3.35 [14]. In addition to the actual code propagation protocol overhead, there is also overhead from the MAC layer, both in terms of packet headers and control traffic. The TR1001 provides a low-level interface to the CPU, which enabled us to measure only the current draw of the receiver. We first measured the time required for receiving one byte of data from the radio. To produce the graph in the figure, we measured the current draw of an ESB board which we had programmed to turn on receive mode and busy-wait for the time corresponding to the reception time of 1000 bytes. When measuring the reception current draw of the CC2420, we could not measure the time required for receiving one byte because the CC2420 does not provide an interface at the bit level. Instead, we used two Telos Sky boards and programmed one to continuously send back-to-back packets with 100 bytes of data. We programmed the other board to turn on receive mode when the on-board button was pressed. The receiver would receive 1000 bytes of data, corresponding to 10 packets, before turning the receiver off. We placed the two boards next to each other on a table to avoid packet drops. We produced the graph in Figure 5 by measuring the current draw of the receiver Telos Sky board. To ensure that we did not get spurious packet drops, we repeated the measurement five times without obtaining differing results. Table 2 shows the lower bounds on the time and energy consumption for receiving data with the TR1001 and CC2420 transceivers. The results show that while the current draw of the CC2420 is higher than that of the TR1001, the energy efficiency in terms of energy per byte of the CC2420 is better because of the shorter time required to receive the data. 6.1.2 Energy Consumption of Dynamic Linking To evaluate the energy consumption of dynamic linking, we measure the energy required for the Contiki dynamic linker to link and load two Contiki programs. Normally, Contiki loads programs from the radio network but to avoid measuring any unrelated radio or network effects, we stored the loadable object files in flash ROM before running the experiments. The loadable objects were stored as ELF files from which all debugging information and symbols that were not needed for run-time linking was removed. At boot-up, 22 0 5 10 15 20 0 0.2 0.4 0.6 0.8 1 Current (mA) Time (s) Current 0 5 10 15 20 0 0.2 0.4 0.6 0.8 1 Current (mA) Time (s) Current Figure 5. Current draw for receiving 1000 bytes with the TR1001 and CC2420, respectively. 0 5 10 15 20 0 0.1 0.2 0.3 0.4 0.5 0.6 Current (mA) Time (s) Current Writing to EEPROM Linking and relocating Writing ROM Executing binary to flash Figure 6. Current draw for writing the Blinker ELF file to EEPROM (0 - 0.166 s), linking and relocating the program (0.166 - 0.418 s), writing the resulting code to flash ROM (0.418 - 0.488 s), and executing the binary (0.488 s and onward). The current spikes delimit the three steps and are intentionally caused by blinking on-board LEDs. The high energy consumption when executing the binary is caused by the green LED. one ELF file was copied into an on-board EEPROM from where the Contiki dynamic linker linked and relocated the ELF file before it loaded the program into flash ROM. Figure 6 shows the current draw when loading the Blinker program, and Figure 7 shows the current draw when loading the Object Tracker program. The current spikes seen in both graphs are intentionally caused by blinking the on-board LEDs. The spikes delimit the four different steps that the loader is going through: copying the ELF object file to EEPROM, linking and relocating the object code, copying the linked code to flash ROM, and finally executing the loaded program. The current draw of the green LED is slightly above 8 mA, which causes the high current draw when executing the blinker program (Figure 6). Similarly, when the object tracking application starts, it turns on the radio for neighbor discovery. This causes the current draw to rise to around 6 mA in Figure 7, and matches the radio current measurements in Figure 5. Table 3 shows the energy consumption of loading and linking the Blinker program. The energy was obtained from integration of the curve from Figure 6 and multiplying it by 0 5 10 15 20 0 0.2 0.4 0.6 0.8 1 1.2 Current (mA) Time (s) Current Writing to EEPROM ROM Writing to flash Executing binary Linking and relocating Figure 7. Current draw for writing the Object Tracker ELF file to EEPROM (0 - 0.282 s), linking and relocating the program (0.282 - 0.882 s), writing the resulting code to flash ROM (0.882 - 0.988 s), and executing the binary (0.988 s and onward). The current spikes delimit the three steps and are intentionally caused by blinking on-board LEDs. The high current draw when executing the binary comes from the radio being turned on. the voltage used in our experiments (4.5 V). We see that the linking and relocation step is the most expensive in terms of energy. It is also the longest step. To evaluate the energy overhead of the ELF file format, we compare the energy consumption for receiving four different Contiki programs using the ELF and CELF formats. In addition to the two programs from Figures 3 and 4 we include the code for the Contiki code propagation mechanism and a network publish/subscribe program that performs periodic flooding and converging of information. The two latter programs are significantly larger. We calculate an estimate of the required energy for receiving the files by using the measured energy consumption of the CC2420 radio transceiver and multiply it by the average overhead by the Deluge code propagation protocol, 3.35 [14]. The results are listed in Table 4 and show that radio reception is more energy consuming than linking and loading a program, even for a small program . Furthermore, the results show that the relative average size and energy overhead for ELF files compared to the code and data contained in the files is approximately 4 whereas the relative CELF overhead is just under 2. 23 ELF ELF ELF radio CELF CELF CELF radio Code Data file file size reception file file size reception Program size size size overhead energy (mJ) size overhead energy (mJ) Blinker 130 14 1056 7.3 17 361 2.5 5.9 Object tracker 344 22 1668 5.0 29 758 2.0 12 Code propagator 2184 10 5696 2.6 92 3686 1.7 59 Flood/converge 4298 42 8456 1.9 136 5399 1.2 87 Table 4. The overhead of the ELF and CELF file formats in terms of bytes and estimated reception energy for four Contiki programs. The reception energy is the lower bound of the radio reception energy with the CC2420 chip, multiplied by the average Deluge overhead (3.35). Blinker Energy Obj. Tr. Energy Step time (s) (mJ) time (s) (mJ) Wrt. EEPROM 0.164 1.1 0.282 1.9 Link & reloc 0.252 1.2 0.600 2.9 Wrt. flash ROM 0.070 0.62 0.106 0.76 Total 0.486 2.9 0.988 5.5 Table 3. Measured energy consumption of the storing, linking and loading of the 1056 bytes large Blinker binary and the 1824 bytes large Object Tracker binary. The size of the Blinker code is 130 bytes and the size of the Object Tracker code is 344 bytes. Module ROM RAM Static loader 670 0 Dynamic linker, loader 5694 18 CVM 1344 8 Java VM 13284 59 Table 5. Memory requirements, in bytes. The ROM size for the dynamic linker includes the symbol table. The RAM figures do not include memory for programs running on top of the virtual machines. 6.2 Memory Consumption Memory consumption is an important metric for sensor nodes since memory is a scarce resource on most sensor node platforms. The ESB nodes feature only 2 KB RAM and 60 KB ROM while Mica2 motes provide 128 KB of program memory and 4 KB of RAM. The less memory required for reprogramming, the more is left for applications and support for other important tasks such as security which may also require a large part of the available memory [28]. Table 5 lists the memory requirements of the static linker, the dynamic linker and loader, the CVM and the Java VM. The dynamic linker needs to keep a table of all core symbols in the system. For a complete Contiki system with process management, networking, the dynamic loader, memory allocation , Contiki libraries, and parts of the standard C library, the symbol table requires about 4 kilobytes of ROM. This is included in the ROM size for the dynamic linker. 6.3 Execution Overhead To measure the execution overhead of the application specific virtual machine and the Java virtual machine, we implemented the object tracking program in Figure 4 in C and Java. We compiled the Java code to CVM code and Java bytecode. We ran the compiled code on the MSP430-equipped ESB board. The native C code was compiled Execution type Execution time (ms) Energy (mJ) Native 0.479 0.00054 CVM 0.845 0.00095 Java VM 1.79 0.0020 Table 6. Execution times and energy consumption of one iteration of the tracking program. Execution type Execution time (ms) Energy (mJ) Native 0.67 0.00075 CVM 58.52 0.065 Java VM 65.6 0.073 Table 7. Execution times and energy consumption of the 8 by 8 vector convolution. with the MSP430 port of GCC version 3.2.3. The MSP430 digitally-controlled oscillator was set to clock the CPU at a speed of 2.4576 MHz. We measured the execution time of the three implementations using the on-chip timer A1 that was set to generate a timer interrupt 1000 times per second. The execution times are averaged over 5000 iterations of the object tracking program. The results in Table 6 show the execution time of one run of the object tracking application from Figure 4. The execution time measurements are averaged over 5000 runs of the object tracking program. The energy consumption is calculated by multiplying the execution time with the average energy consumption when a program is running with the radio turned off. The table shows that the overhead of the Java virtual machine is higher than that of the CVM, which is turn is higher than the execution overhead of the native C code. All three implementations of the tracker program use the same abstract regions library which is compiled as native code. Thus much of the execution time in the Java VM and CVM implementations of the object tracking program is spent executing the native code in the abstract regions library. Essentially, the virtual machine simply acts as a dispatcher of calls to various native functions. For programs that spend a significant part of their time executing virtual machine code the relative execution times are significantly higher for the virtual machine programs. To illustrate this, Table 7 lists the execution times of a convolution operation of two vectors of length 8. Convolution is a common operation in digital signal processing where it is used for algorithms such as filtering or edge detection. We see that the execution time of the program running on the virtual machines is close to ten times that of the native program. 24 Dynamic Full image Step linking (mJ) replacement (mJ) Receiving 17 330 Wrt. EEPROM 1.1 22 Link & reloc 1.4 Wrt . flash ROM 0.45 72 Total 20 424 Table 8. Comparison of energy-consumption of reprogramming the blinker application using dynamic linking with an ELF file and full image replacement methods. Step ELF CELF CVM Java Size (bytes) 1824 968 123 1356 Receiving 29 12 2.0 22 Wrt. EEPROM 1.9 0.80 Link & reloc 2.5 2.5 Wrt . flash ROM 1.2 1.2 4 .7 Total 35 16.5 2.0 26.7 Table 9. Comparison of energy-consumption in mJ of reprogramming for the object tracking application using the four different methods. 6.4 Quantitative Comparison Using our model from Section 6.1 and the results from the above measurements, we can calculate approximations of the energy consumption for distribution, reprogramming, and execution of native and virtual machine programs in order to compare the methods with each other. We set P p , the scale factor of the energy consumption for receiving an object file, to the average Deluge overhead of 3.35. 6.4.1 Dynamic Linking vs Full Image Replacement We first compare the energy costs for the two native code reprogramming models: dynamic linking and full image replacement . Table 8 shows the results for the energy consumption of reprogramming the blinker application. The size of blinker application including the operating system is 20 KB which is about 20 times the size of the blinker application itself. Even though no linking needs to be performed during the full image replacement, this method is about 20 times more expensive to perform a whole image replacement compared to a modular update using the dynamic linker. 6.4.2 Dynamic Linking vs Virtual Machines We use the tracking application to compare reprogramming using the Contiki dynamic linker with code updates for the CVM and the Java virtual machine. CVM programs are typically very small and are not stored in EEPROM, nor are they linked or written to flash. Java uncompressed class files are loaded into flash ROM before they are executed. Table 9 shows the sizes of the corresponding binaries and the energy consumption of each reprogramming step. As expected, the process of updating sensor nodes with native code is less energy-efficient than updating with a virtual machine. Also, as shown in Table 6, executing native code is more energy-efficient than executing code for the virtual machines. By combining the results in Table 6 and Table 9, we can compute break-even points for how often we can execute native code as opposed to virtual machine code for the same 0 20 40 60 80 100 120 140 0 20000 40000 60000 80000 100000 Consumed energy (mJ) Number of program iterations Java VM ELF CVM CELF Figure 8. Break-even points for the object tracking program implemented with four different linking and execution methods. 0 20 40 60 80 100 120 140 0 200 400 600 800 1000 Consumed energy (mJ) Number of program iterations Java VM ELF CVM CELF Figure 9. Break-even points for the vector convolution implemented with four different linking and execution methods. energy consumption. That is, after how many program iterations do the cheaper execution costs outweigh the more expensive code updates. Figure 8 shows the modeled energy consumption for executing the Object Tracking program using native code loaded with an ELF object file, native code loaded with an CELF object file, CVM code, and Java code. We see that the Java virtual machine is expensive in terms of energy and will always require more energy than native code loaded with a CELF file. For native code loaded with an ELF file the energy overhead due to receiving the file makes the Java virtual machine more energy efficient until the program is repeated a few thousand times. Due to the small size of the CVM code it is very energy efficient for small numbers of program iterations . It takes about 40000 iterations of the program before the interpretation overhead outweigh the linking and loading overhead of same program running as native code and loaded as a CELF file. If the native program was loaded with an ELF file, however, the CVM program needs to be run approximately 80000 iterations before the energy costs are the same. At the break-even point, the energy consumption is only about one fifth of the energy consumption for loading the blink program using full image replacement as shown in Table 8. In contrast with Figure 8, Figure 9 contains the break-even points from the vector convolution in Table 7. We assume that the convolution algorithm is part of a program with the same size as in Figure 8 so that the energy consumption for reprogramming is the same. In this case the break-even points are drastically lower than in Figure 8. Here the native code loaded with an ELF file outperforms the Java imple-25 mentation already at 100 iterations. The CVM implementation has spent as much energy as the native ELF implementation after 500 iterations. 6.5 Scenario Suitability We can now apply our results to the software update scenarios discussed in Section 2. In a scenario with frequent code updates, such as the dynamic application scenario or during software development, a low loading overhead is to prefer. From Figure 8 we see that both an application-specific virtual machine and a Java machine may be good choices. Depending on the type of application it may be beneficial to decide to run the program on top of a more flexible virtual machine such as the Java machine. The price for such a decision is higher energy overhead. In scenarios where the update frequency is low, e.g. when fixing bugs in installed software or when reconfiguring an installed application, the higher price for dynamic linking may be worth paying. If the program is continuously run for a long time, the energy savings of being able to use native code outweigh the energy cost of the linking process. Furthermore , with a virtual machine it may not be possible to make changes to all levels of the system. For example, a bug in a low-level driver can usually only be fixed by installing new native code. Moreover, programs that are computation-ally heavy benefit from being implemented as native code as native code has lower energy consumption than virtual machine code. The results from Figures 8 and 9 suggest that a combination of virtual machin code and native code can be energy efficient. For many situations this may be a viable alternative to running only native code or only virtual machine code. 6.6 Portability Because of the diversity of sensor network platforms, the Contiki dynamic linker is designed to be portable between different microcontrollers. The dynamic linker is divided into two modules: a generic part that parses and analyzes the ELF/CELF that is to be loaded, and a microcontroller-specific part that allocates memory for the program to be loaded, performs code and data relocation, and writes the linked program into memory. To evaluate the portability of our design we have ported the dynamic linker to two different microcontrollers: the TI MSP430 and the Atmel AVR. The TI MSP430 is used in several sensor network platforms, including the Telos Sky and the ESB. The Atmel AVR is used in the Mica2 motes. Table 10 shows the number of lines of code needed to implement each module. The dramatic difference between the MSP430-specific module and the AVR-specific module is due to the different addressing modes used by the machine code of the two microcontrollers. While the MSP430 has only one addressing mode, the AVR has 19 different addressing modes. Each addressing mode must be handled dif-ferently by the relocation function, which leads to a larger amount of code for the AVR-specific module. Discussion Standard file formats. Our main motivation behind choosing the ELF format for dynamic linking in Contiki was Lines of code, Lines of code, Module total relocation function Generic linker 292 MSP430-specific 45 8 AVR-specific 143 104 Table 10. Number of lines of code for the dynamic linker and the microcontroller-specific parts. that the ELF format is a standard file format. Many compilers and utilities, including all GCC utilities, are able to produce and handle ELF files. Hence no special software is needed to compile and upload new programs into a network of Contiki nodes. In contrast, FlexCup [27] or diff-based approaches require the usage of specially crafted utilities to produce meta data or diff scripts required for uploading software . These special utilities also need to be maintained and ported to the full range of development platforms used for software development for the system. Operating system support. Dynamic linking of ELF files requires support from the underlying operating system and cannot be done on monolithic operating systems such as TinyOS. This is a disadvantage of our approach. For monolithic operating systems, an approach such as FlexCup is better suited. Heterogeneity. With diff-based approaches a binary diff is created either at a base station or by an outside server. The server must have knowledge of the exact software configuration of the sensor nodes on which the diff script is to be run. If sensor nodes are running different versions of their software, diff-based approaches do not scale. Specifically, in many of our development networks we have witnessed a form of micro heterogeneity in the software configuration. Many sensor nodes, which have been running the exact same version of the Contiki operating system , have had small differences in the address of functions and variables in the core. This micro heterogeneity comes from the different core images being compiled by different developers, each having slightly different versions of the C compiler, the C library and the linker utilities. This results in small variations of the operating system image depending on which developer compiled the operating system image. With diff-based approaches micro heterogeneity poses a big problem, as the base station would have to be aware of all the small differences between each node. Combination of native and virtual machine code. Our results suggest that a combination of native and virtual machine code is an energy efficient alternative to pure native code or pure virtual machine code approaches. The dynamic linking mechanism can be used to load the native code that is used by the virtual machine code by the native code interfaces in the virtual machines. Related Work Because of the importance of dynamic reprogramming of wireless sensor networks there has been a lot of effort in the area of software updates for sensor nodes both in the form of system support for software updates and execution environments that directly impact the type and size of updates as well as distribution protocols for software updates. 26 Mainwaring et al. [26] also identified the trade-off between using virtual machine code that is more expensive to run but enables more energy-efficient updates and running native code that executes more efficiently but requires more costly updates. This trade-off has been further discussed by Levis and Culler [19] who implemented the Mate virtual machine designed to both simplify programming and to leverage energy-efficient large-scale software updates in sensor networks. Mate is implemented on top of TinyOS. Levis and Culler later enhanced Mate by application specific virtual machines (ASVMs) [20]. They address the main limitations of Mate: flexibility, concurrency and propagation . Whereas Mate was designed for a single application domain only, ASVM supports a wide range of application domains. Further, instead of relying on broadcasts for code propagation as Mate, ASVM uses the trickle algorithm [21]. The MagnetOS [23] system uses the Java virtual machine to distribute applications across an ad hoc network of laptops. In MagnetOS, Java applications are partitioned into distributed components. The components transparently communicate by raising events. Unlike Mate and Contiki, MagnetOS targets larger platforms than sensor nodes such as PocketPC devices. SensorWare [1] is another script-based proposal for programming nodes that targets larger platforms. VM* is a framework for runtime environments for sensor networks [18]. Using this framework Koshy and Pandey have implemented a subset of the Java Virtual Machine that enables programmers to write applications in Java, and access sensing devices and I/O through native interfaces. Mobile agent-based approaches extend the notion of injected scripts by deploying dynamic, localized and intelligent mobile agents. Using mobile agents, Fok et al. have built the Agilla platform that enables continuous reprogramming by injecting new agents into the network [9]. TinyOS uses a special description language for composing a system of smaller components [10] which are statically linked with the kernel to a complete image of the system. After linking, modifying the system is not possible [19] and hence TinyOS requires the whole image to be updated even for small code changes. Systems that offer loadable modules besides Contiki include SOS [12] and Impala [24]. Impala features an application updater that enables software updates to be performed by linking in updated modules. Updates in Impala are coarse-grained since cross-references between different modules are not possible. Also, the software updater in Impala was only implemented for much more resource-rich hardware than our target devices. The design of SOS [12] is very similar to the Contiki system: SOS consists of a small kernel and dynamically-loaded modules. However, SOS uses position independent code to achieve relocation and jump tables for application programs to access the operating system kernel. Application programs can register function pointers with the operating system for performing inter-process communication. Position independent code is not available for all platforms, however, which limits the ap-plicability of this approach. FlexCup [27] enables run-time installation of software components in TinyOS and thus solves the problem that a full image replacement is required for reprogramming TinyOS applications. In contrast to our ELF-based solution, FlexCup uses a non-standard format and is less portable. Further, FlexCup requires a reboot after a program has been installed, requiring an external mechanism to save and restore the state of all other applications as well as the state of running network protocols across the reboot. Contiki does not need to be rebooted after a program has been installed. FlexCup also requires a complete duplicate image of the binary image of the system to be stored in external flash ROM. The copy of the system image is used for constructing a new system image when a new program has been loaded. In contrast, the Contiki dynamic linker does not alter the core image when programs are loaded and therefore no external copy of the core image is needed. Since the energy consumption of distributing code in sensor networks increases with the size of the code to be distributed several attempts have been made to reduce the size of the code to be distributed. Reijers and Langendoen [31] produce an edit script based on the difference between the modified and original executable. After various optimiza-tions including architecture-dependent ones, the script is distributed . A similar approach has been developed by Jeong and Culler [15] who use the rsync algorithm to generate the difference between modified and original executable. Koshy and Pandey's diff-based approach [17] reduces the amount of flash rewriting by modifying the linking procedure so that functions that are not changed are not shifted. XNP [16] was the previous default reprogramming mechanism in TinyOS which is used by the multi-hop reprogramming scheme MOAP (Multihop Over-the-Air Programming) developed to distribute node images in the sensor network. MOAP distributes data to a selective number of nodes on a neighbourhood-by-neighbourhood basis that avoids flooding [34]. In Trickle [21] virtual machine code is distributed to a network of nodes. While Trickle is restricted to single packet dissemination, Deluge adds support for the dissemination of large data objects [14]. Conclusions We have presented a highly portable dynamic linker and loader that uses the standard ELF file format and compared the energy-efficiency of run-time dynamic linking with an application specific virtual machine and a Java virtual machine . We show that dynamic linking is feasible even for constrained sensor nodes. Our results also suggest that a combination of native and virtual machine code provide an energy efficient alternative to pure native code or pure virtual machine approaches. The native code that is called from the virtual machine code can be updated using the dynamic linker, even in heterogeneous systems. Acknowledgments This work was partly financed by VINNOVA, the Swedish Agency for Innovation Systems, and the European Commission under contract IST-004536-RUNES. Thanks to our paper shepherd Feng Zhao for reading and commenting on the paper. 27 References [1] A. Boulis, C. Han, and M. B. Srivastava. Design and implementation of a framework for efficient and programmable sensor networks. In Proceedings of The First International Conference on Mobile Systems, Applications, and Services (MOBISYS `03), May 2003. [2] Chipcon AS. CC2420 Datasheet (rev. 1.3), 2005. http://www.chipcon.com/ [3] TIS Committee. Tool Interface Standard (TIS) Executable and Linking Format (ELF) Specification Version 1.2, May 1995. [4] H. Dai, M. Neufeld, and R. Han. Elf: an efficient log-structured flash file system for micro sensor nodes. In SenSys, pages 176187, 2004. [5] A. Dunkels, B. Gronvall, and T. Voigt. Contiki - a lightweight and flexible operating system for tiny networked sensors. In Proceedings of the First IEEE Workshop on Embedded Networked Sensors, Tampa, Florida, USA, November 2004. [6] A. Dunkels, O. Schmidt, T. Voigt, and M. Ali. Protothreads: Simplifying event-driven programming of memory-constrained embedded systems. In Proceedings of the 4th International Conference on Embedded Networked Sensor Systems, SenSys 2006, Boulder, Colorado, USA, 2006. [7] D. Estrin (editor). Embedded everywhere: A research agenda for networked systems of embedded computers. National Academy Press, 1st edition, October 2001. ISBN: 0309075688 [8] G. Ferrari, J. Stuber, A. Gombos, and D. Laverde, editors. Programming Lego Mindstorms with Java with CD-ROM. Syngress Publishing, 2002. ISBN: 1928994555 [9] C. Fok, G. Roman, and C. Lu. Rapid development and flexible deploy-ment of adaptive wireless sensor network applications. In Proceedings of the 24th International Conference on Distributed Computing Systems , June 2005. [10] D. Gay, P. Levis, R. von Behren, M. Welsh, E. Brewer, and D. Culler. The nesC language: A holistic approach to networked embedded systems . In Proceedings of the ACM SIGPLAN 2003 conference on Programming language design and implementation, pages 111, 2003. [11] J. A. Gutierrez, M. Naeve, E. Callaway, M. Bourgeois, V. Mitter, and B. Heile. IEEE 802.15.4: A developing standard for low-power low-cost wireless personal area networks. IEEE Network, 15(5):1219, September/October 2001. [12] C. Han, R. K. Rengaswamy, R. Shea, E. Kohler, and M. Srivastava. Sos: A dynamic operating system for sensor networks. In MobiSYS '05: Proceedings of the 3rd international conference on Mobile systems , applications, and services. ACM Press, 2005. [13] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. Culler, and K. Pister. System architecture directions for networked sensors. In Proceedings of the 9th International Conference on Architectural Support for Programming Languages and Operating Systems, November 2000. [14] J. W. Hui and D. Culler. The dynamic behavior of a data dissemination protocol for network programming at scale. In Proc. SenSys'04, Baltimore, Maryland, USA, November 2004. [15] J. Jeong and D. Culler. Incremental network programming for wireless sensors. In Proceedings of the First IEEE Communications Society Conference on Sensor and Ad Hoc Communications and Networks IEEE SECON (2004), October 2004. [16] J. Jeong, S. Kim, and A. Broad. Network reprogramming . TinyOS documentation, 2003. Visited 2006-04-06. http://www.tinyos.net/tinyos-1 .x/doc/NetworkReprogramming.pdf [17] J. Koshy and R. Pandey. Remote incremental linking for energy-efficient reprogramming of sensor networks. In Proceedings of the second European Workshop on Wireless Sensor Networks, 2005. [18] J. Koshy and R. Pandey. Vm*: Synthesizing scalable runtime environments for sensor networks. In Proc. SenSys'05, San Diego, CA, USA, November 2005. [19] P. Levis and D. Culler. Mate: A tiny virtual machine for sensor networks . In Proceedings of ASPLOS-X, San Jose, CA, USA, October 2002. [20] P. Levis, D. Gay, and D Culler. Active sensor networks. In Proc. USENIX/ACM NSDI'05, Boston, MA, USA, May 2005. [21] P. Levis, N. Patel, D. Culler, and S. Shenker. Trickle: A self-regulating algorithm for code propagation and maintenance in wireless sensor networks. In Proc. NSDI'04, March 2004. [22] J. Lilius and I. Paltor. Deeply embedded python, a virtual machine for embedded systems. Web page. 2006-04-06. http://www.tucs.fi/magazin/output.php?ID=2000.N2.LilDeEmPy [23] H. Liu, T. Roeder, K. Walsh, R. Barr, and E. Gun Sirer. Design and implementation of a single system image operating system for ad hoc networks. In MobiSys, pages 149162, 2005. [24] T. Liu, C. Sadler, P. Zhang, and M. Martonosi. Implementing software on resource-constrained mobile sensors: Experiences with Impala and ZebraNet. In Proc. Second Intl. Conference on Mobile Systems, Applications and Services (MOBISYS 2004), June 2004. [25] G. Mainland, L. Kang, S. Lahaie, D. C. Parkes, and M. Welsh. Using virtual markets to program global behavior in sensor networks. In Proceedings of the 2004 SIGOPS European Workshop, Leuven, Belgium, September 2004. [26] A. Mainwaring, J. Polastre, R. Szewczyk, D. Culler, and J. Anderson. Wireless sensor networks for habitat monitoring. In First ACM Workshop on Wireless Sensor Networks and Applications (WSNA 2002), Atlanta, GA, USA, September 2002. [27] P. Jose Marron, M. Gauger, A. Lachenmann, D. Minder, O. Saukh, and K. Rothermel. Flexcup: A flexible and efficient code update mechanism for sensor networks. In European Workshop on Wireless Sensor Networks, 2006. [28] A. Perrig, R. Szewczyk, V. Wen, D. E. Culler, and J. D. Tygar. SPINS: security protocols for sensor netowrks. In Mobile Computing and Networking , pages 189199, 2001. [29] J. Polastre, R. Szewczyk, and D. Culler. Telos: Enabling ultra-low power wireless research. In Proc. IPSN/SPOTS'05, Los Angeles, CA, USA, April 2005. [30] N. Ramanathan, E. Kohler, and D. Estrin. Towards a debugging system for sensor networks. International Journal for Network Management , 3(5), 2005. [31] N. Reijers and K. Langendoen. Efficient code distribution in wireless sensor networks. In Proceedings of the 2nd ACM international conference on Wireless sensor networks and applications, pages 6067, 2003. [32] RF Monolithics. 868.35 MHz Hybrid Transceiver TR1001, 1999. http://www.rfm.com [33] J. Schiller, H. Ritter, A. Liers, and T. Voigt. Scatterweb - low power nodes and energy aware routing. In Proceedings of Hawaii International Conference on System Sciences, Hawaii, USA, 2005. [34] T. Stathopoulos, J. Heidemann, and D. Estrin. A remote code update mechanism for wireless sensor networks. Technical Report CENS-TR -30, University of California, Los Angeles, Center for Embedded Networked Computing, November 2003. [35] M. Welsh and G. Mainland. Programming sensor networks using abstract regions. In Proc. USENIX/ACM NSDI'04, San Francisco, CA,, March 2004. [36] G. Werner-Allen, P. Swieskowski, and M. Welsh. Motelab: A wireless sensor network testbed. In Proc. IPSN/SPOTS'05, Los Angeles, CA, USA, April 2005. 28

Wireless sensor networks;Embedded systems;Dynamic linking;Operating systems;Virtual machines

171

S2DB : A Novel Simulation-Based Debugger for Sensor Network Applications

Sensor network computing can be characterized as resource-constrained distributed computing using unreliable, low bandwidth communication. This combination of characteristics poses significant software development and maintenance challenges. Effective and efficient debugging tools for sensor network are thus critical. Existent development tools, such as TOSSIM, EmStar, ATEMU and Avrora, provide useful debugging support, but not with the fidelity, scale and functionality that we believe are sufficient to meet the needs of the next generation of applications. In this paper, we propose a debugger, called S2DB, based on a distributed full system sensor network simulator with high fidelity and scalable performance, DiSenS. By exploiting the potential of DiSenS as a scalable full system simulator, S2DB extends conventional debugging methods by adding novel device level, program source level, group level, and network level debugging abstractions. The performance evaluation shows that all these debugging features introduce overhead that is generally less than 10% into the simulator and thus making S2DB an efficient and effective debugging tool for sensor networks.

INTRODUCTION Sensor networks, comprised of tiny resource-constrained devices connected by short range radios and powered by batteries, provide an innovative way to implement pervasive and non-intrusive envi-ronmental instrumentation and (potentially) actuation. The resource-constrained nature of sensor network devices poses significant software development and maintenance challenges. To prolong battery life and promote miniaturization, most devices have little memory, use low-power and unreliable radios, and run long duty cycles. In addition to these per-device constraints, by definition sensor networks are also distributed systems, with all of the concomitant synchronization and consistency concerns that distributed coordination implies. For these reasons, effective debugging support is critical. A number of sensor network development systems [2, 18, 3, 17, 13, 6] provide debugging support for individual devices and/or the complete network. However, they all have their limitations. Some rely on hardware support, subject to the same resource constraints that as the programs on which they operate. Some only monitor the network radio traffic. And most importantly, as networks scale, these tools become difficult to apply to the details of collections of interacting sensor nodes. In this paper, we present a new approach that is based on scalable full system sensor network simulation with enhanced debugging features. Our debugging tool is called S 2 DB (where S 2 stands for Simulation and Sensor network). The goal of S 2 DB is to adapt conventional debugging methods to sensor network applications so that we can have better control of hardware details and debug the complete sensor network in a coordinated way. Our approach relies upon four principle innovations in the area of debugging resource constrained devices. At the single device level, we introduce the concept of debugging point a generalized notion of break point, watch point, and state interrogation that permits state display from all sensor device subsystems (flash pages, buffers, etc.); Also at the device level, we introduce virtual registers within the simulator to support source level instrumentation and tracing . The access to these registers does not affect the correct functioning of other components; At the multi-device level, we introduce a coordinated break condition, which enables the coordinated execution control of multiple devices; Finally, at the network level, we provide a "time traveling" facility to use with network level trace analysis, so that error site can be rapidly restored for detailed inspection. S 2 DB is built upon DiSenS [25], a scalable distributed full system sensor network simulator DiSenS has a distributed simulation framework. Individual sensor devices are emulated in separated operating system threads. DiSenS then partitions and schedules these device emulations to the computer nodes of a cluster, and simulates inter-device communication at the radio level (i.e. below the communication protocol stack and radio hardware device interfaces). Sensor device emulations in DiSenS are cycle-accurate. Moreover, a plugin mechanism allows the insertion of power models and radio models with different fidelity levels. Thus DiSenS is capable of accurate , large-scale sensor network simulation where the application and operating system code can be executed, unmodified on native hardware. DiSenS benefits our design and implementation in many aspects. Its simulator infrastructure gives us the full control of device states, which enables the design of debugging points. Its high performance makes our debugger execute efficiently. Its scalability enables us to debug large-scale sensor networks. While the availability of a high-fidelity radio model for sensor network radio remains elusive (making many senor network implementors reluctant to embrace simulation and/or emulation), we believe the ability to debug sensor network programs at scale as a precursor to actual deployment will cut development time and reduce the amount of in situ debugging that will be required in an actual deployment. We also wish to emphasize that in this paper we do not claim S 2 DB adequately addresses many of the thorny difficulties associated with all debugging tools (e.g. the ability to debug optimized code). Rather our focus is on innovations that we believe are important to the development of large-scale senor network deploy-ments and that also improve the current state-of-the-practice in sensor network debugging. In Section 2, we first give the background of sensor network debugging. In Section 3, we briefly introduce the features and details of DiSenS that are relevant to our debugging purpose. In Section 4, we introduce the debugging point and its use with break conditions. We also present the design of virtual hardware based source level instrumentation. In Section 5, we discuss how to control the execution of multiple devices in a coordinated way. We focus on the implementation detail in DiSenS infrastructure. In Section 6, we talk about the checkpoint implementations for fast time traveling. We evaluate the performance of our enhancing techniques in Section 7. And we conclude our work in Section 8. RELATED WORK Like most embedded devices, sensor network devices can be debugged with special hardware support. For motes (e.g. Mica2 and MicaZ), Atmel's AVR JTAG ICE (In-Circuit Emulator) [2] is one of the popular hardware-based debuggers. Atmel's AVR family of microcontrollers (that are currently used as the processing elements in many mote implementations) has built-in debugging support , called On-Chip Debugging (OCD). Developers can access the OCD functions via JTAG [10] hardware interface. With JTAG ICE, developers can set break points, step-execute program and query hardware resources. JTAG ICE can also be used with GUI interfaces or a GDB debugging console. Hardware-based approaches such as JTAG ICE typically have their limitations. For example, it is not possible to synchronize the states of program execution with I/O systems in debugging. This is because when the program execution is stopped in JTAG ICE, the I/O system continues to run at full speed [1]. Also since the debugging support is only provided with the processing unit (i.e. the microcontroller), it is not easy to interrogate the state of other on-board systems, like flash memory. In contrast, by working with the full system DiSenS simulations, S 2 DB does not suffer from these limitations. At network level, many monitoring and visualization tools like Sympathy [18, 19], SpyGlass [3], Surge Network Viewer [22] and Mote-VIEW [16] provide a way to trace, display and analyze network activities for a physical sensor network. These tools usually use a software data collecting module running on sensor nodes in the network. The collected data is transferred using flooding or multihop routing to the gateway node. The gateway node then forwards the data to a PC class machine for analysis or visualization. These tools are useful for displaying the network topology and and analyzing the dynamics of data flow, particularly with respect to specific inter-node communication events. Tools like Sympathy even specialize in detecting and localizing sensor network failures in data collection applications. However, these monitoring may be intrusive in that they share many of the scarce device resources they use with the applications they are intended to instrument. These tools may complement what we have with S 2 DB . When a communication anomaly is detected, for example, often a program-level debugger may still be necessary to pinpoint the exact location of error in code. More generally, while debugging on real hardware is the ultimate way to verify the correctness of sensor network applications , simulation based debuggers provide complementary advantages that have been successfully demonstrated by other projects. Many sensor network simulators, like TOSSIM [13], ATEMU [17], Avrora [23] and EmStar [6], provide significant debugging capabilities . TOSSIM is a discrete event simulator for TinyOS applications . It translates the TinyOS code into emulation code and links with the emulator itself. So debugging with TOSSIM is actually debugging the emulator. Developers have to keep in their mind the internal representation of device states. While discrete event simulators are useful for verifying functional correctness, they typically do not capture the precise timing characteristics of device hardware, and thus have limited capability in exposing errors in program logic. In contrast, full system simulators, such as ATEMU and Avrora, have much higher fidelity. ATEMU features a source level debugger XATDB, which has a graphic frontend for easy use. XATDB can debug multiple sensor devices, but can only focus on one at a time. Avrora provides rich built-in support for profiling and instrumentation. User code can be inserted at any program address , watches can be attached to memory locations, and specific events can be monitored. These facilities can be quite useful for debugging purposes. Indeed, we extend Avrora's probe and watch concepts in the development of S 2 DB's debugging points (cf. Section 4). In addition to this support for simulator instrumentation, S 2 DB also provides a source code level instrumentation facility, via virtual debugging registers, since it is easier to use for some debugging problems. Time traveling for debugging is currently the subject of much research [11, 20] in the field of software system development and virtualization. Flashback [20] is a lightweight extension for rollback and replay for software debugging. Flashback uses shadow processes to take snapshots of the in-memory states of a running process and logs the process' I/O events with the underlying system to support deterministic rollback or replay. VMM (virtual machine monitor) level logging is used in [11] for replaying the system executing in a virtual machine. Checkpointing the state of a full system simulator is easier than that in a real OS or virtual machine monitor since all the hardware are simulated in software. Our results show that time traveling support in DiSenS has very low overhead due to the simpleness of sensor hardware it emulates. THE DiSenS SIMULATOR S 2 DB is built upon DiSenS [25], a distributed sensor network simulator designed for high fidelity and scalable performance. DiSenS provides sensor network applications an execution environment as "close" to real deployment as possible. DiSenS is also able 103 to simulate a sensor network with hundreds of nodes in real time speed using computer clusters. In this section, we briefly introduce the design aspects of DiSenS that are relevant to the implementation of S 2 DB . The complete discussion and evaluation of DiSenS are in papers [25, 24]. 3.1 Full System Device Simulation The building blocks of DiSenS are full system device simulators, supporting popular sensor network devices, including iPAQ [9], Stargate [21] and Mica2/MicaZ motes [15]. In this paper, we confine our description to the functionality necessary for debugging mote applications. However, the same functionality is implemented for more complex devices such as the iPAQ and Stargate. A more full examination of debugging for heterogeneous sensor devices is the subject of our future work. The mote device simulator in DiSenS supports most of the Mica2 and MicaZ hardware features, including the AVR instruction set, the ATmega128L microcontroller (memories, UARTs, timers, SPI and ADC, etc.), the on-board Flash memory, CC1000 (Mica2) and CC2420 (MicaZ) radio chips and other miscellaneous components (like sensor board, LEDs, etc.). The core of the device simulator is a cycle-accurate AVR instruction emulator. The instruction emulator interacts with other hardware simulation components via memory mapped I/O. When an application binary is executed in the simulator, each machine instruction is fed into the instruction emulator, shifting the internal representation of hardware states accordingly and faithfully. Asyn-chronous state change is modelled as events. Events are scheduled by hardware components and kept in an event queue. The instruction emulator checks the event queue for each instruction execution , triggering timed events. The collection of simulated hardware features is rich enough to boot and execute unmodified binaries of TinyOS [8] and most sensor network applications, including Surge, TinyDB [14] and Deluge [4]. By correctly simulating hardware components, the device simulator ensures the cycle accuracy, providing the basis of faithful simulation of a complete sensor network . The full system device simulator in DiSenS also presents extension points or "hooks" for integrating power and radio models. This extensible architecture provides a way to support the development of new models and to trade simulation speed for level of accuracy. For debugging, this extensibility enables developers to test applications with different settings. For example, radio models representing different environments (like outdoor, indoor, etc.) can be plugged in to test applications under different circumstances. In i t s defaul t confi gurat i on, Di S enS i ncorporat es an accurat e power model from [12], a simple linear battery model, a basic lossless radio model, and a simple parameterized statistical model. The structure of the system, however, incorporates these models as modules that can be replaced with more sophisticated counterparts. 3.2 Scalable Distributed Simulation DiSenS's ability to simulate hundreds of mote devices using distributed cluster computing resources is its most distinctive feature. This level of scalability makes it possible to experiment with large sensor network applications before they are actually deployed and to explore reconfiguration options "virtually" so that only the most promising need to be investigated in situ. As a debugging tool, DiSenS's scalability allows developers to identify and correct problems associated with scale. For example, a data sink application may work well in a network of dozens of nodes, but fails when the network size increases to hundreds, due to the problems such as insufficient queue or buffer size. Even for small scale network, the scalability is useful because it translates into simulation speed, and thus debugging efficiency. DiSenS achieves its scalability by using a simple yet effective synchronization protocol for radio simulation and applying automatic node partition algorithms to spread the simulation/emulation workload across machines in a computer cluster. In DiSenS, sensor nodes are simulated in parallel, each running in its own operating system thread and keeping its own virtual clock. Sensor nodes interact with each other only in the radio transmission, during which radio packets are exchanged. The radio interaction of sensor nodes can be abstracted into two operations: read radio channel and write radio channel. The analysis [25] shows that only when a node reads radio channel, it needs to synchronize its clock with its neighbors (i.e., potential radio transmitters in its radio range). This ensures that each receiving node receives all the packets it is supposed to receive. A primitive called wait on sync is introduced to perform this synchronization, which forces the caller to wait for neighbor nodes to catch up with its current clock time. To implement this protocol, each node also has to keep its neighbors updated about its clock advance by periodically sending out its current clock time. A more detailed description and analysis of this protocol is in [25]. To utilize distributed computing r esources, D iS enS partitions nodes into groups, each simulated on one machine within a cluster. Communication between sensor nodes assigned to the same machine is via a shared-memory communication channel. However, when motes assigned to distinct machines communicate, that communication and synchronization must be implemented via a message pass between machines. Due to the relatively large overhead of remote synchronization via message passing (caused by network latency), partitioning of simulated nodes to cluster machines plays an important role in making the ensemble simulation efficient. To address this problem, graph-partitioning algorithms, originally developed for tightly-coupled data-parallel high-performance computing applications, are employed. DiSenS uses a popular partitioning package [7] to partition nodes nearly optimally. Our S 2 DB debugging tool is built upon DiSenS , whose design has huge impact on how the debugging facilities that we have implemented , including both advantages and limitations. In the next 3 sections, we'll discuss how DiSenS interacts with S 2 DB to support both conventional and novel debugging techniques. DEBUGGING INDIVIDUAL DEVICES S 2 DB was first built as a conventional distributed debugger on the DiSenS simulator. Each group of sensor nodes has a standalone debugging proxy waiting for incoming debugging commands. A debugger console thus can attach to each individual sensor node via this group proxy and perform debugging operations. The basic S 2 DB includes most functions in a conventional debugger, like state (register and memory) checking, break points and step execution , etc. In this section, we discuss how we exploit the potential of a simulation environment to devise novel techniques for debugging single sensor devices. 4.1 Debugging Point Debugging is essentially a process of exposing program's internal states relevant to its abnormal behavior and pinpointing the cause. Visibility of execution states is a determining factor of how difficult the debugging task is. Building upon a full system simulator for each device gives S 2 DB a great potential to expose time synchronized state. Conventional debuggers essentially manipulate three states of a program: register, memory and program counter (PC). Simulators 104 Component Parameters Value Interrupt Watchable Overhead PC (pc) microcontroller none Int No Yes Large Register (reg) microcontroller address Int No Yes Large Memory Read (mem rd) SRAM address Boolean No Yes Small Memory Write (mem wr) SRAM address Boolean No Yes Small Memory (mem) SRAM address Int No Yes Small Flash Access (flash access) Flash command, address Boolean No Yes Small Flash (flash) Flash address Int No Yes Small Power Change (power) Power Model none Float No Yes Small Timer Match (timer) Timers none Boolean Yes No Small Radio Data Ready (spi) SPI (radio) none Boolean Yes No Small ADC Data Ready (adc) ADC (radio/sensor) none Boolean Yes No Small Serial Data Received (uart) UART none Boolean Yes No Small Clock (clock) Virtual none Int No Yes Minimal Radio Packet Ready (packet) Radio Chip none Packet No Yes Small Program Defined (custom) Virtual Debugging Hardware ID Int No Yes Program defined Table 1: The current set of debugging points in S 2 DB . can provide much more abundant state information, which may enable or ease certain debugging tasks. For example, to debug a TinyOS module that manages on-board flash memory, it is important for the internal buffers and flash pages to be displayed directly. It is straightforward for DiSenS but rather difficult in a conventional debugger, which has to invoke complex code sequence to access the flash indirectly. We carefully studied the device states in DiSenS and defined a series of debugging points. A debugging point is the access point to one of the internal states of the simulated device. The device state that is exposed by a debugging point can then be used by the debugger for displaying program status and controlling program execution, e.g., break and watch, as that in a conventional debugger . In this sense, debugging points have extended our debugger's capability of program manipulation. Table 1 lists the current set of debugging points defined in S 2 DB. It is not a complete list since we are still improving our implementation and discovering more meaningful debugging points. In the table, the first column shows the debugging point name and the abbreviated notation (in parentheses) used by the debugger console . The corresponding hardware component that a debugging point belongs to is listed in the second column. The third and fourth columns specify the parameters and return value of a debugging point. For example, the "memory" point returns the byte content by the given memory address. The fifth column tells whether a debugging point has an interrupt associated. And the sixth column specifies whether a watch can be added to the point. The last column estimates the theoretical performance overhead of monitoring a particular debugging point. As we see in the table, the common program states interrogated by convent i onal debugger s , i . e . r egi s t e r, memor y and pr ogr am count er, are also generalized as debugging points in S 2 DB , listed as reg, mem and pc. For memory, we also introduced two extra debugging points, mem rd and mem wr, to monitor the access to memory in terms of direction. Notice that debugging points have different time properties: some are persistent while others are transient. In the memory case, the memory content, mem, is persistent, while memory accesses, mem rd and mem wr, are transient. They are valid only when memory is read or written. Similarly, the on-board flash has two defined debugging points: one for the page content (flash) and the other (flash access) for the flash access, including read, program and erase. The power debugging point is used to access the simulated power state of the device, which may be useful for debugging power-aware algorithms. Four important hardware events are defined as debugging points: timer match event (timer), radio (SPI) data ready (spi), ADC data ready (adc) and serial data ready (uart). They are all transient and all related to an interrupt. These debugging points provide a natural and convenient way to debug sensor network programs since many of these programs are event-driven, such as TinyOS and its application suite. As an example, if we want to break the program execution at the occurrence of a timer match event, we can simply invoke the command: &gt; break when timer() == true In a more conventional debugger, a breakpoint is typically set in the interrupt handling code, the name of which must be known to the programmer. Furthermore, breaking on these event-based debugging points is much more efficient than breaking on a source code line (i.e., a specific program address). This is because matching program addresses requires a comparison after the execution of each instruction while matching event-based debugging points only happens when the corresponding hardware events are triggered , which occur much less frequently. We will discuss how to use debugging points to set break conditions and their overhead in later this subsection. The clock debugging point provides a way for accurate timing control over program execution. It can be used to fast forward the execution to a certain point if we know that the bug of our interest will not occur until after a period of time. It would be rather difficult to implement this in a conventional debugger since there is no easy way to obtain accurate clock timing across device subsystems. It is also possible to analyze the states and data in the simulator to extract useful high-level semantics and use them to build advanced debugging points. An example is the recognition of radio packet. The Mica2 sensor device uses the CC1000 radio chip, which operates at the byte level. Thus an emulator can only see the byte stream transmitted from/to neighbor nodes and not packet boundaries. For application debugging, however, it is often necessary to break program execution when a complete packet has been transmitted or received. A typical debugging strategy is to set a breakpoint in the radio software stack at the the line of code line that finishes a packet reception. However, this process can be both tedious and unreliable (e.g. software stack may change when a new image is installed), especially during development or maintenance of the radio stack itself. Fortunately, in the current TinyOS radio stack implementation , the radio packet has a fixed format. We implemented a tiny radio packet recognizer in the radio chip simulation code. A "radio packet ready" (packet) debugging point is defined to signal the state 105 when a complete packet is received. These extracted high-level semantics are useful because we can debug applications without relying on the source code, especially when the application binary is optimized code and it is hard to associate exact program addresses with specific source code line. However, discovering these semantics using low-level data/states is challenging and non-obvious (at least, to us) and as such continues to be a focus of our on-going research in this area. 4.1.1 Break Conditions Using Debugging Points Debugging points are used in a functional form. For example, if we want to print a variable X, we can use: &gt; print mem(X) To implement conditional break or watch points, they can be included in imperatives such as: &gt; break when flash_access(erase, 0x1) which breaks the execution when the first page of the flash is erased. It is also possible to compose them: &gt; break when timer() && mem(Y) &gt; 1 which breaks when a timer match event occurs and a state variable Y , like a counter, is larger than 1. The basic algorithm for monitoring and evaluating break conditions is as follows. Each debugging point maintains a monitor queue. Whenever a break point is set, its condition is added to the queue of every debugging point that is used by the condition. Every time the state changes at a debugging point, the conditions in its queue is re-evaluated to check whether any of them is satisfied. If so, one of the break points is reached and the execution is suspended . Otherwise, the execution continues. Note that the monitoring overhead varies for different debugging points revealing the possibility for optimization of the basic condition evaluating algorithm. The monitoring overhead is determined by the frequency of state change at a debugging point. Obviously, pc has the largest overhead because it changes at each instruction execution. Event related debugging points have very low overhead since hardware events occur less frequently. For example, the timer event may be triggered for every hundreds of cycles. Clock logi-cally has a large overhead since it changes every clock cycle. However , in simulation, clock time is checked anyway for event triggering . By implementing the clock monitoring itself as an event, we introduce no extra overhead for monitoring clock debugging point. Thus we are able to optimize the implementation of condition evaluation. For example, considering the following break condition : &gt; break when pc() == foo && mem(Y) &gt; 1 Using the basic algorithm, the overhead of monitoring the condition is the sum of pc's overhead and mem's overhead. However, since the condition is satisfied when both debugging points match their expression, we could only track mem since it has smaller overhead than pc. When mem is satisfied, we then continue to check pc. In this way, the overall overhead reduces. Now we present the general condition evaluation algorithm. Given a condition as a logic expression, C, it is first converted into canonical form using product of maxterms: C = t 1 t 2 ... t n (1) where t i is a maxterm. The overhead function f ov is defined as the total overhead to monitor all the debugging points in a maxterm. Then we sort the maxterms by the value of f ov (t i ) in incremental order, say, t k 1 , ..., t k n . We start the monitoring of C first using maxterm t k 1 by adding C to all the debugging points that belong to t k 1 . When t k 1 is satisfied, we re-evaluate C and stop if it is true. Otherwise, we remove C from t k 1 's debugging points and start monitoring t k 2 . If t k n is monitored and C is still not satisfied, we loop back to t k 1 . We repeat this process until C is satisfied. If C is unsatisfiable, this process never ends. Debugging points give us powerful capability to debug sensor network programs at a level between the hardware level and the source-code level. However, a direct instrumentation of the source code i s somet i m es easi est and m ost s t r ai ght - f or war d debuggi ng met hod. The typical methodology for implementing source-level instrumentation is to use print statements to dump states. Printing, however, can introduce considerable overhead that can mask the problem being tracked. In S 2 DB we include an instrumentation facility based on virtual registers that serves the same purpose with reduced overhead. We introduce our instrumentation facility in the next subsection. 4.2 Virtual Hardware Based Source Code Instrumentation Sensor devices are usually resource-constrained, lacking the necessary facility for debugging in both hardware and software. On a Mica2 sensor device, the only I/O method that can be used for display internal status by the program is to flash the three LEDs, which is tedious and error-prone to decode. DiSenS faithfully simulates the sensor hardware, thus inheriting this limitation. Because we insist that DiSenS maintain binary transparency with the native hardware it emulates, the simulated sensor network program is not able to perform a simple "printf". To solve this problem, we introduce three virtual registers as an I/O channel for the communication between application and simulator . Their I/O addresses are allocated in the reserved memory space of ATmega128L. Thus the access of these virtual registers will not affect the correct functioning of other components. Table 2 lists the three registers and their functions. Address Name Functionality 0x75 VDBCMD Command Register 0x76 VDBIN Input Register 0x77 VDBOUT Output Register Table 2: Virtual registers for communication between application and simulator. The operation of virtual registers is as follows: an application first issues a command in the command register, VDBCMD; then the output data is transferred via VDBOUT register and the input data is read from em VDBIN register. The simplest application of virtual registers is to print debugging messages by first sending a "PRINT" command and then continuously writing the ASCII characters in a string to the VDBOUT register until a new line is reached. On the simulator side, whenever a command is issued, it either reads from the VDBOUT register or sending data to VDBIN. In the print case, when the simulator gets all the characters (ended by a new line), it will print out on the host console of the simulating machine. A more advanced use of virtual registers is to control a debugging point. We term this combination of virtual registers and debugging points a program defined debugging point (custom, as listed in the last line of Table 1). The state of a custom debugging point is generated by the instrumentation code in the program. To do so, the instrumentation code first sends a "DEBUG" command to the VD-106 BCMD register, then outputs the debugging data on the VDBOUT register, in the form of a tuple, &lt; id, value &gt;. The id is used to identify the instrumentation point in the source code and the value is any value generated by the instrumented code. If there is a break condition registered at this point, it will be checked against the tuple and execution will stop when it is matched. As an example, if we want to break at the 10th entry of a function, we can instrument the function and keep a counter of entries. Every time the counter changes, we output the counter value via virtual registers. The break condition will be satisfied when the value equals to 10. To make it easy to use, we developed a small C library for accessing the virtual registers transparently. Developers can invoke accessing functions on these registers by simply calling the C APIs, for example, in a TinyOS program. Instrumentation via the virtual registers has the minimal intru-siveness on application execution. When generating a debugging point event by sending a &lt; id, value &gt; tuple, only three register accesses are needed if both values in the tuple are 8-bits each (one for the command and two for the data). COORDINATED PARALLEL DEBUGGING OF MULTIPLE DEVICES DiSenS's scalability and performance enables S 2 DB to debug large cooperating ensembles of sensors as a simulated sensor network deployment. Like other debuggers, S 2 DB permits its user to attach to and "focus" on a specific sensor while the other sensors in the ensemble execute independently. However, often, more systematic errors emerge from the interactions among sensor nodes even when individual devices and/or applications are functioning correctly. To reveal these kinds of errors, developers must be able to interrogate and control multiple sensor devices in a coordinated way. Debugging a program normally involves displaying program status , breaking program execution at arbitrary points, step-executing, etc. By extending this concept to parallel debugging, we want to be able to: 1. Display the status of multiple devices in parallel; 2. Break the execution of multiple devices at certain common point; 3. Step-execute multiple devices at the same pace. The first and third items in the above "wish list" are easy to implement in a simulation context. S 2 DB can simply "multicast" its debugging commands to a batch of sensor nodes once their execution stop at a certain common point. As for the second item, since DiSenS is, in effect, executing multiple parallel simulations without a centralized clock, implementing a time-correlated and common breakpoint shares the same coordination challenges with in parallel debugging counterpart. The simplest form of coordinated break is to pause the execution of a set of involved nodes at a specific virtual time, T : &gt; :break when clock() == T where the colon before "break" indicates that it is a batch command and will be sent to all the nodes in a global batch list (maintained by other commands). It is necessary to review DiSenS's synchronization mechanism first. We summarize the major rules as follows: 1. A node that receives or samples radio channel must wait for all its neighbors to catch up with its current clock time; A C B D Y Receive Receive Update Transmit X clock update & byte transmission clock update Figure 1: Illustration of synchronization between sensor nodes in DiSenS . Dashed arrows indicate the update and transmission messages. A &lt; B &lt; C &lt; D. 2. All nodes must periodically broadcast their clock updates to neighbors; 3. Before any wait, a node must first send its clock update (to avoid loop waiting); 4. Radio byte is always sent with a clock update at the end of its last bit transmission. Figure 1 illustrates the process. At time point A, node X receives. It first sends an update of its clock and wait for its neighbor Y (rule 3 & 1). Y runs to B and sends its clock update (rule 2), which wakes up X. X proceeds to C and receives again. Y starts a byte transmission at B. At D, the last bit transmitted and so the byte along with a clock update is sent to X (rule 4). X receives the byte, knowing Y passes its current time, and proceeds. B Y X A Receive C Break point Last update Next update update for break Figure 2: Break at a certain point of time. Dashed arrows indicate the update and transmission messages. B &lt; A &lt; C. Now, let's see what happens when we ask multiple nodes to stop at the same time. Figure 2 shows one case of the situation. X receives at time A and sends an update and waits for Y . Y sends an update at B. Its next update time is C. But we want to break at a point before C but after A. Since Y breaks (thus waits), it sends an update (rule 3). X receives the update, wakes up, proceeds to the break point and stops. Now both X and Y are stopped at the same time point. In Figure 3, the situation is similar to the case in Figure 2. The difference is that now the break point is in the middle of a byte transmission for Y . Y can not just send an update to X and let X proceeds to break point as in Figure 2. because if X gets the update from Y , it believes Y has no byte to send up to the break point and will continue its radio receiving logic. Thus the partial byte from Y is lost. This problem is caused by rule 4. We solve it by relaxing the rule: Whenever a node is stopped (thus it waits) in the middle of a byte transmission, the byte is pre-transmitted with the clock update. We can do this because mote radio always transmits in byte unit. Once a byte transmission starts, we already 107 B Y X A Receive C Break point Transmit start Transmit update & pre-transmit Figure 3: Extension to the synchronization protocol: pre-transmission . Dashed arrows indicate the update and transmission messages. B &lt; A &lt; C. know its content. Also, in DiSenS , each byte received by a node is buffered with timestamp. It will be processed only when the time matches the local clock. With this relaxed rule, we are now able to stop multiple sensor nodes at the same virtual time. The next question is how to perform a conditional break on multiple nodes. Notice that we cannot simply implement: &gt; :break when mem(X) &gt; 3 because it asks the nodes to break independently. Whenever a node breaks at some point, other nodes with direct or indirect neighborhood relationship with it will wait at indeterminate points due to the synchronization requirement. Whether they all satisfy the condition is not clear. A reasonable version of this command is: &gt; :break when *.mem(X) &gt; 3 or &gt; :break when node1.mem(X) &gt; 3 && ... && nodek.mem(X) &gt; 3 which means "break when X &gt; 3 for all the nodes". In the general form, we define a coordinated break as a break with condition cond 1 cond 2 ... cond k , where cond i is a logic expression for node i. 000000000000000 000000000000000 111111111111111 111111111111111 000000000 000000000 111111111 111111111 00000000000000 00000000000000 11111111111111 11111111111111 X Y Z A B C D Condition satisfied Figure 4: Coordinated break. The shaded boxes represent the time range during which a local condition is satisfied. Between C and D, the global condition is satisfied. A &lt; B &lt; C &lt; D. Figure 4 illustrates the meaning of this form of breakpoint. The shaded boxes are the time period during which the local condition for a node is satisfied. In Figure 4, the global condition, i.e. cond x cond y cond z , is satisfied between time C and D. Time C is the exact point where we want to break. Before we present the algorithm that implements coordinated break, we need to first introduce a new synchronization scheme. We call it partially ordered synchronization. By default DiSenS implements peer synchronization: all the nodes are running in arbitrary order except synchronized during receiving or sampling. The new scheme imposes a partial order. In this scheme, a node master is first specified. Then all the other nodes proceed by following the master node. That is, at any wall clock time t wall (i.e., the real world time), for any node i , clock i &lt;= clock master . A C B D A C E B D X Y X Y Figure 5: TOP: peer synchronization in DiSenS . A &lt; B &lt; D &lt; C. BOTTOM: partially ordered synchronization for S 2 DB. A &lt; C &lt; E, B = C, D = E. Dashed arrows indicate the update and transmission messages. (Some update messages are omitted) Figure 5 illustrates the two synchronization schemes. The top part shows DiSenS's peer synchronization scheme. Node X waits at A. Y sends update at B and wakes X. Then Y waits at D, waken by X's update at C. X and Y proceed in parallel afterwards. The bottom part shows S 2 DB's partially ordered synchronization scheme. Here Y is the master. X first waits at A. Y sends its update at B. X receives the update and runs to the updated point, which is C (=B). Then X waits again. When Y runs to D and sends update. X can proceed to E (=D). If Y needs to wait to receive, X will wake it up when X reaches E according to rule 3. Obviously, in this scheme, X always follows Y . Now we can give our algorithm for coordinated break. Using Figure 4 as the example, we first designate X as the master. At point A, X's condition is satisfied. X stops at A. Since Y and Z follow X, they all stop at A. Then we choose the next node as the new master, whose condition is not satisfied yet. It is Y . X and Z follow Y until Y reaches B. Next, similarly, we choose Z as new master. At time C, we find cond x cond y cond z = true. We break the execution and C is exactly our break point. In this algorithm, the aforementioned pre-transmission also plays an important role in that it enables us to stop all nodes at the same time point precisely. Coordinated break, however, does not work with arbitrary conditions . Consider the case where the local conditions in Figure 4 are connected by injunction instead of conjunction. The break point now should be at time A. Since we are not able to predict which node will first satisfy its condition, it is not possible for us to stop all the nodes together at time A unless we synchronize all the nodes cycle by cycle, which would limit the scalability and the performance significantly. For the same reason, we can not set up multiple coordinated break points. We reiterate that these limita-108 tions are a direct result of our desire to scale DiSenS and to use S 2 DB on large-scale simulated networks. That is, we have sacri-ficed generality in favor of the performance gained through parallel and distributed-memory implementation. Although the generality of coordinated break is limited, it is still useful in many situations. For example, for a data sink application, we may want to determine why data is lost when a surge of data flows to the sink node. In this case, we would break the execution of the sink node based on the condition that its neighbor nodes have sent data to it. Then we step-execute the program running on the sink node to determine why the data is being lost. To implement the condition of data sent on neighbor nodes we can simply use source code instrumentation exporting a custom debugging point. Thus this example also illustrates how the single-device debugging features discussed in the previous section can be integrated with the group debugging features. FAST TIME TRAVELING FOR REPLAYABLE DEBUGGING Even with the ability to perform coordinated breakpoints, the normal debugging cycle of break/step/print i s still cumbersome when the complete sensor network is debugged, especially if the size of network is large. The high level nature of some systematic errors requires a global view of the interactions among sensor nodes. An alternative model for debugging sensor networks is: A simulation is conducted with tracing. Trace log is analyzed to pinpoint the anomaly. Quickly return to the point when the anomaly occurs to perform detailed source code level debugging. To achieve this, we need to trace the simulation and restore the state of network at any point in the trace. The debugging points and virtual hardware based instrumentation discussed in section 4 can be used to trace the simulation in a way similar to [23]. In this section, we present the S 2 DB's design of fast time traveling, which enables the restoration of network states. The basic mechanism required to implement time traveling is a periodic checkpoint. A checkpoint of a simulation is a complete copy of the state of the simulated sensor network. DiSenS is an object oriented framework in representing device components. When a checkpoint is initiated, the state saving function is invoked first at the highest level "machine" object. Recursively, the sub-components in the "machine" invoke their own state saving functions . The saved state is comprised of registers, memories (SRAM, EEPROM, etc.) and auxiliary state variables in each component. It also includes some simulation related states. For example, we need to save the event queue content, the received radio byte queue in the radio model and the status of the power model, etc. The complete binary of the state is saved into a timestamped file. The result checkpoint file for DiSenS has a size of 4948 bytes, mostly comprised of SRAM ( 4KB) content. Checkpoint for the on-board flash has to be handled differently. Motes have a 512KB flash chip used for sensor data logging and in-network programming. If flash content is saved as other components , the checkpoint file will be as large as over half megabyte, which is 128 times larger than the one without flash. So if flash is also saved in a snapshot way, it is both extremely space and time inefficient for a large scale sensor network. We solve this problem by saving flash operations in a log file. Since most sensor network applications use flash infrequently and flash content is updated in page unit, the overhead of saving log is much smaller than saving flash snapshots. Notice that the flash buffers have to be saved in the snapshot checkpoint file. Once a simulation is finished, we have a set of snapshot checkpoints and a continuous flash log. Given an arbitrary time point T , to restore the state of system includes the following steps: 1. Restore: find the latest checkpoint CP that is prior to T and load the snapshot checkpoint file; 2. Replay: if flash is used, replay the flash operation log up to CP 's time; 3. Re-run: start from CP , re-run the simulation until time T . Checkpoints can also be initiated by methods other than the need to take a periodic snapshot. For example, under S 2 DB a break point can be associated with a checkpoint so that once the execution breaks, a checkpoint is generated. Thus, a developer can move between the checkpoints to find the exact point when error occurs during a replayed simulation. Checkpoint can also be initiated by a debugging point, especially custom debugging points. By allowing checkpoint to be triggered in conjunction with debugging points S 2 DB integrates the replay and state-saving capabilities needed to efficiently re-examine an error condition with the execution control over state changes. EVALUATION Since S 2 DB is built upon DiSenS , its performance is highly dependent on DiSenS itself. We begin this section by focusing on the the performance of DiSenS simulation/emulation and then show the overhead introduced by various S 2 DB debugging facilities. All experiments described in this section are conducted using a 16-node cluster in which each host has dual 3.2GHz Intel Xeon processors with 1GB memory. The hosts are connected via switched gigabit Ethernet. To make fair comparison, we use the same sensor network application CntToRfm for evaluation. 7.1 Performance of DiSenS For brevity, we present only the typical simulation speed of DiSenS on the cluster. A more thorough examination of scalability and performance under different configurations can be found in paper [25]. Figure 6 shows the performance achieved by DiSenS when simulating various numbers of nodes on the cluster in both 1-D and 2-D topologies. In the figure, the X axis shows the total number of nodes simulated. The Y -axis is the normalized simulation speed (compared to real time speed on hardware). For the 1-D topology, all nodes are oriented on a straight line, 50 meters apart (assuming the maximal radio range is 60 meters). For the 2-D topology, nodes are arranged in a square grid. Again the distance between two nodes is 50 meters. Both performance curves are very close except in the middle part, where 2-D topology has slightly worse performance. The simulation speed drops noticeably from 1 to 4 nodes but then the speed curve keeps flat until 128 nodes are simulated. After that, the speed decreases linearly. The transition from flat to linear decrement is because there is not enough computing resources within the cluster ( 16 hosts). To summarize the results from [25], DiSenS is able to simulate one mote 9 times faster than real time speed, or 160 nodes at near real time speed, or 2048 nodes at nearly a tenth of real time speed. 109 Total number of nodes Normalized simulated clock speed 1 2 4 8 16 32 64 128 256 512 1024 0.01 0.10 1.00 10.00 one dimension two dimensions Figure 6: DiSenS simulation performance in 1-D and 2-D topologies. X-axis is total number of nodes simulated. Y -axis is normalized simulation speed (compared to execution speed on real device). 7.2 Performance of a Break Condition on a Single Device We first evaluate the cost of monitoring debugging points in single-device debugging. Not all the listed (in Table 1) debugging points are evaluated since the overhead for some of them is application dependent. pc memrd memwr power timer spi Debugging Point Relative Simulation Speed 0.0 0.2 0.4 0.6 0.8 1.0 Figure 7: Relative simulation speed for various debugging points. X-axis shows the name of debugging points. Y -axis is the ratio to original simulation speed (without monitoring debugging points). Figure 7 gives the relative simulation speed of evaluating various debugging points. For each one, we set a break condition using the debugging point and run the simulation. The result shows that pc has the largest overhead since the PC change occurs for every instruction execution. Memory related debugging points has less overhead. Power and event-based debugging points have the least overhead since their states change infrequently. 7.3 Performance of a Coordinated Break Condition with Multiple Devices We evaluate the overhead of monitoring the coordinated break condition in this subsection. We run our experiments with a 2-D 4 4 grid of sensor nodes, distributed in 4 groups (hosts). 0.85 0.90 0.95 1.00 Involved Groups Relative Simulation Speed 1 2 3 4 Figure 8: Relative simulation speed of monitoring a coordinated break condition for multiple devices. X-axis is the number of groups (hosts) involved. Y -axis is the ratio to original simulation speed (without condition monitoring). Figure 8 shows the speed ratio between the simulation with monitoring and without. When the group number is 1, only nodes in one group are involved in the break condition. For group number 2, nodes in both groups are used in the break condition, and so on. The speed ratio curve drops when the number of groups increases. The overhead of monitoring coordinated break condition is mostly due to the extra synchronization cost introduced by the new partially ordered synchronization scheme. Obviously, when more nodes (especially remote nodes) involved, the simulation overhead is higher. 7.4 Performance of Checkpointing for Time Traveling We evaluate the overhead of checkpointing in four configurations : 1 1, 4 1, 16 1 and 4 4, where x y means x nodes per group and y groups. For each one, we vary the checkpoint interval from 1/8 up to 4 virtual seconds. Figure 9 shows the relative simulation speed when checkpointing the system periodically. Naturally, the overhead increases when checkpointing more frequently. It is hard to distinguish the single-group curves since their differences are so small. In general, checkpointing in multi-group simulation seems to have larger overhead than single-group. However, the checkpoint overhead is relatively small. All four curves lie above 96% of original simulation speed, which translates to less than 4% of overhead. This result encourages us to use time-traveling extensively in debugging. Developers thus can always return to the last break point or a previous trace point with little cost. To summarize, we find that most of the new debugging facilities we have introduced with S 2 DB have small overhead (less than 10%). As a result, we are able to debug sensor network applications using tools that operate at different levels of abstraction while preserving the high performance and scalability provided by DiSenS . 110 0.96 0.98 1.00 1.02 1.04 Checkpoint Interval in Virtual Time (second) Relative Simulation Speed 1/8 1/4 1/2 1 2 4 1x1 4x1 16x1 4x4 Figure 9: Relative simulation speed for checkpointing. X-axis is the interval between two checkpoints (in terms of virtual clock time of mote device). Y -axis is the ratio to original simulation speed (without checkpointing). CONCLUSION S 2 DB is an efficient and effective sensor network debugger based on DiSenS, a scalable distributed sensor network simulator. S 2 DB makes four innovations to the conventional debugging scheme at different levels of abstraction. For effective debugging of single sensor devices, debugging points are introduced for the interrogation of all interested subsystem states in a sensor device. To facilitate source level tracing and instrumentation, we extend the simulated sensor device hardware with a set of virtual registers providing a way for the communication between simulator and simulated program. At the multi-device level, we discuss the implementation of coordinated break condition in the distributed framework. This new type of break condition enables coordinated parallel execution control of multiple sensor devices. A time traveling facility is introduced for the network level debugging, used for rapid error site restoration when working with sensor network trace analysis . Overall, these debugging features impose overhead of less than 10% (generally) to DiSenS, and thus enable efficient debugging of large scale sensor networks. S 2 DB is still an ongoing project that we think to make it a comprehensive debugging tool for sensor networks, there is still a lot of work to do. The most imperative task is to design and implement a graphic user interface for intuitive and productive debugging. We are planning to build a plugin in the famous Eclipse [5] development environment, which controls the debugging and simulation functions in S 2 DB and DiSenS. We are also interested in incorporating the debugging needs according to people's experiences in sensor network development and discovering new debugging techniques , especially at the network level. REFERENCES [1] Atmel. AVR JTAG ICE User Guide. 2001. http://www.atmel. com/dyn/resources/prod documents/DOC2475.PDF . [2] Atmel's AVR JTAG ICE. http://www.atmel.com/dyn/ products/tools card.asp?tool id=2737 . [3] C. Buschmann, D. Pfisterer, S. Fischer, S. P. Fekete, and A. Kroller. SpyGlass: taking a closer look at sensor networks. In the Proceedings of the 2nd international conference on Embedded networked sensor systems, pages 301302, 2004. New York, NY, USA. [4] A. Chlipala, J. W. Hui, and G. Tolle. Deluge: Dissemination Protocols for Network Reprogramming at Scale. Fall 2003 UC Berkeley class project paper, 2003. [5] Eclipse: an extensible development platform and application frameworks for building software. http://www.eclipse.org. [6] L. Girod, J. Elson, A. Cerpa, T. Stathopoulos, N. Ramanathan, and D. Estrin. EmStar: a Software Environment for Developing and Deploying Wireless Sensor Networks. USENIX Technical Conference, 2004. [7] B. Hendrickson and R. Leland. The Chaco User's Guide: Version 2.0. Technical Report SAND942692, Sandia National Lab, 1994. [8] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. Culler, and K. Pister. System architecture directions for network sensors. International Conference on Architectural Support for Programming Languages and Operating Systems, Oct. 2000. [9] iPAQ devices. http://welcome.hp.com/country/us/en/ prodserv/handheld.html . [10] Boundary-Scan (JTAG) test and in-system programming solutions (IEEE 1149.1). http://www.jtag.com/main.php. [11] S. T. King, G. W. Dunlap, and P. M. Chen. Debugging Operating Systems with Time-Traveling Virtual Machines. In the Proceedings of USENIX Annual Technical Conference 2005, Apr. 2005. Anaheim, CA. [12] O. Landsiedel, K. Wehrle, and S. Gtz. Accurate Prediction of Power Consumption in Sensor Networks. In Proceedings of The Second IEEE Workshop on Embedded Networked Sensors (EmNetS-II), May 2005. Sydney, Australia. [13] P. Levis, N. Lee, M. Welsh, and D. Culler. TOSSIM: Accurate and Scalable Simulation of Entire TinyOS Applications. ACM Conference on Embedded Networked Sensor Systems, Nov. 2003. [14] S. R. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong. The Design of an Acquisitional Query Processor for Sensor Networks. In Proceedings of SIGMOD 2003, June 2003. [15] Mote hardware platform. http://www.tinyos.net/scoop/special/hardware . [16] MOTE-VIEW Monitoring Software. http://www.xbow.com/ Products/productsdetails.aspx?sid=88 . [17] J. Polley, D. Blazakis, J. McGee, D. Rusk, and J. S. Baras. ATEMU: A Fine-grained Sensor Network Simulator. IEEE Communications Society Conference on Sensor and Ad Hoc Communications and Networks, 2004. [18] N. Ramanathan, K. Chang, R. Kapur, L. Girod, E. Kohler, and D. Estrin. Sympathy for the Sensor Network Debugger. In the Proceedings of 3rd ACM Conference on Embedded Networked Sensor Systems (SenSys '05), Nov. 2005. San Diego, California. [19] N. Ramanathan, E. Kohler, and D. Estrin. Towards a debugging system for sensor networks. International Journal of Network Management, 15(4):223234, 2005. [20] S. M. Srinivasan, S. Kandula, C. R. Andrews, and Y. Zhou. Flashback: A Lightweight Extension for Rollback and Deterministic Replay for Software Debugging. In the Proceedings of USENIX Annual Technical Conference 2004, June 2004. Boston, MA. [21] Stargate: a platform X project. http://platformx.sourceforge.net/ . [22] Surge Network Viewer. http://xbow.com/Products/ productsdetails.aspx?sid=86 . [23] B. Titzer and J. Palsberg. Nonintrusive Precision Instrumentation of Microcontroller Software. In the Proceedings of ACM SIGPLAN/SIGBED 2005 Conference on Languages, Compilers, and Tools for Embedded Systems (LCTES'05), June 2005. Chicago, Illinois. [24] Y. Wen, S. Gurun, N. Chohan, R. Wolski, and C. Krintz. SimGate: Full-System, Cycle-Close Simulation of the Stargate Sensor Network Intermediate Node. In Proceedings of International Conference on Embedded Computer Systems: Architectures, MOdeling, and Simulation (IC-SAMOS), 2006. Samos, Greece. [25] Y. Wen, R. Wolski, and G. Moore. DiSenS: Scalable Distributed Sensor Network Simulation. Technical Report CS2005-30, University of California, Santa Barbara, 2005. 111

Sensor Network;Simulation;Debugging

172

Scalable Data Aggregation for Dynamic Events in Sensor Networks

Computing and maintaining network structures for efficient data aggregation incurs high overhead for dynamic events where the set of nodes sensing an event changes with time. Moreover, structured approaches are sensitive to the waiting-time which is used by nodes to wait for packets from their children before forwarding the packet to the sink. Although structure-less approaches can address these issues, the performance does not scale well with the network size. We propose a semi-structured approach that uses a structure-less technique locally followed by Dynamic Forwarding on an implicitly constructed packet forwarding structure to support network scalability. The structure, ToD, is composed of multiple shortest path trees. After performing local aggregation , nodes dynamically decide the forwarding tree based on the location of the sources. The key principle behind ToD is that adjacent nodes in a graph will have low stretch in one of these trees in ToD, thus resulting in early aggregation of packets. Based on simulations on a 2000 nodes network and real experiments on a 105 nodes Mica2-based network, we conclude that efficient aggregation in large scale networks can be achieved by our semi-structured approach.

Introduction Data aggregation is an effective technique for conserving communication energy in sensor networks. In sensor networks, the communication cost is often several orders of magnitude larger than the computation cost. Due to inherent redundancy in raw data collected from sensors, in-network data aggregation can often reduce the communication cost by eliminating redundancy and forwarding only the extracted information from the raw data. As reducing consumption of communication energy extends the network lifetime, it is critical for sensor networks to support in-network data aggregation . Various data aggregation approaches have been proposed for data gathering applications and event-based applications. These approaches make use of cluster based structures [1, 2] or tree based structures [38]. In data gathering applications, such as environment and habitat monitoring [912], nodes periodically report the sensed data to the sink. As the traffic pattern is unchanging, these structure-based approaches incur low maintenance overhead and are therefore suitable for such applications. However, in event-based applications, such as intrusion detection [13, 14] and biological hazard detection [15], the source nodes are not known in advance. Therefore the approaches that use fixed structures can not efficiently aggregate data, while the approaches that change the structure dynamically incur high maintenance overhead [4, 5]. The goal of this paper is to design a scalable and efficient data aggregation protocol that incurs low maintenance overhead and is suited for event-based applications. Constructing an optimal structure for data aggregation for various aggregation functions has been proven to be an NP-hard problem [16, 17]. Although heuristics can be used to construct structures for data aggregation, another problem associated with the convergecast traffic pattern, where nodes transmit their packets to the cluster-head or parent in cluster or tree structures, results in low performance of structure based data aggregation protocols. In [18] the simulation results show that the packet dropping rate in Shortest Path Tree (SPT) is higher because of heavy contention caused by the convergecast traffic. This results in more packet drops and increased delays. As a result, enforcing a fixed order of packet transmissions becomes difficult, which impacts the performance of data aggregation in structured approaches. Typically, packets have to be transmitted in a fixed order 181 from leaves to the root in a tree-like structure to achieve maximum aggregation. Dropped packets not only make the optimal structure sub-optimal, but also waste energy on transmitting packets that are unable to reach the sink. In [19] it shows that the performance gain by using heuristics to create the Steiner Minimum Tree (SMT) for aggregation is not significant compared with using only the Shortest Path Tree (SPT), not to mention that the overhead of constructing such a structure may negate the benefit resulting from data aggregation. However, their conclusions were based on the assumption of randomly located data sources, which is different from the scenarios in event-based sensor networks where a set of close-by nodes is expected to sense an event. Realizing the shortcomings of structured approaches, [20] proposes an anycast based structure-less approach at the MAC layer to aggregate packet. It involves mechanisms to increase the chance of packets meeting at the same node (Spatial Aggregation) at the same time (Temporal Aggregation ). As the approach does not guarantee aggregation of all packets from a single event, the cost of forwarding unaggregated packets increases with the scale of the network and the distance of the event from the sink. To benefit from the strengths of the structured and the structure-less approaches, we propose a semi-structured approach in this paper. The main challenge in designing such a protocol is to determine the packet forwarding strategy in absence of a pre-constructed global structure to achieve early aggregation. Our approach uses a structure-less technique locally followed by Dynamic Forwarding on an implicitly constructed packet forwarding structure to support network scalability . The structure, ToD (Tree on Directed acyclic graph), is composed of multiple shortest path trees. After performing local aggregation, nodes dynamically decide the forwarding tree based on the location of the source nodes. The key principle behind ToD is that adjacent nodes in a graph will have low stretch in at least one of these trees in ToD, thus resulting in early aggregation of packets. This paper makes the following contributions: We propose an efficient and scalable data aggregation mechanism that can achieve early aggregation without incurring any overhead of constructing a structure. We implement the ToD approach on TinyOS and compare its performance against other approaches on a 105 nodes sensor network. For studying the scalability aspects of our approach, we implement ToD in the ns2 simulator and study its performance in networks of up to 2000 nodes. The organization of the rest of the paper is as follows. Section 2 presents background and related work. Section 3 presents the structure-less approach. Section 4 analyzes the performance of ToD in the worst case. The performance evaluation of the protocols using simulations and experiments is presented in Section 5. Finally Section 6 concludes the paper. Related Work Data aggregation has been an active research area in sensor networks for its ability to reduce energy consumption. Some works focus on how to aggregate data from different nodes [2124], some focus on how to construct and maintain a structure to facilitate data aggregation [18,17,2530], and some focus on how to efficiently compress and aggregate data by taking the correlation of data into consideration [17, 3134]. As our work focuses on how to facilitate data aggregation without incurring the overhead of constructing a structure, we briefly describe the structure-based as well as structure-less approaches in current research. In [1,2], the authors propose the LEACH protocol to cluster sensor nodes and let the cluster-heads aggregate data. The cluster-heads then communicate directly with the base station . PEGASIS [26] extends LEACH by organizing all nodes in a chain and letting nodes be the head in turn. [26, 27] extend PEGASIS by allowing simultaneous transmission that balances the energy and delay cost for data gathering. Both LEACH and PEGASIS assume that any node in the network can reach the base-station directly in one-hop, which limits the size of the network for them to be applicable. GIT [3] uses a different approach as compared to LEACH. GIT is built on top of a routing protocol, Directed Diffusion [21,22], which is one of the earliest proposed attribute-based routing protocols. In Directed Diffusion, data can be aggregated opportunistically when they meet at any intermediate node. Based on Directed Diffusion, the Greedy Incremental Tree (GIT) establishes an energy-efficient tree by attaching all sources greedily onto an established energy-efficient path and pruning less energy efficient paths. However due to the overhead of pruning branches, GIT might lead to high cost in moving event scenarios. In [4, 5], the authors propose DCTC, Dynamic Convoy Tree-Based Collaboration, to reduce the overhead of tree migration in mobile event scenarios. DCTC assumes that the distance to the event is known to each sensor and uses the node near the center of the event as the root to construct and maintain the aggregation tree dynamically. However it involves heavy message exchanges which might offset the benefit of aggregation in large-scale networks. From the simulation results in DCTC [5], the energy consumption of tree expansion, pruning and reconfiguration is about 33% of the data collection. In [8], the authors propose an aggregation tree construction algorithm to simultaneously approximate the optimum trees for all non-decreasing and concave aggregation functions . The algorithm uses a simple min-cost perfect matching to construct the tree. [7] also uses similar min-cost matching process to construct an aggregation tree that takes the data fusion cost into consideration. Other works, such as SMT (Steiner Minimum Tree) and MST (Multiple Shared Tree) for multicast algorithms which can be used in data aggregation [17, 19, 30], build a structure in advance for data aggregation . In addition to their complexity and overhead, they are only suitable for networks where the sources are known in advance. Therefore they are not suitable for networks with mobile events. Moreover, fixed tree structure might have long stretch between adjacent nodes. A stretch of two nodes u and v in a tree T on a graph G is the ratio between the distance from node u to v in T and their distance in G. Long stretch 182 implies packets from adjacent nodes have to be forwarded many hops away before they can be aggregated. This problem has been studied as MSST (Minimum Stretch Spanning Tree) [35] and MAST (Minimum Average Stretch Spanning Tree) [36]. They are also NP-hard problems, and it has been shown that for any graph, the lower bound of the average stretch is O (log(n)) [36], and it can be as high as O (n) for the worst case [37]. Even for a grid network, it has been shown that the lower bound for the worst case is O (n) [36]. [38] proposes a polynomial time algorithm to construct a group-independent spanning tree that can achieve O (log(n)) stretch. However the delay in [38] is high in large networks if only nodes near the sink are triggered. [20] is the first proposed structure-less data aggregation protocol that can achieve high aggregation without incurring the overhead of structure approaches. [20] uses anycast to forward packets to one-hop neighbors that have packets for aggregation. It can efficiently aggregate packets near the sources and effectively reduce the number of transmissions. However, it does not guarantee the aggregation of all packets from a single event. As the network grows, the cost of forwarding packets that were unable to be aggregated will negate the benefit of energy saving resulted from eliminating the control overhead. In order to get benefit from structure-less approaches even in large networks, scalability has to be considered in the design of the aggregation protocol. In this paper, we propose a scalable structure-less protocol, ToD, that can achieve efficient aggregation even in large networks. ToD uses a semi-structure approach that does not have the long stretch problem in fixed structure nor incur structure maintenance overhead of dynamic structure, and further improves the performance of the structure-less approach. Scalable Data Aggregation As described before, the goal of our protocol is to achieve aggregation of data near the sources without explicitly constructing a structure for mobile event scenarios. Aggregating packets near the sources is critical for reducing the number of transmissions. Aggregating without using an explicit structure reduces the overhead of construction and maintenance of the structure. In this section, we propose a highly scalable approach that is suitable for very large sensor networks. Our protocol is based on the Data Aware Anycast (DAA) and Randomized Waiting (RW) approaches 1 proposed in [20]. There are two phases in our protocol: DAA and Dynamic Forwarding. In the first phase, packets are forwarded and aggregated to a selected node, termed aggregator, using DAA. In DAA [20], packets were destined to the sink, whereas in our approach they are destined to an aggregator. In the second phase, the leftover un-aggregated or partially aggregated packets are forwarded on a structure, termed Tree on DAG (ToD), for further aggregation. First we briefly describe the DAA protocol proposed in [20]. 3.1 Data Aware Anycast [20] Data Aware Anycast is a structure-less protocol that aggregates packets by improving the Spatial and Temporal con-1 In rest of this paper, we use DAA or Data Aware Anycast to refer to the combination of the two approaches. vergence. Spatial convergence and temporal convergence during transmission are two necessary conditions for aggregation . Packets have to be transmitted to the same node at the same time to be aggregated. Structured approaches achieve these two conditions by letting nodes transmit packets to their parents in the aggregation tree and parents wait for packets from all their children before transmitting the aggregated packets. Without explicit message exchanges in structure-less aggregation, nodes do not know where they should send packets to and how long they should wait for aggregation. Therefore improving spatial or temporal convergence is critical for improving the chance of aggregation. Spatial Convergence is achieved by using anycast to forward packets to nodes that can achieve aggregation. Anycast is a routing scheme whereby packets are forwarded to the best one, or any one, of a group of target destinations based on some routing metrics. By exploiting the nature of wireless radio transmission in sensor networks where all nodes within the transmission range can receive the packet, nodes are able to tell if they can aggregate the transmitting packet, and the anycast mechanism allows the sender to forward packets to any one of them. Transmitting packets to nodes that can achieve aggregation reduces the number of remaining packets in the network, thereby reducing the total number of transmissions. Temporal Convergence is used to further improve the aggregation . Randomized Waiting is a simple technique for achieving temporal convergence, in which nodes wait for a random delay before transmitting. In mobile event triggered networks, nodes are unable to know which nodes are triggered and have packets to transmit in advance. Therefore nodes can not know if they should wait for their upstream nodes and how long they should wait for aggregation. A naive approach of using a fixed delay depending on the distance to the sink may make the detection delay very high. For example, as shown in Fig. 1, nodes closer to the sink must wait longer for packets from possible upstream nodes if fixed waiting time is employed. When events are closer to the sink, the longer delay chosen by nodes closer to the sink is unnecessary. Random delay is used to avoid long delay in large networks while increasing the chance of aggregation. sink ...... ...... = 0 = 1 = n-2 = n-1 = n nodes triggered by an event Figure 1. Longer delay is unnecessary but is inevitable using fixed delay when the event is closer to the sink. Nodes closer to the sink have longer delay ( ) because they have to wait for packets from possible upstream nodes. When a node detects an event and generates a packet for reporting, it picks a random delay between 0 and before transmitting, where is a network parameter that specifies the maximum delay. After delaying the packet, the node 183 broadcasts an RTS packet containing an Aggregation ID. In [20], the timestamp is used as the Aggregation ID, which means that packets generated at the same time can be aggregated . When a node receives an RTS packet, it checks if it has packets with the same Aggregation ID. If it does, it has higher priority for replying with a CTS than nodes that do not have packets for aggregation. The priority is decided by the delay of replying a CTS packet. Nodes with higher priority reply a CTS with shorter delay. If a node overhears any traffic before transmitting its CTS packet, it cancels the CTS transmission in order to avoid collision of multiple CTS responses at the sender. Therefore, nodes can send their packets for aggregation as long as at least one of its neighbors has a packet with the same Aggregation ID. More details and extensions of the DAA approach can be found in [20]. However, DAA can not guarantee that all packets will be aggregated into one packet. When more packets are transmitted from sources to the sink without aggregation, more energy is wasted. This effect becomes more severe when the network is very large and the sources are very far away from the sink. Therefore, instead of forwarding packets directly to the sink when DAA can not aggregate packets any more, we propose the use of Dynamic Forwarding for further packet aggregation. We now describe the Dynamic Forwarding and the construction of ToD. 3.2 Dynamic Forwarding over ToD sink nodes triggered by event B nodes triggered by event A Figure 2. Fixed tree structure for aggregation can have long distance (link-stretch) between adjacent nodes, as in the case of nodes triggered by event B. In this example we assume that nodes in the range of event B are within transmission range of each other. We adopt the pre-constructed structure approach in the second phase to achieve further aggregation. Having a structure to direct all packets to a single node is inevitable if we want to aggregate all packets into one. Constructing a structure dynamically with explicit message exchanges incurs high overhead. Therefore we use an implicitly computed pre-constructed structure that remains unchanged for relatively long time periods (several hours or days). However , using a fixed structure has the long stretch problem as described in Section 2. Take Fig. 2 as an example of pre-computed tree structure where gray nodes are the sources. The fixed tree structure works well if the nodes that generate packets are triggered by event A because their packets can be aggregated immediately on the tree. However, if the nodes that generate packets are triggered by event B, their packets can not be aggregated even if they are adjacent to each other. Therefore we design a dynamic forwarding mechanism over ToD, to avoid the problem of long stretch. 3.2.1 ToD in One Dimensional Networks ...... ........................ ........................ ...... network one row instance of the network sink Figure 3. We illustrate the ToD construction from one row's point of view to simplify the discussion. For illustrating the concept of ToD, we first describe the construction of ToD for a 1-D (a single row of nodes) network , as shown in Fig. 3. We assume that the nodes can communicate with their adjacent nodes in the same row through one hop. We define a cell as a square with side length where is greater than the maximum diameter of the area that an event can span. The network is divided into cells. These cells are grouped into clusters, called F-clusters (First-level clusters). The size of the F-clusters must be large enough to cover the cells an event can span, which is two when we only consider 1-D cells in the network. All nodes in F-clusters send their packets to their cluster-heads, called F-aggregators (First-level aggregators). Note that nodes in the F-cluster can be multiple hops away from the F-aggregator. The formation of the clusters and the election of the aggregators are discussed later in Section 3.2.3. Each F-aggregator then creates a shortest path to the sink. Therefore the structure is a shortest path tree where the root is the sink and the leaves are F-aggregators. We call this tree an F-Tree. Fig. 4(a) shows the construction of the F-Tree. In addition to the F-clusters, we create the second type of clusters, S-clusters (Second-level clusters) for these cells. The size of an S-cluster must also be large enough to cover all cells spanned by an event, and it must interleave with the F-clusters so it can cover adjacent cells in different F-clusters. Each S-cluster also has a cluster-head, S-aggregator , for aggregating packets. Each S-aggregator creates a shortest path to the sink, and forms a second shortest path tree in the network. We call it S-Tree. The illustration of an S-Tree is shown in Fig. 4(b). For all sets of nearby cells that can be triggered by an event, either they will be in the same F-cluster, or they will be in the same S-cluster. This property is exploited by Dynamic Forwarding to avoid the long stretch problem discussed earlier. After the S-Tree is constructed, the F-aggregators connect themselves to the S-aggregators of S-clusters which its F-cluster overlaps with, as shown in Fig. 4(c). For example, in Fig. 4(c), the F-aggregator F4 connects to S-aggregators S3 and S4 because its F-cluster overlaps with S-cluster 3 and 4. Thus, the combination of F-Tree and S-Tree creates a Di-184 A B C D F1 F2 S2 F4 S4 F6 S6 F8 F3 F5 F7 S1 S3 S5 S7 Cells Other nodes in the network F-Aggregators Cells with packets F1 F2 F4 F6 F8 F3 F5 F7 F-Tree S-Tree Overlapping ToD (a) (b) (c) A B C D F-clusters A B C D S-clusters S-Aggregators S2 S4 S6 S1 S3 S5 S7 Figure 4. The construction of F-Tree, S-Tree, and ToD. (a) Leaf nodes are cells. Pairs of neighbor cells define F-clusters. Each F-cluster has an F-aggregator, and F-aggregators form the F-Tree. (b) Each pair of adjacent cells not in the same F-cluster form an S-cluster. Each S-cluster has an S-aggregator, and S-aggregators form the S-Tree. Nodes on the network boundary do not need to be in any S-cluster. (c) Each F-aggregator connects to two S-aggregators of S-clusters which its F-cluster overlaps with. This structure called the Tree on DAG or ToD. F-aggregator in ToD uses Dynamic Forwarding to forward packets to the root, or through an S-aggregator in the S-Tree based on where the packets come from. rected Acyclic Graph, which we refer to as the ToD (Tree on DAG). Nodes first use the Data Aware Anycast (DAA) approach to aggregate as many packets as possible. When no further aggregation can be achieved by DAA, nodes forward their packets to the F-aggregator in its F-cluster. If an event only triggers nodes within a single F-cluster, its packets can be aggregated at the F-aggregator, and be forwarded to the sink using the F-Tree. However, in case the event spans multiple F-clusters, the corresponding packets will be forwarded to different F-aggregators. As we assumed that the event size is not larger than the size of a cell, an event on the boundary of F-clusters will only trigger nodes in cells on the boundary of the F-clusters. By the construction of S-clusters, adjacent cells on the boundary of F-clusters belong to the same S-cluster . Thus, F-aggregators can exploit the information collected from received packets to select the S-aggregator that is best suited for further aggregation. This information is obtained from the source of traffic that can be encoded in the packets. Often such information is readily available in the packet. Otherwise, 4 extra bits can be used to indicate which cell the packet comes from. Consider the example in Fig. 4(c). Since the maximum number of cells an event can span is two, either these two cells are in the same F-cluster, or they are in the same S-cluster . If they are in the same F-cluster, their packets can be aggregated at the F-aggregator. For example, if the event spans A and B, F1 knows that no other F-cluster has packets for aggregation, and it can forward the packets using the F-Tree. If the event spans two cells that are in different F-clusters, the two F-aggregators in the two F-clusters will receive packets only from one of their cells. The F-aggregators then conjecture which F-cluster might also have packets based on which cells the packets come from. For example , if the event spans C and D, F4 will only receive packets from C. Therefore F4 can know either the event happens only in C, or the event spans C and D. Consequently, F4 can forward packets to S4, the S-aggregator of its overlapped S-clusters covering C. Also F5 will forward its packets to S4 if packets only come from D. Therefore these packets can be aggregated at S4. Note that we do not specifically assign cells on the boundary of the network to any S-cluster. They do not need to be in any S-cluster if they are not adjacent to any other F-cluster, or they can be assigned to the same S-cluster as its adjacent cell. The ToD for the one dimensional network has the following property. Property 1. For any two adjacent nodes in ToD in one dimensional network, their packets will be aggregated either at a first level aggregator, or will be aggregated at a second level aggregator. Proof. There are only three possibilities when an event triggers nodes to generate packets. If only nodes in one cell are triggered and generate the packets, their packets can be aggregated at one F-aggregator since all nodes in a cell reside 185 in the same F-cluster, and all packets in an F-cluster will be aggregated at the F-aggregator. If an event triggers nodes in two cells, and these two cells are in the same F-cluster, the packets can be aggregated at the F-aggregator as well. If an event triggers nodes in two cells, but these two cells are in different F-clusters, they must reside in the same S-cluster because S-clusters and F-clusters are interleaved. Moreover, packets in one F-cluster will only originate from the cell that is closer to the other F-cluster that also has packets . Therefore the F-aggregator can forward packets to the S-aggregator for aggregation accordingly, and packets will be aggregated at the S-aggregator. Since the cell is not smaller than the maximum size of an event, it is impossible for an event to trigger more than two cells, and this completes the proof. 3.2.2 ToD in Two Dimensional Networks Section 3.2.1 only demonstrates the construction for one row of nodes to illustrate the basic idea of dynamic forwarding , and it works because each cell is only adjacent to one (or none, if the cell is on the boundary of the network) of the F-clusters . Therefore if an event spans two cells, the two cells are either in the same F-cluster or in the same S-cluster, and the F-aggregator can conjecture whether to forward the packets to the S-aggregator, or to the sink directly. When we consider other cells and F-clusters in the adjacent row, a cell on the boundary of an F-cluster might be adjacent to multiple F-clusters . If an event spans multiple cells, each F-aggregator may have multiple choices of S-aggregators if the cells in their F-cluster are adjacent to multiple F-clusters. If these F-aggregators select different S-aggregators, their packets will not be aggregated. However, the ideas presented in 1D networks can be extended for the 2D networks. But instead of guaranteeing that packets will be aggregated within two steps as in the 1D case (aggregating either at an F-aggregator or an S-aggregator), the ToD in 2D guarantees that the packets can be aggregated within three steps. We first define the cells and clusters in two dimensions. For the ease of understanding, we use grid clustering to illustrate the construction. As defined before, the size of a cell is not less than the maximum size of an event, and an F-cluster must cover all the cells that an event might span, which is four cells in 2D grid-clustering. Therefore the entire network is divided into F-clusters, and each F-cluster contains four cells. The S-clusters have to cover all adjacent cells in different F-clusters. Each F-cluster and S-cluster also has a cluster-head acting as the aggregator to aggregate packets . Fig. 5 shows a 5 5 network with its F-clusters and S-clusters. Since the size of a cell (one side of the square cell) must be greater or equal to the maximum size of an event (diameter of the event), an event can span only one, two, three, or four cells as illustrated in Fig. 6. If the event only spans cells in the same F-cluster, the packets can be aggregated at the F-aggregator. Therefore we only consider scenarios where an event spans cells in multiple F-clusters. (a) F-clusters (c) S-cluters A B C D (b) Cells G H I E F C1 A4 B3 B1 C2 A3 A1 A2 B2 B4 C3 C4 D3 D1 D2 D4 E3 E1 E2 E4 F3 F1 F2 F4 G3 G1 G2 G4 H3 H1 H2 H4 I3 I1 I2 I4 S1 S2 S3 S4 C1 A4 B3 B1 C2 A3 A1 A2 B2 B4 C3 C4 D3 D1 D2 D4 E3 E1 E2 E4 F3 F1 F2 F4 G3 G1 G2 G4 H3 H1 H2 H4 I3 I1 I2 I4 2 2 2 Figure 5. Grid-clustering for a two-dimension network. (a) The network is divided into 5 5 F-clusters. (b) Each F-cluster contains four cells. For example the F-cluster A in (a) contains cell A1, A2, A3, and A4. (c) The S-clusters have to cover all adjacent cells in different F-clusters. Each S-cluster contains four cells from four different F-clusters. Figure 6. The possible numbers of cells an event may span in 2 2 cells, which are one, two, three, and four from left to right. The four cells in each case are any instance of four cells in the network. They may be in the same F-cluster or different F-clusters. Fig. 7 shows four basic scenarios that an F-aggregator may encounter when collecting all packets generated in its F-cluster. All other scenarios are only different combinations of these four scenarios. If packets originate from three or four cells in the same F-cluster, the F-aggregator knows that no other nodes in other F-clusters have packets, and it can forward the packets directly to the sink. If only one or two cells generate packets, it is possible that other F-clusters also have packets. We assume that the region spanned by an event is contiguous. So simultaneous occurrence of scenarios of (a) and (c), or (b) and (d), is impossible in the F-cluster. However, such scenarios are possible in presence of losses in a real environment where packets from third or fourth cluster are lost. In such cases the F-aggregator can just forward the packets directly to the sink because no other F-cluster will have packets from the same event. Figure 7. All possible scenarios in an F-aggregator's point of view. Each case shows 3 3 F-clusters, and the aggregator of the center F-cluster is making the decision. The dark grayed squares are cells that generate packets, and the light grayed squares represent the corresponding S-cluster of the dark grayed cells. When the F-aggregator collects all packets within its cluster , it knows which cells the packets come from and forwards the packets to best suited S-aggregator for further aggregation . For example, if the packets only come from one cell as in case (a) in Fig. 7, the F-aggregator can forward the packet to the S-aggregator of the S-cluster that covers that 186 cell. However, if packets come from two cells in an F-cluster, the two cells must be in different S-clusters. For example, as in Fig. 8 where the F-aggregator of F-cluster X receives packets from two cells, is the combination of case (a) and (b) in Fig. 7. It is possible that the F-aggregator of F-cluster Y may receive packets from cells as in Fig. 7 (c), (d), or both. Since the F-aggregator of F-cluster X does not know which case the F-aggregator of F-cluster Y encounters, it does not know which S-aggregator to forward packets to. To guarantee the aggregation, the F-aggregator of F-cluster X forwards the packet through two S-aggregators that covers cell C1 and C2, therefore packets can meet at least at one S-aggregator. If both F-aggregators receive packets from two cells in its cluster, to guarantee that the packets can meet at least at one S-aggregator, these two F-aggregators must select the S-aggregator deterministically. The strategy is to select the S-aggregator that is closer to the sink. If the packets meet at the first S-aggregator, it does not need to forward packets to the second S-aggregator. The S-aggregator only forwards packets to the second S-aggregator if the packets it received only come from two cells in one F-cluster. We will present a simplified construction later (in Section 3.2.3) for the selection of S-aggregators. F-cluster X F-cluster Y S-cluster I S-cluster II C1 C2 C3 Figure 8. The F-aggregators have two choices for S-aggregators if they receive packets from two cells. To guarantee that the packets can meet at least at one S-aggregator , the second S-aggregator must wait longer than the first S-aggregator. Therefore, if the S-aggregator receives packets from only one cell, it waits longer to wait for possible packets forwarded by the other S-aggregator because it could be the second S-aggregator of the other F-aggregator. Fig. 9 shows an example of one F-aggregator sending packets to the first S-aggregator and then the second S-aggregator, while the other F-aggregator sends packets directly to the second S-aggregator. As long as the second S-aggregator waits suf-ficiently longer than the first S-aggregator the packets can be aggregated at the second S-aggregator. F-aggregators 1 st S-aggregators 2 nd S-aggregators Figure 9. Depending on how many cells generate packets in its F-cluster , one F-aggregator sends packets to two S-aggregators while the other F-aggregator sends packets to only one S-aggregator. We assume that the sink is located at bottom-left of the network. The ToD for the two dimension networks has the following property. Property 2. For any two adjacent nodes in ToD, their packets will be aggregated at the F-aggregator, at the 1 st S-aggregator , or at the 2 nd S-aggregator. Proof. First we define the F-aggregator X as the aggregator of F-cluster X and S-aggregator I as the aggregator of S-cluster I, and so forth. For packets generated only in one F-cluster, their packets can be aggregated at the F-aggregator since all packets in the F-cluster will be sent to the F-aggregator. If an event triggers nodes in different F-clusters, there are only three cases. First, only one cell in each F-cluster generates packets. In this case, all cells having packets will be in the same S-cluster since the adjacent cells in different F-clusters are all in the same S-cluster. Therefore their packets can be aggregated at the S-aggregator. Second, the event spans three cells, C1, C2, and C3, and two of them are in one F-cluster and one of them is in the other F-cluster. Without loss of generality, we assume that C1 and C2 are in the same F-cluster, F-cluster X , and C3 is in the other F-cluster, F-cluster Y . Moreover C3 must be adjacent to either C1 or C2, and let us assume that it is C2. From the ToD construction we know that C2 and C3 will be in the same S-cluster, S-cluster II, and C1 will be in another S-cluster, S-cluster I. Fig. 8 illustrates one instance of this case. First the F-aggregator X will aggregate packets from C1 and C2 because they are in the same F-cluster, and forward the aggregated packets through S-aggregator I to S-aggregator II, or the other way around, because C1 is in S-cluster I and C2 is in S-cluster II. F-aggregator Y will aggregate packets from C3 and forward packets to S-aggregator II because C3 is in S-cluster II. Because packets of F-aggregator Y only come from C3, they will have longer delay in S-aggregator II in order to wait for packets being forwarded through the other S-aggregator. In the mean time, if F-aggregator X forwards packets to S-aggregator II first, the packets can be aggregated at S-aggregator II. If F-aggregator X forwards packets to S-aggregator I first, S-aggregator I will forward packets to S-aggregator II with shorter delay because the packets come from two cells in one F-cluster, therefore their packets can also be aggregated at S-aggregator II. In the third case, the event spans four cells. Two of them will be in one F-cluster and the other two will be in the other F-cluster. Without loss of generality, we can assume that cells C1 and C2 are in F-cluster X and cells C3 and C4 are in F-cluster Y , and C1 and C3 are adjacent, C2 and C4 are adjacent. From the ToD construction, C1 and C3 will be in one S-cluster, S-cluster I, and C2 and C4 will be in the other S-cluster, S-cluster II. Because from S-aggregator I and II, F-aggregator X and Y choose one that is closer to the sink as the first S-aggregator, they will choose the same S-aggregator . Therefore their packets can be aggregated at the first S-aggregator, and this completes the proof. Though in this section we assume that the size of an event is smaller than the size of the cell, our approach can still work 187 correctly and perform more efficiently than DAA even if the size of the event is not known in advance. This is because the nodes will use Dynamic Forwarding over ToD only at second phase where the aggregation by DAA is no longer achievable. Therefore at worst our approach just falls back to DAA. Section 5.1 shows that in experiments, ToD improves the performance of DAA by 27% even if the size of the event is greater than the size of a cell. 3.2.3 Clustering and Aggregator Selection In this paper we use grid-clustering to construct the cells and clusters. Although other clustering methods, such as clustering based on hexagonal or triangular tessellation, can also be used, we do not explore them further in this paper. In principle any clustering would work as long as they satisfy the following conditions. First, the size of the cell must be greater than or equal to the maximum size of an event. Second, the F-cluster and S-cluster must cover the cells that an event may span, and the S-cluster must cover the adjacent cells in different F-clusters. As opposed to defining an arbitrary clustering, using grid-clustering has two advantages. First, the size of the grid can be easily determined by configuring the grid size as a network parameter. Second, as long as the geographic location is known to the node, the cell, F-cluster and S-cluster it belongs to can be determined immediately without any communication . Geographic information is essential in sensor networks, therefore we assume that sensor nodes know their physical location by configuration at deployment, a GPS device , or localization protocols [39, 40]. As a consequence, all the cells, F-clusters, and S-clusters can be implicitly constructed . After the grids are constructed, nodes in an F-cluster and S-cluster have to select an aggregator for their cluster. Because the node that acts as the aggregator consumes more energy than other nodes, nodes should play the role of aggregator in turn in order to evenly distribute the energy consumption among all nodes. Therefore the aggregator selection process must be performed periodically. However the frequency of updating the aggregator can be very low, from once in several hours to once in several days, depending on the capacity of the battery on the nodes. Nodes can elect themselves as the cluster-head with probability based on metrics such as the residual energy, and advertise to all nodes in its cluster. In case two nodes select themselves as the cluster-head, the node-id can be used to break the tie. The other approach is that the nodes in a cluster use a hash function to hash the current time to a node within that cluster, and use that node as the aggregator. Nodes have to know the address of all nodes in its F-cluster and sort them by their node id. A hash function hashes the current time to a number k from 1 to n where n is the number of nodes in its cluster, and nodes use the k th node as the aggregator. Because the frequency of changing the aggregator could be low, the time used could be in hours or days, therefore the time only needs to be coarsely synchronized, and the cluster-head election overhead can be avoided. However, the Dynamic Forwarding approach requires that each F-aggregator knows the location of S-aggregators of S-clusters that its F-cluster overlaps with. Therefore each time the S-aggregator changes, it has to notify the F-aggregators. To simplify the cluster-head selection process and avoid the overhead of propagating the update information, we delegate the role of S-aggregators to F-aggregators. Instead of selecting a node as the S-aggregator and changing it periodically for an S-cluster, we choose an F-cluster, called Aggregating Cluster, for each S-cluster, and use the F-aggregator of the Aggregating Cluster as its S-aggregator. The Aggregating Cluster of an S-cluster is the F-cluster which is closest to the sink among all F-clusters that the S-cluster overlaps with, as shown in Fig. 10(a), assuming that the sink is located on the bottom-left corner. Therefore as the F-aggregator changes, the corresponding S-aggregator changes as well. When an F-aggregator forwards a packet to an S-aggregator, it forwards the packet toward the Aggregating Cluster of that S-aggregator. When the packet reaches the Aggregating Cluster, nodes in that F-cluster know the location of its F-aggregator and can forward the packet to it. Therefore no aggregator update has to be propagated to neighboring clusters . F-cluster S-cluster The common aggregator for both the shaded F-cluster and S-cluster (a) (b) F-aggregator F-aggregator and 1 st S-aggregator 2 nd S-aggregator Figure 10. (a) The S-cluster selects the F-cluster closest to the sink among its overlapped F-clusters, assuming that the sink is located at the bottom-left corner of the network. (b) The white F-aggregator selects the F-cluster containing the gray F-aggregator as the aggregating cluster. Now the role of S-aggregators is passed on to the F-aggregators , and the F-cluster selected by an S-aggregator is the one closer to the sink. When an F-aggregator wants to forward packets to both S-aggregators, it selects the F-cluster that is closer to itself as the aggregating cluster of the first S-aggregator (could be itself) to reduce the number of transmissions between aggregators, as shown in Fig. 10(b). This selection does not affect the property that packets will eventually be aggregated at one aggregator because the S-clusters that cover the cells in two F-clusters are the same, therefore the aggregating cluster selected by two F-aggregators will be the same. The benefits of using this approach are five-fold. First, no leader election is required for S-clusters, which eliminates the leader election overhead. Second, nodes only need to know the F-aggregator of its F-cluster, which make this approach very scalable. Third, when the F-aggregator changes, the S-aggregator changes as well, but the change does not need to be propagated to other F-clusters or S-clusters . Fourth, if nodes choose the aggregator by hashing current time to get a node id of the aggregator in its cluster, 188 only nodes within the same F-cluster need to be synchronized with each other. And last, since the Aggregating Clusters of S-clusters are statically computed, there is no packet overhead for computing the Aggregating Clusters. Performance Analysis In this section we show that the maximum distance between any two adjacent nodes in ToD only depends on the size of the cells, and is independent of the size of the network . We ignore the cost from the aggregator to the sink since for perfect aggregation, only one packet will be forwarded to the sink from the aggregator, therefore the cost is comparatively small. Compared to the lower bound O (n) [36] of the grid network for a fixed tree, ToD can achieve constant factor even in the worst case. u v s f u f v Figure 11. The worst case scenario for ToD. The worst case in ToD is illustrated in Fig. 11 where only two adjacent nodes, u and v, in the corner of two different F-clusters generate packets, and their F-aggregators, f u and f v , are located at the opposite corner. We assume a dense deployment of sensor nodes, therefore the distance between two nodes can be transferred to the cost of transmitting a packet between these nodes. Fig. 11 is the worst case since if more nodes are generating packets in one cluster, it will only amortize the cost of sending packets from the F-aggregator to the S-aggregator, and more nodes in multiple F-clusters generating packets will only lower the average distance. We assume that the length of one side of the cell is , and two nodes are adjacent if their distance is less than a unit of distance. Therefore in Fig. 11 the distance that packets from u and v have to be forwarded before they are aggregated at s is the sum of distances between u to f u to s and v to f v to s, and is (22 + 42) + (22 + 4) = 82 + 4. Therefore in the optimal approach, only one transmission is required because u and v are adjacent. In ToD, 8 2 + 4 number of transmission is required for the worst case. However, since we use DAA as the aggregation technique, packets from adjacent nodes will be aggregated immediately . Therefore for the worst cast to happen, the distance between u and v must be at least 2 units, and our protocol has 4 2+2 7.66 times number of transmissions than optimal. The upper bound is only dependent on the size of a cell, and the size of the cell is dependent on the size of an event. This value is independent of the size of the network and therefore is very suitable for large-scale networks. On average, the number of transmissions will be much less than 4 2 + 2 because first, typically there will be many nodes generating packets. Second, the distance between a node and its F-aggregator is not always 2 2, and the distances between the F-aggregators and the S-aggregator are shorter, too. Third, the DAA approach can efficiently aggregate packets from adjacent nodes thereby further reducing the number of transmissions. Therefore we expect the average distance for nodes generating packets to be much less than the worst case. Performance Evaluation In this section we use experiments and simulations to evaluate the performance of our semi-structured approach and compare it with other protocols. 5.1 Testbed Evaluation We conduct experiments with 105 Mica2-based nodes on a sensor testbed. The testbed consists of 105 Mica2-based motes and each mote is hooked onto a Stargate. The Stargate is a 32-bit hardware device from CrossBow [41] running Linux, which has an Ethernet interface and a serial port for connecting a mote. The Stargates are connected to the server using wired Ethernet. Therefore we can program these motes and send messages and signals to them through Stargates via Ethernet connection. The 105 nodes form a 7 15 grid network with 3 feet spacing. The radio signal using default transmission power covers a lot of nodes in the testbed. In our experiments we do not change the transmission power but limit nodes only to receive packets from nodes within two grid neighbors away, i.e. each node has maximum 12 neighbors. We implement an Anycast MAC protocol on top of the Mica2 MAC layer. The Anycast MAC layer has its own backoff and retransmission mechanisms and we disable the ACK and backoff of the Mica2 MAC module. The Anycast MAC implements the RTS-CTS-DATA-ACK for anycast . An event is emulated by broadcasting a message on the testbed to the Stargates, and the Stargates send the message to the Mica2 nodes through serial port. The message contains a unique ID distinguishing packets generated at different time. When a node is triggered by an event, an event report is generated. If the node has to delay its transmission, it stores the packet in a report queue. Both the application layer and Anycast MAC layer can access the queue, therefore they can check if the node has packets for aggregation, or aggregate the received packets to packets in the queue. First we evaluate the following protocols on the testbed and the codes are available on-line 2 : Dynamic Forwarding over ToD (ToD). The semi-structured approach we proposed in this paper. DAA is used to aggregate packets in each F-cluster, and aggregated packets are forwarded to the sink on ToD. Data Aware Anycast (DAA). The structure-less approach proposed in [20]. Shortest Path Tree (SPT). Nodes send packets to the sink through the shortest path tree immediately after sensing an event. Aggregation is opportunistic and happens only if two packets are at the same node at the same time. The shortest path tree is constructed immediately after the network is deployed. A message is 2 http://www.cse.ohio-state.edu/ fank/research/tod.tar.gz 189 broadcast from the sink and flooded into the network to create a shortest path tree from all nodes to the sink. Shortest Path Tree with Fixed Delay (SPT-D) Same as the SPT approach, but nodes delay their transmission according to their height in the tree to wait for packets from their children. Due to the scale of the testbed, in ToD we only divide the network into two F-clusters, which forces the smallest cell to have only 9 sensor nodes. However we do not limit the size of an event to be smaller than the cell size. The event size is larger than the cell size in all following experiments. We use the normalized number of transmissions as the metric to compare the performance of these protocols. The normalized number of transmissions is the average number of transmissions performed in the entire network to deliver one unit of useful information from sources to the sink. It can be converted to the normalized energy consumption if we know the sending and receiving cost of one transmission, therefore the energy spent on data collection for one packet can be derived. We do not consider energy consumption on idle listening here since all nodes are fully active for all protocols in the experiments and simulations, and the idle energy consumption would be similar for all protocols. To reduce the energy consumption on idle listening, various duty cycling protocols have been proposed. However, due to the page limitation, we are unable to describe how to integrate those works. Fig. 12 shows the normalized number of transmissions for different event sizes. We fixed the location of the event and vary its diameter from 12 ft to 36 ft where nodes within two grid-hops to six grid-hops of the event will be triggered respectively and send packets to the sink located at one corner of the network. We use 6 seconds as maximum delay for all protocols except SPT. For event size less than 12 ft, there are too little nodes been triggered (less than five), and all triggered nodes are within transmission range. Data aggregation is not so interesting in such scenario therefore we do not evaluate it. Actually DAA can perform best since all packets can be aggregated because all triggered nodes are within transmission range of each other. All protocols have better performance when the size of the event increases because packets have more chances of being aggregated. ToD performs best among all protocols in all scenarios. This shows that DAA can efficiently achieve early aggregation and the Dynamic Forwarding over ToD can effectively reduce the cost of directly forwarding unaggregated packets to the sink in DAA. In SPT-D, when the event size is smaller, the long stretch effect is more significant than in larger event scenario. When event size is large, for example , two-third of nodes in the network are triggered when the diameter of the event is 36 feet, most of the packets can be aggregated to their parent with one transmission. This indicates that in applications where most nodes are transmitting, the fixed structure such as SPT-D is better, but when only a small subset of nodes are transmitting, their performance degrades because of the long stretch problem. We notice that the variance of some results in SPT and SPT-D is very high. For example, when the event size is 12 feet in diameter, the maximum normalized number of trans-0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 36 30 24 18 12 Normalized Number of Transmissions Event Size (ft) ToD DAA SPT SPT-D Figure 12. The normalized number of transmissions for different event sizes from experiments on 105 sensors. missions in SPT-D is 3 .41, and the minimum value is 2.41. By tracing into the detail experiment logs we found that the high variance is because of the different shortest path trees. The tree is re-constructed for each experiment, and therefore may change from experiment to experiment. We found that SPT-D always gets better performance in one tree where all sources are under the same subtree, and performs badly in the other tree where sources are located under two or three different subtrees. This further supports our claims that the long stretch problem in fixed structured approaches affects their performance significantly. The second experiment evaluates the performance of these protocols for different values of maximum delay. We vary the delay from 0 to 8 seconds, and all nodes in the network generate one packet every 10 seconds. Fig. 13 shows the results. As we described, the performance of the structure-based approaches heavily depends on the delay . The SPT-D performs worse than ToD when the maximum delay is less than five seconds, and the performance increases as the delay increases. On the contrary, the performance of ToD and DAA does not change for different delays, which is different from results observed in [20]. We believe that this is because with the default transmission power, a large number of nodes are in interference range when nodes transmit. Therefore even if nodes do not delay their transmissions , only one node can transmit at any given time. Other nodes will be forced to delay, which has the same effect as the Randomized Waiting. 5.2 Large Scale Simulation To evaluate and compare the performance and scalability of ToD with other approaches requires a large sensor network, which is currently unavailable in real experiments. Therefore we resort to simulations. In this section we use the ns2 network simulator to evaluate these protocols. Besides ToD, DAA, and SPT, we evaluate OPT, Optimal Aggregation Tree, to replace the SPT-D protocol. In OPT, nodes forward their packets on an aggregation tree rooted at the center of the event. Nodes know where to forward packets to and how long to wait. The tree is constructed in advance and changes when the event moves assuming the location and mobility of the event are known. 190 1 1.2 1.4 1.6 1.8 2 2.2 0 1 2 3 4 5 6 7 8 Normalized Number of Transmissions Maximum Delay (s) ToD DAA SPT SPT-D Figure 13. The normalized number of transmissions for different maximum delays from experiments on 105 sensors. Ideally only n - 1 transmissions are required for n sources. This is the lower bound for any structure, therefore we use it as the optimal case. This approach is similar to the aggregation tree proposed in [4] but without its tree construction and migration overhead. We do not evaluate SPT-D in simulation because SPT-D is not practical in the large scale network. In the largest simulation scenario, the network is a 58-hop network . According to the simulation in smaller network, SPT-D gets best performance when the delay of each hop is about 0 .64 seconds. This makes nodes closer to the sink have about 36 seconds delay in SPT-D, which is not advisable. We perform simulations of these protocols on a 2000m 1200m grid network with 35m node separation, therefore there are a total of 1938 nodes in the network. The data rate of the radio is 38 .4Kbps and the transmission range of the nodes is slightly higher than 50m. An event moves in the network using the random way-point mobility model at the speed of 10m /s for 400 seconds. The event size is 400m in diameter. The nodes triggered by an event will send packets every five seconds to the sink located at (0,0). The aggregation function evaluated here is perfect aggregation, i.e. all packets can be aggregated into one packet without increasing the packet size. 5.3 Event Size We first evaluate the performance for these protocols on different number of nodes generating the packets. This simulation reflects the performance of each protocol for different event sizes. We study the performance for 4 mobility scenarios and show the average, maximum, and minimum values of the results. Fig. 14(a) shows the result of normalized number of transmissions. ToD improves the performance of DAA and SPT by 30% and 85%, and is 25% higher than OPT. However OPT has the best performance by using the aggregation tree that keeps changing when event moves but its overhead is not considered in the simulation. SPT has very poor performance since its aggregation is opportunistic. Except the SPT, the performance of all other protocols is quite steady. This shows that they are quite scalable in terms of the event size. Fig. 14(b) and (c) show the total number of transmissions and total units of useful information received by the sink. As observed in [20], DAA and ToD have higher number of received packets than OPT due to the ability of structure-less aggregation to aggregate packets early and scatter them away from each other to reduce contention. ToD performs better than DAA in terms of the normalized number of transmissions because of its ability to aggregate packets at nodes closer to the source, and thus it reduces the cost of forwarding packets from sources to the sink. 5.4 Scalability In this set of simulations we evaluate the scalability of our protocol since our goal is to design a scalable data aggregation protocol. If a protocol is not scalable, its performance will degrade as the size of the network increases. To evaluate the scalability of a protocol, we limit an event to move only in a bounded region at a certain distance from the sink to simulate the effect of different network sizes. We limit an event to move within a 400m 1200m rectangle, and change the distance of the rectangle to the sink from 200m to 1400m, as shown in Fig. 15. In order to be fair to all scenarios, we limit the event not to move closer than 200m to the network boundary such that the number of nodes triggered by the event does not change drastically. 2000m 1200m 200m 400m 200m Figure 15. The simulation scenario for scalability. The event is limited to move only within a small gray rectangle in each simulation. Fig. 16 shows the results of the scalability simulation. The performance of ToD and OPT remains steady, and ToD is 22% higher than OPT. This shows that ToD is quite scalable as its performance does not degrade as the size of the network increases. The performance of both DAA and SPT degrades as the size of the network increases. The normalized number of transmissions for DAA and SPT doubled when the event moves from the closest rectangle (to the sink) to the farthest rectangle. Fig. 16(c) shows the number of packets received at the sink per event. If all packets can be aggregated near the event and forwarded to the sink, the sink will receive only one packet. Conversely, more packets received at the sink shows that fewer aggregations happened in the network. The cost of forwarding more packets to the sink increases rapidly as the size of the network increases. We can see that in both DAA and SPT the sink receives many packets. Though the number of packets received at the sink remains quite steady, the total number of transmissions increases linearly as the distance from the sources to the sink increases. Ideally the number of received packets at sink is 1, if all packets can be aggregated at the aggregator. However the number of received packets at sink is higher than 1 in ToD 191 0 5 10 15 20 25 30 35 500 400 300 200 Normalized Number of Transmissions Event Size (m) ToD DAA SPT OPT 0 10000 20000 30000 40000 50000 60000 70000 80000 90000 100000 500 400 300 200 Number of Total Transmissions Event Size (m) ToD DAA SPT OPT 0 2000 4000 6000 8000 10000 12000 14000 16000 500 400 300 200 Unit of Received Information Event Size (m) ToD DAA SPT OPT (a) Normalized number of transmissions (b) Number of transmissions (c) Unit of received information Figure 14. The simulation results for different event sizes. and OPT. This is because the delay in CSMA-based MAC protocol can not be accurately predicted therefore the aggregator might send the packet to the sink before all packets are forwarded to it. Though the cost of forwarding the unaggregated packets from aggregator to the sink in ToD and OPT also increases when the size of the network increases, the increase is comparably smaller than DAA and SPT because few packets are forwarded to the sink without aggregation . We observe that the number of received packets at the sink in ToD is higher when the event is closer to the sink. In our simulation, nodes in the same F-cluster as the sink will always use sink as the F-aggregator because we assume that the sink is wire powered therefore there is no need to delegate the role of aggregator to other nodes in order to evenly distribute the energy consumption. 5.5 Cell Size The above simulations use the maximum size of an event as the cell size. As we described in Section 3.2.2, this ensures that the Dynamic Forwarding can aggregate all packets at an S-aggregator, and the cost of forwarding the aggregated packets to the sink is minimized. However, large cell size increases the cost of aggregating packets to the aggregator as we use DAA as the aggregation technique in an F-cluster. In this section we evaluate the impact of the size of a cell on the performance of ToD. We vary the cell size from 50m 50m to 800m 800m and run simulations for three different event sizes, which are 200m, 400m, and 600m in diameter. The results are collected from five different event mobility patterns and shown in Fig. 17. When the size of cell is larger than the event size, the performance is worse because the cost of aggregating packets to F-aggregator increases, but the cost of forwarding packets from S-aggregator does not change. When the size of cell is too small, the cost of forwarding packets to sink increases because packets will be aggregated at different F-aggregators and more packets will be forwarded to the sink without further aggregation. In general, when the size of the F-cluster is small enough to only contain one node, or when the size of the F-cluster is large enough to include all nodes in the network, ToD just downgrades to DAA. ToD has the best performance when the cell size is 100m 100m (F-cluster size is 200m200m) when the event size is 200m in diameter. When the diameter of an event is 400m and 600m, using 200m 200m as the cell size has the best performance (F-cluster size is 400m 400m). This shows that the ToD performance can be further optimized by selecting the appropriate cell size. To explore the relation between the event and cell size for optimization will be part of our future work. Conclusion In this paper we propose a semi-structured approach that locally uses a structure-less technique followed by Dynamic Forwarding on an implicitly constructed packet forwarding structure, ToD, to support network scalability. ToD avoids the long stretch problem in fixed structured approaches and eliminates the overhead of construction and maintenance of dynamic structures. We evaluate its performance using real experiments on a testbed of 105 sensor nodes and simulations on 2000 node networks. Based on our studies we find that ToD is highly scalable and it performs close to the optimal structured approach. Therefore, it is very suitable for conserving energy and extending the lifetime of large scale sensor networks. References [1] W. Heinzelman, A. Chandrakasan, and H. Balakrishnan, "Energy-Efficient Communication Protocol for Wireless Microsensor Networks," in Proceedings of the 33rd Annual Hawaii International Conference on System Sciences, vol. 2, January 2000. [2] W. Heinzelman, A. Chandrakasan, and H. Balakrishnan, "An Application-Specific Protocol Architecture for Wireless Microsensor Networks," in IEEE Transactions on Wireless Communications, vol. 1, October 2002, pp. 660670. [3] C. Intanagonwiwat, D. Estrin, and R. Goviindan, "Impact of Network Density on Data Aggregation in Wireless Sensor Networks," in Technical Report 01-750, University of Southern California, November 2001. [4] W. Zhang and G. Cao, "Optimizing Tree Reconfiguration for Mobile Target Tracking in Sensor Networks," in Proceedings of INFOCOM 2004, vol. 4, March 2004, pp. 24342445. 192 0 5 10 15 20 25 30 35 40 400 600 800 1000 1200 1400 1600 Normalized Number of Transmissions Distance to the Sink (m) ToD DAA SPT OPT 0 2 4 6 8 10 400 600 800 1000 1200 1400 1600 Normalized Number of Transmissions Distance to the Sink (m) ToD DAA OPT 0 5 10 15 20 400 600 800 1000 1200 1400 1600 Number of Received Packets Distance to the Sink (m) ToD DAA SPT OPT (a) Normalized number of transmission (b) Zoom in of Fig. 16(a) (c) Number of received packets Figure 16. The simulation results for difference distances from the event to the sink. 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 100 200 300 400 500 600 700 800 Normalized Number of Transmissions Cell Size (m) 200m 400m 600m 0 10000 20000 30000 40000 50000 60000 70000 80000 90000 100 200 300 400 500 600 700 800 Number of Total Transmissions Cell Size (m) 200m 400m 600m 0 100 200 300 400 500 600 700 800 900 1000 1100 100 200 300 400 500 600 700 800 Number of Received Packets Cell Size (m) 200m 400m 600m (a) Normalized number of transmissions (b) Number of transmissions (c) Number of received packets Figure 17. The simulation results for difference cell sizes. [5] W. Zhang and G. Cao, "DCTC: Dynamic Convoy Tree-based Collaboration for Target Tracking in Sensor Networks," in IEEE Transactions on Wireless Communications, vol. 3, September 2004, pp. 16891701. [6] M. Ding, X. Cheng, and G. Xue, "Aggregation Tree Construction in Sensor Networks," in Proceedings of the 58th IEEE Vehicular Technology Conference, vol. 4, October 2003, pp. 21682172. [7] H. Luo, J. Luo, and Y. Liu, "Energy Efficient Routing with Adaptive Data Fusion in Sensor Networks," in Proceedings of the Third ACM/SIGMOBILEe Workshop on Foundations of Mobile Computing, August 2005. [8] A. Goel and D. Estrin, "Simultaneous Optimization for Concave Costs: Single Sink Aggregation or Single Source Buy-at-Bulk," in Proceedings of the 14th Annual ACM-SIAM Symposium on Discrete Algorithms, 2003. [9] "Networked Infomechanical Systems," http://www.cens.ucla.edu . [10] "Center for Embedded Networked Sensing at UCLA," http://www.cens.ucla.edu . [11] J. Polastre, "Design and Implementation of Wireless Sensor Networks for Habitat Mon itoring," Master's Thesis, University of California at Berkeley, Spring 2003. [12] A. Mainwaring, R. Szewczyk, J. Anderson, and J. Polastre, "Habitat Monitoring on Great Duck Island," http://www.greatduckisland.net. [13] A. Arora, P. Dutta, and S. Bapat, "Line in the Sand: A Wireless Sensor Network for Target Detection, Classification, and Tracking," OSU-CISRC-12/03-TR71, 2003. [14] "ExScal," http://www.cast.cse.ohio-state.edu/exscal/. [15] S. Corporation, "Chemical/Bio Defense and Sensor Networks," http://www.sentel.com/html/chemicalbio.html . [16] E. L. Lawler, J. K. Lenstra, A. H. G. R. Kan, and D. B. Shmoys, The Traveling Salesman Problem : A Guided Tour of Combinatorial Optimization. John Wiley & Sons, 1985. [17] R. Cristescu, B. Beferull-Lozano, and M. Vetterli, "On Network Correlated Data Gathering," in Proceedings of the 23rd Annual Joint Conference of the IEEE Computer and Communications Societies, vol. 4, March 2004, pp. 25712582. [18] K. W. Fan, S. Liu, and P. Sinha, "Structure-free Data Aggregation in Sensor Networks," in OSU-CISRC-4/06-TR35, Technical Report, Dept of CSE, OSU, April 2006. [19] Y. Zhu, K. Sundaresan, and R. Sivakumar, "Practical Limits on Achievable Energy Improvements and 193 Useable Delay Tolerance in Correlation Aware Data Gathering in Wireless Sensor Networks," in IEEE Second Annual IEEE Communications Society Conference on Sensor and Ad Hoc Communications and Networks, September 2005. [20] K. W. Fan, S. Liu, and P. Sinha, "On the potential of Structure-free Data Aggregation in Sensor Networks," in To be appear in Proceedings of INFOCOM 2006, April 2006. [21] C. Intanagonwiwat, R. Govindan, and D. Estrin, "Directed Diffusion: A Scalable and Robust Communication Paradigm for Sensor Networks," in Proceedings of the 6th Annual International Conference on Mobile Computing and Networking, August 2000, pp. 5667. [22] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and F. Silva, "Directed Diffusion for Wireless Sensor Networking," in IEEE/ACM Transactions on Networking, vol. 11, February 2003, pp. 216. [23] S. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong, "TAG: a Tiny AGgregation Service for Ad-Hoc Sensor Networks," in Proceedings of the 5th symposium on Operating systems design and implementation, December 2002, pp. 131146. [24] S. Madden, R. Szewczyk, M. J. Franklin, and D. Culler, "Supporting Aggregate Queries Over Ad-Hoc Wireless Sensor Networks," in Proceedings of the 4th IEEE Workshop on Mobile Computing Systems and Applications, June 2004, pp. 4958. [25] S. Lindsey, C. S. Raghavendra, and K. M. Sivalingam, "Data Gathering in Sensor Networks using the Energy*Delay Metric," in Proceedings 15th International Parallel and Distributed Processing Symposium, April 2001, pp. 20012008. [26] S. Lindsey and C. Raghavendra, "PEGASIS: Power-efficient gathering in sensor information systems," in Proceedings of IEEE Aerospace Conference, vol. 3, March 2002, pp. 11251130. [27] S. Lindsey, C. Raghavendra, and K. M. Sivalingam, "Data Gathering Algorithms in Sensor Networks Using Energy Metrics," in IEEE Transactions on Parallel and Distributed Systems, vol. 13, September 2002, pp. 924935. [28] J. Wong, R. Jafari, and M. Potkonjak, "Gateway placement for latency and energy efficient data aggregation," in 29th Annual IEEE International Conference on Local Computer Networks, November 2004, pp. 490497. [29] B. J. Culpepper, L. Dung, and M. Moh, "Design and Analysis of Hybrid Indirect Transmissions (HIT) for Data Gathering in Wireless Micro Sensor Networks," in ACM SIGMOBILE Mobile Computing and Communications Review, vol. 8, January 2004, pp. 6183. [30] H. F. Salama, D. S. Reeves, and Y. Viniotis, "Evaluation of Multicast Routing Algorithms for Real-time Communication on High-speed Networks," in IEEE Journal on Selected Area in Communications, vol. 15, April 1997. [31] A. Scaglione and S. D. Servetto, "On the Interdependence of Routing and Data Compression in Multi-Hop Sensor Networks," in Proceedings of the 8th Annual International Conference on Mobile Computing and Networking, September 2002, pp. 140147. [32] A. Scaglione, "Routing and Data Compression in Sensor Networks: Stochastic Models for Sensor Data that Guarantee Scalability," in Proceedings of IEEE International Symposium on Information Theory, June 2003, p. 174. [33] R. Cristescu and M. Vetterli, "Power Efficient Gathering of Correlated Data: Optimization, NP-Completeness and Heuristics," in Summaries of MobiHoc 2003 posters, vol. 7, July 2003, pp. 3132. [34] S. Pattern, B. Krishnamachari, and R. Govindan, "The Impact of Spatial Correlation on Routing with Compression in Wireless Sensor Networks," in Proceedings of the 3rd International Symposium on Information Processing in Sensor Networks, April 2004, pp. 2835. [35] L. Cai and D. Corneil, "Tree Spanners," in SIAM Journal of Discrete Mathematics, vol. 8, 1995. [36] N. Alon, R. M. Karp, D. Peleg, and D. West, "A graph theoretic game and its application to the k-server problem," in SIAM Journal of Computing, vol. 24, 1995. [37] D. Peleg and D. Tendler, "Low Stretch Spanning Trees for Planar Graphs," in Technical Report MCS01-14, Mathematics & Computer Sience, Weizmann Institute Of Sience, 2001. [38] L. Jia, G. Noubir, R. Rajaraman, and R. Sundaram, "GIST: Group-Independent Spanning Tree for Data Aggregation in Dense Sensor Networks," in International Conference on Distributed Computing in Sensor Systems, June 2006. [39] N. Bulusu, J. Heidemann, and D. Estrin, "GPS-less Low Cost Outdoor Localization For Very Small Devices," in IEEE Personal Communications, Special Issue on "Smart Spaces and Environments", vol. 7, October 2000. [40] D. Moore, J. Leonard, D. Rus, and S. Teller, "Robust Distributed Network Localization with Noisy Range Measurements," in Proceedings of 2nd ACM Sensys, pp. 5061. [41] Crossbow, "Crossbow," http://www.xbow.com. 194

ToD;Structure-free;Anycasting;Data Aggregation

173

Scalable Mining of Large Disk-based Graph Databases

Mining frequent structural patterns from graph databases is an interesting problem with broad applications. Most of the previous studies focus on pruning unfruitful search subspaces effectively, but few of them address the mining on large, disk-based databases. As many graph databases in applications cannot be held into main memory, scalable mining of large, disk-based graph databases remains a challenging problem. In this paper, we develop an effective index structure, ADI (for adjacency index), to support mining various graph patterns over large databases that cannot be held into main memory. The index is simple and efficient to build. Moreover, the new index structure can be easily adopted in various existing graph pattern mining algorithms. As an example , we adapt the well-known gSpan algorithm by using the ADI structure. The experimental results show that the new index structure enables the scalable graph pattern mining over large databases. In one set of the experiments, the new disk-based method can mine graph databases with one million graphs, while the original gSpan algorithm can only handle databases of up to 300 thousand graphs. Moreover, our new method is faster than gSpan when both can run in main memory.

INTRODUCTION Mining frequent graph patterns is an interesting research problem with broad applications, including mining structural patterns from chemical compound databases, plan databases, XML documents, web logs, citation networks, and so forth. Several efficient algorithms have been proposed in the previous studies [2, 5, 6, 8, 11, 9], ranging from mining graph patterns, with and without constraints, to mining closed graph patterns. Most of the existing methods assume implicitly or explic-itly that the databases are not very large, and the graphs in the database are relatively simple. That is, either the databases or the major part of them can fit into main memory , and the number of possible labels in the graphs [6] is small. For example, [11] reports the performance of gSpan, an efficient frequent graph pattern mining algorithm, on data sets of size up to 320 KB, using a computer with 448 MB main memory. Clearly, the graph database and the projected databases can be easily accommodated into main memory. Under the large main memory assumption, the computation is CPU-bounded instead of I/O-bounded. Then, the algorithms focus on effective heuristics to prune the search space. Few of them address the concern of handling large graph databases that cannot be held in main memory. While the previous studies have made excellent progress in mining graph databases of moderate size, mining large, disk-based graph databases remains a challenging problem. When mining a graph database that cannot fit into main memory, the algorithms have to scan the database and navigate the graphs repeatedly. The computation becomes I/O-bounded . For example, we obtain the executable of gSpan from the authors and test its scalability. In one of our experiments 1 , we increase the number of graphs in the database to test the scalability of gSpan on the database size. gSpan can only handle up to 300 thousand graphs. In another experiment , we increase the number of possible labels in graphs. We observe that the runtime of gSpan increases exponentially . It finishes a data set of 300 thousand graphs with 636 seconds when there are only 10 possible labels, but needs 15 hours for a data set with the same size but the number of possible labels is 45! This result is consistent with the results reported in [11]. Are there any real-life applications that need to mine large graph databases? The answer is yes. For example, in data integration of XML documents or mining semantic web, it is often required to find the common substructures from a huge collection of XML documents. It is easy to see applications with collections of millions of XML documents. There are 1 Details will be provided in Section 6 316 Research Track Paper hundreds of even thousands of different labels. As another example, chemical structures can be modeled as graphs. A chemical database for drug development can contain millions of different chemical structures, and the number of different labels in the graphs can easily go to up to 100. These large databases are disk-based and often cannot be held into main memory. Why is mining large disk-based graph databases so challenging ? In most of the previous studies, the major data structures are designed for being held in main memory. For example, the adjacency-list or adjacency-matrix representations are often used to represent graphs. Moreover, most of the previous methods are based on efficient random accesses to elements (e.g., edges and their adjacent edges) in graphs. However, if the adjacency-list or adjacency-matrix representations cannot be held in main memory, the random accesses to them become very expensive. For disk-based data, without any index, random accesses can be extremely costly. Can we make mining large, disk-based graph databases feasible and scalable? This is the motivation of our study. Since the bottleneck is the random accesses to the large disk-based graph databases, a natural idea is to index the graph databases properly. Designing effective and efficient index structures is one of the most invaluable exercises in database research. A good index structure can support a general category of data access operations. Particularly, a good index should be efficient and scalable in construction and maintenance, and fast for data access. Instead of inventing new algorithms to mine large, disk-based graph patterns, can we devise an efficient index structure for graph databases so that mining various graph patterns can be conducted scalably? Moreover, the index structure should be easy to be adopted in various existing methods with minor adaptations. Stimulated by the above thinking, in this paper, we study the problem of efficient index for scalable mining of large, disk-based graph databases, and make the following contributions . By analyzing the frequent graph pattern mining problem and the typical graph pattern mining algorithms (taking gSpan as an example), we identify several bottleneck data access operations in mining large, disk-based graph databases. We propose ADI (for adjacency index), an effective index structure for graphs. We show that the major operations in graph mining can be facilitated efficiently by an ADI structure. The construction algorithm of ADI structure is presented. We adapt the gSpan algorithm by using the ADI structure on mining large, disk-based graph databases, and achieve algorithm ADI-Mine. We show that ADI-Mine outperforms gSpan in mining complex graph databases and can mine much larger databases than gSpan. A systematic performance study is reported to verify our design. The results show that our new index structure and algorithm are scalable on large data sets. The remainder of the paper is organized as follows. We define the problem of frequent graph pattern mining in Section 2. The idea of minimum DFS code and algorithm gSpan b b a a y z x x x x z a a b v3 v2 v1 v0 b b a a y z x x v3 v2 v0 v1 b b a a y z x x (a) Graph (b) Subgraph (c) DFS-tree (d) DFS-tree G G T 1 T 2 Figure 1: Subgraph and DFS codes are reviewed in Section 3, and the major data access operations in graph mining are also identified. The ADI structure is developed in Section 4. The efficient algorithm ADI-Mine for mining large, disk-based graph databases using ADI is presented in Section 5. The experimental results are reported in Section 6. The related work is discussed in Section 7. Section 8 concludes the paper. PROBLEM DEFINITION In this paper, we focus on undirected labeled simple graphs. A labeled graph is a 4-tuple G = (V, E, L, l), where V is a set of vertices, E V V is a set of edges, L is a set of labels, and l : V E L is a labeling function that assigns a label to an edge or a vertex. We denote the vertex set and the edge set of a graph G by V (G) and E(G), respectively. A graph G is called connected if for any vertices u, v V (G), there exist vertices w 1 , . . . , w n V (G) such that {(u, w 1 ), (w 1 , w 2 ), . . . , (w n-1 , w n ), (w n , v) } E(G). Frequent patterns in graphs are defined based on subgraph isomorphism. Definition 1 (Subgraph isomorphism). Given graphs G = (V, E, L, l) and G = (V , E , L , l ). An injective function f : V V is called a subgraph isomorphism from G to G if (1) for any vertex u V , f(u) V and l (u) = l(f(u)); and (2) for any edge (u, v) E , (f(u), f(v)) E and l (u, v) = l(f (u), f (v)). If there exists a subgraph isomorphism from G to G, then G is called a subgraph of G and G is called a supergraph of G , denoted as G G. For example, the graph G in Figure 1(b) is a subgraph of G in Figure 1(a). A graph database is a set of tuples (gid, G), where gid is a graph identity and G is a graph. Given a graph database GDB, the support of a graph G in GDB, denoted as sup(G ) for short, is the number of graphs in the database that are supergraphs of G , i.e., |{(gid, G) GDB|G G }|. For a support threshold min sup (0 min sup |GDB|), a graph G is called a frequent graph pattern if sup(G ) min sup . In many applications, users are only interested in the frequent recurring components of graphs. Thus, we put a constraint on the graph patterns: we only find the frequent graph patterns that are connected. Problem definition. Given a graph database GDB and a support threshold min sup. The problem of mining frequent connected graph patterns is to find the complete set of connected graphs that are frequent in GDB. 317 Research Track Paper MINIMUM DFS CODE AND GSPAN In [11], Yan and Han developed the lexicographic ordering technique to facilitate the graph pattern mining. They also propose an efficient algorithm, gSpan, one of the most efficient graph pattern mining algorithms so far. In this section, we review the essential ideas of gSpan, and point out the bottlenecks in the graph pattern mining from large disk-based databases. 3.1 Minimum DFS Code In order to enumerate all frequent graph patterns efficiently , we want to identify a linear order on a representation of all graph patterns such that if two graphs are in identical representation, then they are isomorphic. Moreover, all the (possible) graph patterns can be enumerated in the order without any redundancy. The depth-first search tree (DFS-tree for short) [3] is popularly used for navigating connected graphs. Thus, it is natural to encode the edges and vertices in a graph based on its DFS-tree. All the vertices in G can be encoded in the pre-order of T . However, the DFS-tree is generally not unique for a graph. That is, there can be multiple DFS-trees corresponding to a given graph. For example, Figures 1(c) and 1(d) show two DFS-trees of the graph G in Figure 1(a). The thick edges in Figures 1(c) and 1(d) are those in the DFS-trees, and are called forward edges, while the thin edges are those not in the DFS-trees, and are called backward edges. The vertices in the graph are encoded v 0 to v 3 according to the pre-order of the corresponding DFS-trees. To solve the uniqueness problem, a minimum DFS code notation is proposed in [11]. For any connected graph G, let T be a DFS-tree of G. Then, an edge is always listed as (v i , v j ) such that i &lt; j. A linear order on the edges in G can be defined as follows. Given edges e = (v i , v j ) and e = (v i , v j ). e e if (1) when both e and e are forward edges (i.e., in DFS-tree T ), j &lt; j or (i &gt; i j = j ); (2) when both e and e are backward edges (i.e., edges not in DFS-tree T ), i &lt; i or (i = i j &lt; j ); (3) when e is a forward edge and e is a backward edge, j i ; or (4) when e is a backward edge and e is a forward edge, i &lt; j . For a graph G and a DFS-tree T , a list of all edges in E(G) in order is called the DFS code of G with respect to T , denoted as code(G, T ). For example, the DFS code with respect to the DFS-tree T 1 in Figure 1(c) is code(G, T 1 ) = (v 0 , v 1 , x, a, x)-(v 1 , v 2 , x, a, z)-(v 2 , v 0 , z, b, x)-(v 1 , v 3 , x, b, y) , where an edge (v i , v j ) is written as (v i , v j , l(v i ), l(v i , v j ), l(v j )), i.e., the labels are included. Similarly, the DFS code with respect to the DFS-tree T 2 in Figure 1(d) is code(G, T 2 ) = (v 0 , v 1 , y, b, x)-(v 1 , v 2 , x, a, x)-(v 2 , v 3 , x, b, z) (v 3 , v 1 , z, a, x) . Suppose there is a linear order over the label set L. Then, for DFS-trees T 1 and T 2 on the same graph G, their DFS codes can be compared lexically according to the labels of the edges. For example, we have code(G, T 1 ) &lt; code(G, T 2 ) in Figures 1(c) and 1(d). The lexically minimum DFS code is selected as the representation of the graph, denoted as min(G). In our example in Figure 1, min(G) = code(G, T 1 ). Minimum DFS code has a nice property: two graphs G and G are isomorphic if and only if min(G) = min(G ). Moreover, with the minimum DFS code of graphs, the prob-Input : a DFS code s, a graph database GDB and min sup Output: the frequent graph patterns Method: if s is not a minimum DFS code then return; output s as a pattern if s is frequent in GDB; let C = ; scan GDB once, find every edge e such that e can be concatenated to s to form a DFS code s e and s e is frequent; C = C {s e}; sort the DFS codes in C in lexicographic order; for each s e C in lexicographic order do call gSpan(s e, GDB, min sup); return; Figure 2: Algorithm gSpan. lem of mining frequent graph patterns is reduced to mining frequent minimum DFS codes, which are sequences, with some constraints that preserve the connectivity of the graph patterns. 3.2 Algorithm gSpan Based on the minimum DFS codes of graphs, a depth-first search, pattern-growth algorithm, gSpan, is developed in [11], as shown in Figure 2. The central idea is to conduct a depth-first search of minimum DFS codes of possible graph patterns, and obtain longer DFS codes of larger graph patterns by attaching new edges to the end of the minimum DFS code of the existing graph pattern. The anti-monotonicity of frequent graph patterns, i.e., any super pattern of an infrequent graph pattern cannot be frequent, is used to prune. Comparing to the previous methods on graph pattern mining, gSpan is efficient, since gSpan employs the smart idea of minimum DFS codes of graph patterns that facilitates the isomorphism test and pattern enumeration. Moreover , gSpan inherits the depth-first search, pattern-growth methodology to avoid any candidate-generation-and-test. As reported in [11], the advantages of gSpan are verified by the experimental results on both real data sets and synthetic data sets. 3.3 Bottlenecks in Mining Disk-based Graph Databases Algorithm gSpan is efficient when the database can be held into main memory. For example, in [11], gSpan is scalable for databases of size up to 320 KB using a computer with 448 MB main memory. However, it may encounter difficulties when mining large databases. The major overhead is that gSpan has to randomly access elements (e.g., edges and vertices) in the graph database as well as the projections of the graph database many times. For databases that cannot be held into main memory, the mining becomes I/O bounded and thus is costly. Random accesses to elements in graph databases and checking the isomorphism are not unique to gSpan. Instead, such operations are extensive in many graph pattern mining algorithms , such as FSG [6] (another efficient frequent graph pattern mining algorithm) and CloseGraph [9] (an efficient algorithm for mining frequent closed graph patterns). In mining frequent graph patterns, the major data access operations are as follows. 318 Research Track Paper OP1: Edge support checking. Find the support of an edge (l u , l e , l v ), where l u and l v are the labels of vertices and l e is the label of the edge, respectively; OP2: Edge-host graph checking. For an edge e = (l u , l e , l v ), find the graphs in the database where e appears ; OP3: Adjacent edge checking. For an edge e = (l u , l e , l v ), find the adjacent edges of e in the graphs where e appears, so that the adjacent edges can be used to expand the current graph pattern to larger ones. Each of the above operations may happen many times during the mining of frequent graph patterns. Without an appropriate index, each of the above operations may have to scan the graph database or its projections. If the database and its projections cannot fit into main memory, the scanning and checking can be very costly. Can we devise an index structure so that the related information can be kept and all the above operations can be achieved using the index only, and thus without scanning the graph database and checking the graphs? This motivates the design of the ADI structure. THE ADI STRUCTURE In this section we will devise an effective data structure, ADI (for adjacency index), to facilitate the scalable mining of frequent graph patterns from disk-based graph databases. 4.1 Data Structure The ADI index structure is a three-level index for edges, graph-ids and adjacency information. An example is shown in Figure 3, where two graphs, G 1 and G 2 , are indexed. 4.1.1 Edge Table There can be many edges in a graph database. The edges are often retrieved by the labels during the graph pattern mining, such as in the operations identified in Section 3.3. Therefore, the edges are indexed by their labels in the ADI structure. In ADI, an edge e = (u, v) is recorded as a tuple (l(u), l(u, v), l(v)) in the edge table, and is indexed by the labels of the vertices, i.e., l(u) and l(v), and the label of the edge itself, i.e., l(u, v). Each edge appears only once in the edge table, no matter how many times it appears in the graphs. For example, in Figure 3, edge (A, d, C) appears once in graph G 1 and twice in graph G 2 . However, there is only one entry for the edge in the edge table in the ADI structure. All edges in the edge table in the ADI structure are sorted. When the edge table is stored on disk, a B+-tree is built on the edges. When part of the edge table is loaded into main memory, it is organized as a sorted list. Thus, binary search can be conducted. 4.1.2 Linked Lists of Graph-ids For each edge e, the identities of the graphs that contain e form a linked list of graph-ids. Graph-id G i is in the list of edge e if and only if there exists at least one instance of e in G i . For example, in Figure 3, both G 1 and G 2 appear in the list of edge (A, d, C), since the edge appears in G 1 once and in G 2 twice. Please note that the identity of graph G i G1 G2 G1 G2 G2 G1 A B C D a d d b 1 2 3 4 G1 A B C C D B a c d d d 1 2 3 4 5 G2 Edges Block 1 Block 2 Graph-ids (on disk) 1 2 2 3 1 4 3 4 1 2 1 4 1 6 2 3 4 5 (A, a, B) (A, d, C) (B, b, D) G2 G1 (B, c, C) (B, d, D) (C, d, D) Adjacency (on disk) 6 Figure 3: An ADI structure. appears in the linked list of edge e only once if e appears in G i , no matter how many times edge e appears in G i . A list of graph-ids of an edge are stored together. Therefore , given an edge, it is efficient to retrieve all the identities of graphs that contain the edge. Every entry in the edge table is linked to its graph-id linked list. By this linkage, the operation OP2: edge-host graph checking can be conducted efficiently. Moreover, to facilitate operation OP1: edge support checking, the length of the graph-id linked list, i.e., the support of an edge, is registered in the edge table. 4.1.3 Adjacency Information The edges in a graph are stored as a list of the edges encoded. Adjacent edges are linked together by the common vertices, as shown in Figure 3. For example, in block 1, all the vertices having the same label (e.g., 1) are linked together as a list. Since each edge has two vertices, only two pointers are needed for each edge. Moreover, all the edges in a graph are physically stored in one block on disk (or on consecutive blocks if more space is needed), so that the information about a graph can be retrieved by reading one or several consecutive blocks from disk. Often, when the graph is not large, a disk-page (e.g., of size 4k) can hold more than one graph. Encoded edges recording the adjacency information are linked to the graph-ids that are further associated with the edges in the edge table. 4.2 Space Requirement The storage of an ADI structure is flexible. If the graph database is small, then the whole index can be held into main memory. On the other hand, if the graph database is large and thus the ADI structure cannot fit into main 319 Research Track Paper memory, some levels can be stored on disk. The level of adjacency information is the most detailed and can be put on disk. If the main memory is too small to hold the graph-id linked lists, they can also be accommodated on disk. In the extreme case, even the edge table can be held on disk and a B+-tree or hash index can be built on the edge table. Theorem 1 (Space complexity). For graph database GDB = {G 1 , . . . , G n }, the space complexity is O( n i=1 |E(G i ) |). Proof. The space complexity is determined by the following facts. (1) The number of tuples in the edge table is equal to the number of distinct edges in the graph database, which is bounded by n i=1 |E(G i ) |; (2) The number of entries in the graph-id linked lists in the worst case is the number of edges in the graph database, i.e., n i=1 |E(G i ) | again; and (3) The adjacency information part records every edge exactly once. Please note that, in many application, it is reasonable to assume that the edge table can be held into main memory. For example, suppose we have 1, 000 distinct vertex labels and 1, 000 distinct edge labels. There can be up to 1000 999 2 1000 = 4.995 10 8 different edges, i.e., all possible combinations of vertex and edge labels. Suppose up to 1% edges are frequent, there are only less than 5 million different edges, and thus the edge table can be easily held into main memory. In real applications, the graphs are often sparse, that is, not all possible combinations of vertex and edge labels appear in the graphs as an edge. Moreover, users are often interested in only those frequent edges. That shrinks the edge table substantially. 4.3 Search Using ADI Now, let us examine how the ADI structure can facilitate the major data access operations in graph pattern mining that are identified in Section 3.3. OP1: Edge support checking Once an ADI structure is constructed, this information is registered on the edge table for every edge. We only need to search the edge table, which is either indexed (when the table is on disk) or can be searched using binary search (when the table is in main memory). In some cases, we may need to count the support of an edge in a subset of graphs G G. Then, the linked list of the graph-ids of the edge is searched. There is no need to touch any record in the adjacency information part. That is, we do not need to search any detail about the edges. Moreover, for counting supports of edges in projected databases, we can maintain the support of each edge in the current projected database and thus we do not even search the graph-id linked lists. OP2: Edge-host graph checking We only need to search the edge table for the specific edge and follow the link from the edge to the list of graph-ids. There is no need to search any detail from the part of adjacency information. OP3: Adjacent edge checking Again, we start from an entry in the edge table and follow the links to find the list of graphs where the edge appears. Then, only Input: a graph database GDB and min sup Output: the ADI structure Method: scan GDB once, find the frequent edges; initialize the edge table for frequent edges; for each graph do remove infrequent edges; compute the mininmum DFS code [11]; use the DFS-tree to encode the vertices; store the edges in the graph onto disk and form the adjacency information; for each edge do insert the graph-id to the graph-id list associated with the edge; link the graph-id to the related adjacency information; end for end for Figure 4: Algorithm of ADI construction. the blocks containing the details of the instances of the edge are visited, and there is no need to scan the whole database. The average I/O complexity is O(log n + m + l), where n is the number of distinct edges in the graph, m is the average number of graph-ids in the linked lists of edges, and l is the average number of blocks occupied by a graph. In many applications, m is orders of magnitudes smaller than the n, and l is a very small number (e.g., 1 or 2). The algorithms for the above operations are simple. Limited by space, we omit the details here. As can be seen, once the ADI structure is constructed, there is no need to scan the database for any of the above operations. That is, the ADI structure can support the random accesses and the mining efficiently. 4.4 Construction of ADI Given a graph database, the corresponding ADI structure is easy to construct by scanning the database only twice. In the first scan, the frequent edges are identified. According to the apriori property of frequent graph patterns, only those frequent edges can appear in frequent graph patterns and thus should be indexed in the ADI structure. After the first scan, the edge table of frequent edges is initialized. In the second scan, graphs in the database are read and processed one by one. For each graph, the vertices are encoded according to the DFS-tree in the minimum DFS code, as described in [11] and Section 3. Only the vertices involved in some frequent edges should be encoded. Then, for each frequent edge, the graph-id is inserted into the corresponding linked list, and the adjacency information is stored. The sketch of the algorithm is shown in Figure 4. Cost Analysis There are two major costs in the ADI construction: writing the adjacency information and updating the linked lists of graph-ids. Since all edges in a graph will reside on a disk page or several consecutive disk pages, the writing of adjacency information is sequential. Thus, the cost of writing adjacency information is comparable to that of making a 320 Research Track Paper 4 3 2 1 D C A B a d 1 d 3 d 4 b 2 a 1 b 3 d 4 a 2 4 C 3 D 2 B 1 A d d b (a) The graph and the adjacency-lists 1 A 2 B 3 D 4 C 1 A 0 a 0 d 2 B a 0 b 0 3 D 0 b 0 d 4 C d 0 d 0 (b) The adjacency-matrix Figure 5: The adjacency-list and adjacency-matrix representations of graphs. copy of the original database plus some bookkeeping. Updating the linked lists of graph-ids requires random accesses to the edge table and the linked lists. In many cases, the edge table can be held into main memory, but not the linked list. Therefore, it is important to cache the linked lists of graph-ids in a buffer. The linked lists can be cached according to the frequency of the corresponding edges. Constructing ADI for large, disk-based graph database may not be cheap. However, the ADI structure can be built once and used by the mining many times. That is, we can build an ADI structure using a very low support threshold, or even set min sup = 1. 2 The index is stored on disk. Then, the mining in the future can use the index directly, as long as the support threshold is no less than the one that is used in the ADI structure construction. 4.5 Projected Databases Using ADI Many depth-first search, pattern-growth algorithms utilize proper projected databases. During the depth-first search in graph pattern mining, the graphs containing the current graph pattern P should be collected and form the P projected database. Then, the further search of larger graph patterns having P as the prefix of their minimum DFS codes can be achieved by searching only the P -projected database. Interestingly, the projected databases can be constructed using ADI structures. A projected database can be stored in the form of an ADI structure. In fact, only the edge table and the list of graph-ids should be constructed for a new projected database and the adjacency information residing on disk can be shared by all projected databases. That can save a lot of time and space when mining large graph databases that contain many graph patterns, where many projected databases may have to be constructed. 4.6 Why Is ADI Good for Large Databases? In most of the previous methods for graph pattern mining, the adjacency-list or adjacency-matrix representations are used to represent graphs. Each graph is represented by an adjacency-matrix or a set of adjacency-lists. An example is shown in Figure 5. 2 If min sup = 1, then the ADI structure can be constructed by scanning the graph database only once. We do not need to find frequent edges, since every edge appearing in the graph database is frequent. In Figure 5(a), the adjacency-lists have 8 nodes and 8 pointers. It stores the same information as Block 1 in Figure 3, where the block has 4 nodes and 12 pointers. The space requirements of adjacency-lists and ADI structure are comparable. From the figure, we can see that each edge in a graph has to be stored twice: one instance for each vertex. (If we want to remove this redundancy, the tradeoff is the substantial increase of cost in finding adjacency information). In general, for a graph of n edges, the adjacency-list representation needs 2n nodes and 2n pointers . An ADI structure stores each edge once, and use the linkage among the edges from the same vertex to record the adjacency information. In general, for a graph of n edges, it needs n nodes and 3n pointers. Then, what is the advantage of ADI structure against adjacency-list representation? The key advantage is that the ADI structure extracts the information about containments of edges in graphs in the first two levels (i.e., the edge table and the linked list of graph-ids). Therefore, in many operations , such as the edge support checking and edge-host graph checking, there is no need to visit the adjacency information at all. To the contrast, if the adjacency-list representation is used, every operation has to check the linked lists. When the database is large so that either the adjacency-lists of all graphs or the adjacency information in the ADI structure cannot be accommodated into main memory, using the first two levels of the ADI structure can save many calls to the adjacency information, while the adjacency-lists of various graphs have to be transferred between the main memory and the disk many times. Usually, the adjacency-matrix is sparse. The adjacency-matrix representation is inefficient in space and thus is not used. ALGORITHM ADI-MINE With the help from the ADI structure, how can we improve the scalability and efficiency of frequent graph pattern mining? Here, we present a pattern-growth algorithm ADI-Mine , which is an improvement of algorithm gSpan. The algorithm is shown in Figure 6. If the ADI structure is unavailable, then the algorithm scans the graph database and constructs the index. Otherwise , it just uses the ADI structure on the disk. The frequent edges can be obtained from the edge table in the ADI structure. Each frequent edge is one of the smallest frequent graph patterns and thus should be output. Then, the frequent edges should be used as the "seeds" to grow larger frequent graph patterns, and the frequent adjacent edges of e should be used in the pattern-growth. An edge e is a frequent adjacent edge of e if e is an adjacent edge of e in at least min sup graphs. The set of frequent adjacent edges can be retrieved efficiently from the ADI structure since the identities of the graphs containing e are indexed as a linked-list, and the adjacent edges are also indexed in the adjacency information part in the ADI structure. The pattern growth is implemented as calls to procedure subgraph-mine. Procedure subgraph-mine tries every frequent adjacent edge e (i.e., edges in set F e ) and checks whether e can be added into the current frequent graph pattern G to form a larger pattern G . We use the DFS code to test the redundancy. Only the patterns G whose DFS code is minimum is output and further grown. All other patterns G are either found before or will be found later at other 321 Research Track Paper Input: a graph database GDB and min sup Output: the complete set of frequent graph patterns Method: construct the ADI structure for the graph database if it is not available; for each frequent edge e in the edge table do output e as a graph pattern; from the ADI structure, find set F e , the set of frequent adjacent edges for e; call subgraph-mine(e, F e ); end for Procedure subgraph-mine Parameters: a frequent graph pattern G, and the set of frequent adjacent edges F e // output the frequent graph patterns whose // minimum DFS-codes contain that of G as a prefix Method: for each edge e in F e do let G be the graph by adding e into G; compute the DFS code of G ; if the DFS code is not minimum, then return; output G as a frequent graph pattern; update the set F e of adjacent edges; call subgraph-mine(G , F e ); end for return; Figure 6: Algorithm ADI-Mine. branches. The correctness of this step is guaranteed by the property of DFS code [11]. Once a larger pattern G is found, the set of adjacent edges of the current pattern should be updated, since the adjacent edges of the newly inserted edge should also be considered in the future growth from G . This update operation can be implemented efficiently, since the identities of graphs that contain an edge e are linked together in the ADI structure, and the adjacency information is also indexed and linked according to the graph-ids. Differences Between ADI-Mine and gSpan At high level, the structure as well as the search strategies of ADI-Mine and gSpan are similar. The critical difference is on the storage structure for graphs--ADI-Mine uses ADI structure and gSpan uses adjacency-list representation. In the recursive mining, the critical operation is finding the graphs that contain the current graph pattern (i.e., the test of subgraph isomorphism) and finding the adjacent edges to grow larger graph patterns. The current graph pattern is recorded using the labels. Thus, the edges are searched using the labels of the vertices and that of the edges. In gSpan, the test of subgraph isomorphism is achieved by scanning the current (projected) database. Since the graphs are stored in adjacency-list representation, and one label may appear more than once in a graph, the search can be costly. For example, in graph G 2 in Figure 3, in order to find an edge (C, d, A), the adjacency-list for vertices 4 and 6 may have to be searched. If the graph is large and the labels appear multiple times in a graph, there may be many adjacency-lists for vertices of the same label, and the adjacency-lists are long. Moreover, for large graph database that cannot be held into main memory, the adjacency-list representation of a graph has to be loaded into main memory before the graph can be searched. In ADI-Mine, the graphs are stored in the ADI structure. The edges are indexed by their labels. Then, the graphs that contain the edges can be retrieved immediately. Moreover, all edges with the same labels are linked together by the links between the graph-id and the instances. That helps the test of subgraph isomorphism substantially. Furthermore, using the index of edges by their labels, only the graphs that contain the specific edge will be loaded into main memory for further subgraph isomorphism test. Irrelevant graphs can be filtered out immediately by the index. When the database is too large to fit into main memory, it saves a substantial part of transfers of graphs between disk and main memory. EXPERIMENTAL RESULTS In this section, we report a systematic performance study on the ADI structure and a comparison of gSpan and ADI-Mine on mining both small, memory-based databases and large, disk-based databases. We obtain the executable of gSpan from the authors. The ADI structure and algorithm ADI-Mine are implemented using C/C++. 6.1 Experiment Setting All the experiments are conducted on an IBM NetFinity 5100 machine with an Intel PIII 733MHz CPU, 512M RAM and 18G hard disk. The speed of the hard disk is 10, 000 RPM. The operating system is Redhat Linux 9.0. We implement a synthetic data generator following the procedure described in [6]. The data generator takes five parameters as follows. D: the total number of graphs in the data set T : the average number of edges in graphs I: the average number of edges in potentially frequent graph patterns (i.e., the frequent kernels) L: the number of potentially frequent kernels N : the number of possible labels Please refer to [6] for the details of the data generator. For example, a data set D10kN 4I10T 20L200 means that the data set contains 10k graphs; there are 4 possible labels; the average number of edges in the frequent kernel graphs is 10; the average number of edges in the graphs is 20; and the number of potentially frequent kernels is 200. Hereafter in this section, when we say "parameters", it means the parameters for the data generator to create the data sets. In [11], L is fixed to 200. In our experiments, we also set L = 200 as the default value, but will test the scalability of our algorithm on L as well. Please note that, in all experiments, the runtime of ADI-Mine includes both the ADI construction time and the mining time. 6.2 Mining Main Memory-based Databases In this set of experiments, both gSpan and ADI-Mine run in main memory. 322 Research Track Paper 6.2.1 Scalability on Minimum Support Threshold We test the scalability of gSpan and ADI-Mine on the minimum support threshold. Data set D100kN 30I5T 20L200 is used. The minimum support threshold varies from 4% to 10%. The results are shown in Figure 7(a). As can be seen, both gSpan and ADI-Mine are scalable, but ADI-Mine is about 10 times faster. We discussed the result with Mr. X. Yan, the author of gSpan. He confirms that counting frequent edges in gSpan is time consuming. On the other hand, the construction of ADI structure is relatively efficient. When the minimum support threshold is set to 1, i.e., all edges are indexed, the ADI structure uses approximately 57M main memory and costs 86 seconds in construction. 6.2.2 Scalability on Database Size We test the scalability of gSpan and ADI-Mine on the size of databases. We fix the parameters N = 30, I = 5, T = 20 and L = 200, and vary the number of graphs in database from 50 thousand to 100 thousand. The minimum support threshold is set to 1% of the number of graphs in the database. The results are shown in Figure 7(b). The construction time of ADI structure is also plotted in the figure. Both the algorithms and the construction of ADI structure are linearly scalable on the size of databases. ADI-Mine is faster. We observe that the size of ADI structure is also scalable. For example, it uses 28M when the database has 50 thousand graphs, and 57M when the database has 100 thousand graphs. This observation concurs with Theorem 1. 6.2.3 Effects of Data Set Parameters We test the scalability of the two algorithms on parameter N --the number of possible labels. We use data set D100kN 20-50I5T 20L200, that is, the N value varies from 20 to 50. The minimum support threshold is fixed at 1%. The results are shown in Figure 7(c). Please note that the Y -axis is in logarithmic scale. We can observe that the runtime of gSpan increases exponentially as N increases. This result is consistent with the result reported in [11]. 3 When there are many possible labels in the database, the search without index becomes dramatically more costly. Interestingly, both ADI-Mine and the construction of ADI structure are linearly scalable on N . As discussed before, the edge table in ADI structure only indexes the unique edges in a graph database. Searching using the indexed edge table is efficient. The time complexity of searching an edge by labels is O(log n), where n is the number of distinct edges in the database. This is not affected by the increase of the possible labels. As expected, the size of the ADI structure is stable, about 57M in this experiment. We use data set D100kN 30I5T 10-30L200 to test the scalability of the two algorithms on parameter T --the average number of edges in a graph. The minimum support threshold is set to 1%. The results are shown in Figure 7(d). As the number of edges increases, the graph becomes more complex. The cost of storing and searching the graph also increases accordingly. As shown in the figure, both algorithms and the construction of ADI are linearly scalable. We also test the effects of other parameters. The experimental results show that both gSpan and ADI-Mine are not 3 Please refer to Figures 5(b) and 5(c) in the UIUC technical report version of [11]. sensitive to I--the average number of edges in potentially frequent graph patterns--and L--the number of potentially frequent kernels. The construction time and space cost of ADI structures are also stable. The reason is that the effects of those two parameters on the distribution in the data sets are minor. Similar observations have been reported by previous studies on mining frequent itemsets and sequential patterns. Limited by space, we omit the details here. 6.3 Mining Disk-based Databases Now, we report the experimental results on mining large, disk-based databases. In this set of experiments, we reserve a block of main memory of fixed size for ADI structure. When the size is too small for the ADI-structure, some levels of the ADI structure are accommodated on disk. On the other hand, we do not confine the memory usage for gSpan. 6.3.1 Scalability on Database Size We test the scalability of both gSpan and ADI-Mine on the size of databases. We use data set D100k-1mN 30I5T 20L200. The number of graphs in the database is varied from 100 thousand to 1 million. The main memory block for ADI structure is limited to 250M. The results are shown in Figure 8(a). The construction time of ADI structure is also plotted. Please note that the Y -axis is in logarithmic scale. The construction runtime of ADI structure is approximately linear on the database size. That is, the construction of the ADI index is highly scalable. We also measure the size of ADI structure. The results are shown in Figure 8(b). We can observe that the size of the ADI structure is linear to the database size. In this experiment, the ratio size of ADI structure in megabytes number of graphs in thousands is about 0.6. When the database size is 1 million, the size of ADI structure is 601M, which exceeds the main memory size of our machine. Even in such case, the construction runtime is still linear. As explained before, the construction of ADI structure makes sequential scans of the database and conducts a sequential write of the adjacency information. The overhead of construction of edge table and the linked lists of graph-ids is relatively small and thus has a minor effect on the construction time. While gSpan can handle databases of only up to 300 thousand graphs in this experiment, ADI-Mine can handle databases of 1 million graphs. The curve of the runtime of ADI-Mine can be divided into three stages. First, when the database has up to 300 thousand graphs, the ADI structure can be fully accommodated in main memory . ADI-Mine is faster than gSpan. Second, when the database has 300 to 600 thousand graphs, gSpan cannot finish. The ADI structure cannot be fully held in main memory. Some part of the adjacency information is put on disk. We see a significant jump in the runtime curve of ADI-Mine between the databases of 300 thousand graphs and 400 thousand graphs. Last, when the database has 800 thousand or more graphs, even the linked lists of graph-ids cannot be fully put into main memory. Thus, another significant jump in the runtime curve can be observed. 6.3.2 Tradeoff Between Efficiency and Main Memory Consumption It is interesting to examine the tradeoff between efficiency and size of available main memory. We use data set 323 Research Track Paper 0 200 400 600 800 1000 1200 0 2 4 6 8 10 Runtime (second) min_sup (%) gSpan ADI-Mine 0 200 400 600 800 1000 1200 50 55 60 65 70 75 80 85 90 95 100 Runtime (second) Number of graphs (thousand) gSpan ADI-Mine ADI-construction 10 100 1000 10000 100000 20 25 30 35 40 45 50 Runtime (second) N gSpan ADI-Mine ADI-construction 0 500 1000 1500 2000 10 15 20 25 30 Runtime (second) T gSpan ADI-Mine ADI-construction (a) scalability on min sup (b) Scalability on size (c) Scalability on N (d) Scalability on T D100kN 30I5T 20L200 D50-100kN 30I5T 20L200 D100kN 20-50I5T 20L200 D100kN 30I5T 10-30L200 min sup = 1% min sup = 1% min sup = 1% Figure 7: The experimental results of mining main memory-based databases. 100 1000 10000 100000 100 200 300 400 500 600 700 800 900 1000 Runtime (second) Number of graphs (thousand) gSpan ADI-Mine ADI-construction 0 100 200 300 400 500 600 700 100 200 300 400 500 600 700 800 900 1000 Size (M) Number of graphs (thousand) ADI structure 0 200 400 600 800 1000 0 20 40 60 80 100 120 140 160 Runtime (s) Size of available main memory (M) ADI-Mine (a) scalability on size (b) Size of ADI structure (c) Runtime vs. main memory D100k-1mN 30I5T 20L200 D100k-1mN 30I5T 20L200 D100kN 30I5T 20L200 min sup = 1% min sup = 1% min sup = 1% Figure 8: The experimental results of mining large disk-based databases. D100kN 30I5T 20L200, set the minimum support threshold to 1%, vary the main memory limit from 10M to 150M for ADI structure, and measure the runtime of ADI-Mine. The results are shown in Figure 8(c). In this experiment, the size of ADI structure is 57M. The construction time is 86 seconds. The highest watermark of main memory usage for gSpan in mining this data set is 87M. gSpan uses 1161 seconds in the mining if it has sufficient main memory. When the ADI structure can be completely loaded into main memory (57M or larger), ADI-Mine runs fast. Further increase of the available main memory cannot reduce the runtime. When the ADI structure cannot be fully put into main memory, the runtime increases. The more main memory, the faster ADI-Mine runs. When the available main memory is too small to even hold the linked lists of graph-ids, the runtime of ADI-Mine increases substantially. However, it still can finish the mining with 10M main memory limit in 2 hours. 6.3.3 Number of Disk Block Reads In addition to runtime, the efficiency of mining large disk-based databases can also be measured by the number of disk block read operations. Figure 9(a) shows the number of disk block reads versus the minimum support threshold. When the support threshold is high (e.g., 9% or up), the number of frequent edges is small. The ADI structure can be held into main memory and thus the I/O cost is very low. As the support threshold goes down, larger and larger part of the ADI structure is stored on disk, and the I/O cost increases. This curve is consistent with the trend in Figure 7(a). Figure 9(b) shows the number of disk block reads versus the number of graphs in the database. As the database size goes up, the I/O cost increases exponentially. This explains the curve of ADI-Mine in Figure 8(a). We also test the I/O cost on available main memory. The result is shown in Figure 9(c), which is consistent with the trend of runtime curve in Figure 8(c). 6.3.4 Effects of Other Parameters We also test the effects of the other parameters on the efficiency. We observe similar trends as in mining memory-based databases. Limited by space, we omit the details here. 6.4 Summary of Experimental Results The extensive performance study clearly shows the following . First, both gSpan and ADI-Mine are scalable when database can be held into main memory. ADI-Mine is faster than gSpan. Second, ADI-Mine can mine very large graph databases by accommodating the ADI structure on disk. The performance of ADI-Mine on mining large disk-based databases is highly scalable. Third, the size of ADI structure is linearly scalable with respect to the size of databases. Fourth, we can control the tradeoff between the mining efficiency and the main memory consumption. Last, ADI-Mine is more scalable than gSpan in mining complex graphs--the graphs that have many different kinds of labels. RELATED WORK The problem of finding frequent common structures has been studied since early 1990s. For example, [1, 7] study the the problem of finding common substructures from chemical compounds. SUBDUE [4] proposes an approximate algorithm to identify some, instead of the complete set of, 324 Research Track Paper 0 100000 200000 300000 400000 500000 600000 700000 0 2 4 6 8 10 Number of blocks read min_sup (%) ADI-Mine 0 5e+07 1e+08 1.5e+08 2e+08 2.5e+08 0 200 400 600 800 1000 Number of blocks read Number of graphs (thousand) ADI-Mine 0 5e+06 1e+07 1.5e+07 2e+07 2.5e+07 3e+07 3.5e+07 4e+07 0 20 40 60 80 100 120 140 160 Number of blocks read Size of available main memory (M) ADI-Mine (a) # blocks vs. support threshold (b) # blocks vs. database size (c) # blocks vs. main memory size D100kN 30I5T 20L200 D100k-1mN 30I5T 20L200 D100kN 30I5T 20L200 min sup = 1% min sup = 1% Figure 9: The number of disk blocks read in the mining. frequent substructures. However, these methods do not aim at scalable algorithms for mining large graph databases. The problem of mining the complete set of frequent graph patterns is firstly explored by Inokuchi et al. [5]. An Apriori-like algorithm AGM is proposed. Kuramochi and Karypis [6] develop an efficient algorithm, FSG, for graph pattern mining . The major idea is to utilize an effective graph representation , and conduct the edge-growth mining instead of vertex-growth mining. Both AGM and FSG adopt breadth-first search. Recently, Yan and Han propose the depth-first search approach , gSpan [11] for graph mining. They also investigate the problem of mining frequent closed graphs [9], which is a non-redundant representation of frequent graph patterns. As a latest result, Yan et al. [10] uses frequent graph patterns to index graphs. As a special case of graph mining, tree mining also receives intensive research recently. Zaki [12] proposes the first algorithm for mining frequent tree patterns. Although there are quite a few studies on the efficient mining of frequent graph patterns, none of them addresses the problem of effective index structure for mining large disk-based graph databases. When the database is too large to fit into main memory, the mining becomes I/O bounded, and the appropriate index structure becomes very critical for the scalability. CONCLUSIONS In this paper, we study the problem of scalable mining of large disk-based graph database. The ADI structure, an effective index structure, is developed. Taking gSpan as a concrete example, we propose ADI-Mine, an efficient algorithm adopting the ADI structure, to improve the scalability of the frequent graph mining substantially. The ADI-Mine structure is a general index for graph mining . As future work, it is interesting to examine the effect of the index structure on improving other graph pattern mining methods, such as mining frequent closed graphs and mining graphs with constraints. Furthermore, devising index structures to support scalable data mining on large disk-based databases is an important and interesting research problem with extensive applications and industrial values. Acknowledgements We are very grateful to Mr. Xifeng Yan and Dr. Jiawei Han for kindly providing us the executable of gSpan and answering our questions promptly. We would like to thank the anonymous reviewers for their insightful comments, which help to improve the quality of the paper. REFERENCES [1] D.M. Bayada, R. W. Simpson, and A. P. Johnson. An algorithm for the multiple common subgraph problem. J. of Chemical Information & Computer Sci., 32:680685, 1992. [2] C. Borgelt and M.R. Berthold. Mining molecular fragments: Finding relevant substructures of molecules. In Proc. 2002 Int. Conf. Data Mining (ICDM'02), Maebashi TERRSA, Maebashi City, Japan, Dec. 2002. [3] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. Introduction to Algorithms, Second Edition. MIT Press and McGraw-Hill, 2002. [4] L. B. Holder, D. J. Cook, and S. Djoko. Substructure discovery in the subdue system. In Proc. AAAI'94 Workshop Knowledge Discovery in Databases (KDD'94), pages 359370, Seattle, WA, July 1994. [5] A. Inokuchi, T. Washio, and H. Motoda. An apriori-based algorithm for mining frequent substructures from graph data. In Proc. 2000 European Symp. Principle of Data Mining and Knowledge Discovery (PKDD'00), pages 1323, Lyon, France, Sept. 2000. [6] M. Kuramochi and G. Karypis. Frequent subgraph discovery. In Proc. 2001 Int. Conf. Data Mining (ICDM'01), pages 313320, San Jose, CA, Nov. 2001. [7] Y. Takahashi, Y. Satoh, and S. Sasaki. Recognition of largest common fragment among a variety of chemical structures. Analytical Sciences, 3:2338, 1987. [8] N. Vanetik, E. Gudes, and S.E. Shimony. Computing frequent graph patterns from semistructured data. In Proc. 2002 Int. Conf. Data Mining (ICDM'02), Maebashi TERRSA, Maebashi City, Japan, Dec. 2002. [9] X. Yan and J. Han. Closegraph: Mining closed frequent graph patterns. In Proceedings of the 9th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD'03), Washington, D.C, 2003. [10] X. Yan, P.S. Yu, and J. Han. Graph indexing: A frequent structure-based approach. In Proc. 2004 ACM SIGMOD Int. Conf. on Management of Data (SIGMOD'04), Paris, France, June 2004. [11] Y. Yan and J. Han. gspan: Graph-based substructure pattern mining. In Proc. 2002 Int. Conf. on Data Mining (ICDM'02), Maebashi, Japan, December 2002. [12] M.J. Zaki. Efficiently mining frequent trees in a forest. In Proc. 2002 Int. Conf. on Knowledge Discovery and Data Mining (KDD'02), Edmonton, Alberta, Canada, July 2002. 325 Research Track Paper

index;Edge table;Graph mining;Subgraph mine;Frequent graph pattern mining;Adjacency list representation;graph database;DFS code;ADI Index structure;frequent graph pattern;Gspan algorithm;Disk bases databases;GRaph databases;Memory based databases

174

Secure Access to IP Multimedia Services Using Generic Bootstrapping Architecture (GBA) for 3G & Beyond Mobile Networks

The IP Multimedia Subsystem (IMS) defined by Third Generation Partnership Projects (3GPP and 3GPP2) is a technology designed to provide robust multimedia services across roaming boundaries and over diverse access technologies with promising features like quality-of-service (QoS), reliability and security. The IMS defines an overlay service architecture that merges the paradigms and technologies of the Internet with the cellular and fixed telecommunication worlds. Its architecture enables the efficient provision of an open set of potentially highly integrated multimedia services, combining web browsing, email, instant messaging, presence, VoIP, video conferencing, application sharing, telephony, unified messaging, multimedia content delivery, etc. on top of possibly different network technologies. As such IMS enables various business models for providing seamless business and consumer multimedia applications. In this communication converged world, the challenging issues are security, quality of service (QoS) and management & administration. In this paper our focus is to manage secure access to multimedia services and applications based on SIP and HTTP on top of IP Multimedia Subsystem (IMS). These services include presence, video conferencing, messaging, video broadcasting, and push to talk etc. We will utilize Generic Bootstrapping Architecture (GBA) model to authenticate multimedia applications before accessing these multimedia services offered by IMS operators. We will make enhancement in GBA model to access these services securely by introducing Authentication Proxy (AP) which is responsible to implement Transport Layer Security (TLS) for HTTP and SIP communication. This research work is part of Secure Service Provisioning (SSP) Framework for IP Multimedia System at Fokus Fraunhofer IMS 3Gb Testbed.

Introduction With the emergence of mobile multimedia services, such as unified messaging, click to dial, across network multiparty conferencing and seamless multimedia streaming services, the convergence of networks (i.e. fixedmobile convergence and voicedata integration) has started, leading to an overall Internet Telecommunications convergence. In prospect of these global trends, the mobile communications world has defined within the evolution of cellular systems an All-IP Network vision which integrates cellular networks and the Internet. This is the IP Multimedia System (IMS) [1], namely overlay architecture for the provision of multimedia services, such as VoIP (Voice over Internet Protocol) and videoconferencing on top of globally emerging 3G (Third Generation) broadband packet networks. The IP Multimedia System (IMS) which is standardized by Third Generation Partnership Project (3GPP & 3GGP2) in releases 5 is an overlay network on top of GPRS/UMTS (General Packet Radio Systems/Universal Mobile Telecommunication Systems) networks and extended by ETSI TISPAN [2] for fixed line access network within the Next Generation Network (NGN) architecture. The IMS provides all IP Service Delivery Platform (SDP) for mobile multimedia services provisioning e.g. VoIP, Video-telephony , Multimedia conferencing, Mobile Content, Push-to-Talk etc. and it is based on IETF protocols like SIP for session control, Diameter for AAA (Authentication, Authorization, and Auditing) and SDP (Service Delivery Protocol), RTP etc. Different components and parts of IMS are highlighted in figure 1 consisting IMS Core (P-CSCF, I-CSCF, S-CSCF), IMS Client (UE) and Application & Media Servers along with the concept of home network and visited network for roaming users on top of different access networks technologies. The security and data privacy is a big challenge when there is integration of different networks and technologies. The integration of different access technologies causes much vulnerability and hackers get access to steal financial and confidential information. As these hackers networks are often beyond the law enforcement agencies of the today's communication world. So the question arises how to prevent these hackers for performing such attacks on the corporate networks. In order to provide confidentiality, security and privacy, the 3G authentication infrastructure is a valuable and milestone development and asset for 3G operators. This infrastructure consists of authentication centre (AuC), the USIM (Universal Subscriber Identity Module) or ISIM (IP Multimedia Services Identity Module) and AKA (Authentication and Key Agreement) Procedure. It has recognized that this infrastructure could utilize to enable application function in the network and on the user side to enable shared keys. Therefore, Third Generation Partnership Project (3GPP) has provided the bootstrapping of application security to authenticate the subscriber by defining a Generic Bootstrapping Architecture (GBA) [3] based on Authentication and Key Agreement (AKA) protocol. The GBA model can be utilized to authenticate subscriber before accessing multimedia services and applications over HTTP. The candidate applications to use this bootstrapping mechanism include but are not restricted to subscriber certificate distribution. These certificates supports services including presence, conferencing, messaging and push to talk etc. provided by mobile operators. The GBA model has enhanced by implementing Generic Authentication Architecture (GAA) [4] to provide secure assess over HTTP using TLS (Transport Layer Security). In prospective of the advancement of telecommunication, the Fraunhofer Fokus established a Third Generation & beyond (3Gb) Testbed and IMS Testbed [5] for research & development and educational purpose to provide state-of-the-art knowledge to engineers, researchers, educationists and technologists in this area of modern telecommunication. Fokus Fraunhofer has developed a Secure Service Provisioning (SSP) Framework [6] for IMS Testbed to provide security, privacy and authentication of subscriber as well as confidential and protection to the network resources of 3G operators. The paper is organised as: section 2 is about IMS as platform for multimedia services, sections 3, 4 and 5 explain generic bootstrapping architecture, bootstrapping authentication procedure and its application usage procedure respectively. Section 6 discusses the use of authentication proxy for implementing TLS for securing multimedia services. In section 7, we will discus briefly the IMS Testbed at Fokus and than concludes the paper in last section. IMS - Platform for Next Generation Multimedia Services The IMS defines service provision architecture, and it can be considered as the next generation service delivery platform. It consists of modular design with open interfaces and enables the flexibility for providing multimedia services over IP technology. The IMS does not standardize specific services but uses standard service enablers e.g. presence, GLMS/XDMS etc. and supports inherently multimedia over IP, VoIP, Internet Multimedia and presence. In IMS architecture, SIP protocol use as the standard signaling protocol that establishes controls, modifies and terminates voice, video and messaging sessions between two or more participants. The related signaling servers in the architecture are referred to as Call State Control Functions (CSCFs) and distinguished by their specific functionalities. It is important to note that an IMS compliant end user system has to provide the necessary IMS protocol support, namely SIP, and the service related media codecs for multimedia applications in addition to basic connectivity support, e.g. GPRS, WLAN, etc. The IMS is designed to provide number of key capabilities required to enable new IP services via mobile and fixed networks. The important key functionalities which enable new mobile IP services are: Multimedia session negotiation and management Quality of service management Mobility management Service execution, control and interaction Privacy and security management Figure 1:- IP Multimedia Subsystem (IMS) Architecture In IMS specification, Application Server (AS) provides the service logic and service creation environment for applications and services. The AS is intended to influence and maintain the various IMS SIP sessions on behalf of the services. It can behave as a termination point for signaling, redirecting or forwarding SIP requests. It also can act as third party call control unit. Services in this instance refer to IMS services, which are based on the IMS reference points (e.g. instant messaging, presence, conferencing etc.). The advantage of application server is to enable IMS to operate in a more flexible and dynamic way, whereas the AS provides more intelligence to the system. Most Application Servers are closed boxes which map network functions (e.g. via OSA gateways) or signaling protocols (SIP) onto application programming interfaces based on a particular technology (Java, 18 CORBA, web-services). An alternative approach pursued by the Open Mobile Alliance (OMA) is strongly related to the service oriented methodology, which follows the top-down approach beginning with service design down to service mapping over the underlying network technologies. The SIP services can be developed and deployed on a SIP application server using several technologies such as SIP servlets, Call Processing Language (CPL) script, SIP Common Gateway Interface (CGI) and JAIN APIs. Generic Bootstrapping Architecture (GBA) Different 3G Multimedia Services including video conferencing, presence, push to talk etc. has potential usage of Generic Bootstrapping Architecture (GBA) to distribute subscriber certificates. These certificates are used by mobile operators to authenticate the subscriber before accessing the multimedia services and applications. Now we discuss components, entities and interfaces of GBA. 3.1 GBA Components and Entities The GBA consists of five entities: UE (User Equipment), NAF (Network Authentication Function), BSF (Bootstrapping Server Function) and HSS (Home Subscriber Server) and are explained below as specified in 3GPP standards (shown in figure 2). User Equipment: UE is UICC (Universal Integrated Circuit Card) containing USIM or ISIM related information that supports HTTP Digest AKA (Authentication & Key Agreement) and NAF (Network Authentication Function) specific protocols. A USIM (Universal Subscriber Identity Module) is an application for UMTS mobile telephony running on a UICC smartcard which is inserted in a 3G mobile phone. It stores user subscriber information, authentication information and provides with storage space for text messages. An IP Multimedia Services Identity Module (ISIM) is an application running on a UICC smartcard in a 3G telephone in the IP Multimedia Subsystem (IMS). It contains parameters for identifying and authenticating the user to the IMS. The ISIM application can co-exist with SIM and USIM on the same UICC making it possible to use the same smartcard in both GSM networks and earlier releases of UMTS. Bootstrapping Server Function (BSF): It hosts in a network element under the control of mobile network operator. The BSF, HSS, and UEs participate in GBA in which a shared secret is established between the network and a UE by running the bootstrapping procedure. The shared secret can be used between NAFs and UEs, for example, for authentication purposes. A generic Bootstrapping Server Function (BSF) and the UE shall mutually authenticate using the AKA protocol, and agree on session keys that are afterwards applied between UE and a Network Application Function (NAF). The BSF shall restrict the applicability of the key material to a specific NAF by using the key derivation procedure. The key derivation procedure may be used with multiple NAFs during the lifetime of the key material. The lifetime of the key material is set according to the local policy of the BSF. The BSF shall be able to acquire the GBA User security Settings (GUSS) from HSS [3]. Figure 2: Network Entities of GBA Network Authentication Function: NAF has the functionality to locate and communicate securely with subscriber's BSF (Bootstrapping Server Function). It should be able to acquire a shared key material established between the UE and the BSF during application specific protocol runs. Home Subscriber Server: HSS stores GBA user security settings (GUSSs). The GUSS is defined in such a way that interworking of different operators for standardized application profiles is possible. It also supports operator specific application profiles without the standardized of existing application profiles. The GUSS shall be able to contain application-specific USSs that contain parameters that relates to key selection indication, identification or authorization information of one or more applications hosted by one ore more NAFs. Any other types of parameters are not allowed in the application-specific USS [3]. Diameter-Proxy: In case where UE has contacted NAF of visited network than home network, this visited NAF will use diameter proxy (D-Proxy) of NAFs network to communicate with subscriber's BSF (i.e. home BSF). D-Proxy's general functionality requirements include [3]: D-Proxy functions as a proxy between visited NAF and subscriber's home BSF and it will be able to locate subscriber's home BSF and communicate with it over secure channel. The D-Proxy will be able to validate that the visited NAF is authorized to participate in GBA and shall be able to assert to subscriber's home BSF the visited NAFs DNS name. The D-Proxy shall also be able to assert to the BSF that the visited NAF is authorized to request the GBA specific user profiles contained in the NAF request. 19 Figure 3: Bootstrapping Authentication Procedure 3.2 GBA Reference Points Ub: The reference point Ub is between the UE and the BSF and provides mutual authentication between them. It allows the UE to bootstrap the session keys based on 3GPP AKA infrastructure. The HTTP Digest AKA protocol is used on the reference point Ub. It is based on the 3GPP AKA [7] protocol. Ua: The reference point Ua carries the application protocol, which is secured using the keys material agreed between UE and BSF as a result of running of HTTP Digest AKA over reference point Ub. For instance, in case of support for subscriber certificates, it is a protocol, which allows the user to request certificates from NAF. In this case, NAF would be PKI portal. Zh: The reference point Zh used between BSF and HSS. It allows BSF to fetch the required authentication information and all GBA user security settings from HSS. The interface to 3G Authentication Centre is HSS-internal, and it need not be standardised as part of this architecture. Zn: The reference point Zn is used by the NAF to fetch the key material agreed during a previous HTTP Digest AKA protocol run over the reference point Ub from the UE to the BSF. It is also used to fetch application-specific user security settings from the BSF, if requested by the NAF. Bootstrapping Authentication Procedure The UE and Network Authentication Function (NAF) have to decide whether to use GBA before the start of communication between them. When UE wants to interact with NAF, it starts communication with NAF over Ua interface without GBA parameters. If NAF requires the use of shared keys obtained by means of GBA, but the request from UE does not include GBA-related parameters, the NAF replies with a bootstrapping initiation message [3]. When UE wants to interact with NAF, and it knows 20 that the bootstrapping procedure is needed, it shall first perform a bootstrapping authentication as shown in figure 3. Otherwise, the UE shall perform a bootstrapping authentication only when it has received bootstrapping initiation required message or a bootstrapping negotiation indication from the NAF, or when the lifetime of the key in UE has expired. The UE sends an HTTP request to the BSF and the BSF retrieves the complete set of GBA user security settings and one Authentication Vector (AV) [8] as given in equation 1 over the reference point Zh from the HSS. AV = RAND||AUTN||XRES||CK||IK ------------------- Eq. 1 After that BSF forwards the RAND and AUTN to the UE in the 401 message without the CK, IK and XRES. This is to demand the UE to authenticate itself. The UE checks AUTN to verify that the challenge is from an authorized network; the UE also calculates CK, IK and RES [8]. This will result in session keys IK and CK in both BSF and UE. The UE sends another HTTP request to the BSF, containing the Digest AKA response which is calculated using RES. The BSF authenticates the UE by verifying the Digest AKA response. The BSF generates key material Ks by concatenating CK and IK and it also generates B-TID (Bootstrapping Transaction Identifier) which is used to bind the subscriber identity to the keying material in reference points Ua, Ub and Zn. The BSF shall send a 200 OK message, including a B-TID to the UE to indicate the success of the authentication and the lifetime of the key Ks. The key material Ks is generated in UE by concatenating CK and IK. Both the UE and the BSF shall use the Ks to derive the key material Ks-NAF which will be used for securing the reference point Ua. The Ks-NAF is computed as equation 2. Ks-NAF = f KD (Ks, &quot;gba-me&quot;, RAND, IMPI, NAF-ID) ----- Eq. 2 where f KD is the key derivation function and will be implemented in the ME, and the key derivation parameters consist of user's IMPI, NAF-ID and RAND. The NAF-ID consists of the full DNS name of the NAF, concatenated with the Ua security protocol identifier. The UE and the BSF shall store the key Ks with the associated B-TID for further use, until the lifetime of Ks has expired, or until the key Ks is updated [3]. Bootstrapping Usage Procedure Before communication between the UE and the NAF can start, the UE and the NAF first have to agree whether to use shared keys obtained by means of the GBA. If the UE does not know whether to use GBA with this NAF, it uses the initiation of bootstrapping procedure. Once the UE and the NAF have decided that they want to use GBA then every time the UE wants to interact with NAF. The UE starts communication over reference point Ua with the NAF by supplying the B-TID to the NAF to allow the NAF to retrieve the corresponding keys from the BSF. The NAF starts communication over reference point Zn with BSF. The NAF requests key material corresponding to the B-TID supplied by the UE to the NAF over reference point Ua. With the key material request, the NAF shall supply NAF's public hostname that UE has used to access NAF to BSF, and BSF shall be able to verify that NAF is authorized to use that hostname. The NAF may also request one or more application-specific USSs for the applications, which the request received over Ua from UE may access. The BSF derives the keys required to protect the protocol used over reference point Ua from the key Ks and the key derivation parameters. Than it supplies requested key Ks-NAF, bootstrapping time and the lifetime of the key to NAF. If the key identified by the B-TID supplied by the NAF is not available at the BSF, the BSF shall indicate this in the reply to the NAF. The NAF then indicates a bootstrapping renegotiation request to the UE. The BSF may also send the private user identity (IMPI) and requested USSs to NAF according to the BSF's policy. The NAF continues with the protocol used over the reference point Ua with the UE. Once the run of the protocol used over reference point Ua is completed the purpose of bootstrapping is fulfilled as it enabled UE and NAF to use reference point Ua in a secure way. Figure 4: Bootstrapping Application Authentication Proxy Usage for Multimedia Services Authentication Proxy (AP) is like a Network Authentication Function (NAF) and performs the function of HTTP proxy for the UE. It is responsible to handle the Transport Layer Security (TLS) and implement the secure HTTP channel between AP and UE as shown in figure 5. It utilizes the generic bootstrapping architecture to assure the application servers (ASs) that the request is coming from an authorized subscriber of mobile 21 network operator. When HTTPS request is sent to AS through AP, the AP performs UE authentication. The AP may insert the user identity when it forwards the request to application server. Figure 5b presents the architecture view of using AP for different IMS SIP services e.g. presence, messaging, conferencing etc. Figure 5: Authentication Proxy The UE shall manipulate own data such as groups, through the Ua/Ut reference point [4]. The reference point Ut will be applicable to data manipulation of IMS based SIP services, such as Presence, Messaging and Conferencing services. When the HTTPS client starts communication via Ua reference point with the NAF, it shall establish a TLS tunnel with the NAF. The NAF is authenticated to the HTTPS client by means of a public key certificate. The HTTPS client will verify that the server certificate corresponds to the FQDN (Fully Qualified Domain Name) of the AP it established the tunnel with. We explain the procedure briefly as: The HTTPS client sends an HTTP request to NAF inside the TLS tunnel. In response to HTTP request over Ua interface, the AP will invoke HTTP digest with HTTPS client in order to perform client authentication using the shared keys. On the receipt of HTTPS digest from AP, the client will verify that the FDQN corresponds the AP it established the TLS connection with, if not the client will terminate the TLS connection with the AP. In this way the UE and AP are mutually authenticated as the TLS tunnel endpoints. Now we discuss an example that application residing on UICC (Universal Integrated Circuit Card) may use TLS over HTTP in Generic Authentication Architecture (GAA) mechanism to secure its communication with Authentication Proxy (AP). The GBA security association between a UICC-based application and AP could establish as: Figure 6: HTTPS and BIP (Bearer Independent Protocol) Procedures The ME (Mobile Equipment) executes the bootstrapping procedure with the BSF supporting the Ub reference point. The UICC, which hosts the HTTPS client, runs the bootstrapping usage procedure with AP supporting the Ua reference point [9]. Figure 6 shows the use of BIP (Bearer Independent Protocol) to establish the HTTPS connection between UICC and AP. When UICC opens channel with AP as described in [10] than an active TCP/IP connection establishes between UICC and AP. Fokus IMS Testbed In face of the current challenges within telecommunications market are mainly consequences of insufficient early access to new enabling technologies by all market players. , In this development Fraunhofer Institute FOKUS, known as a leading research institute in the field of open communication systems, has established with support of German Ministry of Education and Research (BMBF) a 3G beyond Testbed, known as "National Host for 3Gb Applications". This Testbed provides technologies and related know-how in the field of fixed and wireless next generation network technologies and related service delivery 22 platforms. As a part of 3Gb Testbed, the FOKUS Open IMS Playground is deployed as an open technology test field with the target to validate existing and emerging IMS standards and to extend the IMS appropriately to be used on top of new access networks as well as to provide new seamless multimedia applications [11]. All major IMS core components, i.e., x-CSCF, HSS, MG, MRF, Application Servers, Application Server Simulators, service creation toolkits, and demo applications are integrated into one single environment and can be used and extended for R&D activities by academic and industrial partners. All these components can be used locally on top of all available access technologies or can be used over IP tunnels remotely. Users of the "Open IMS playground" can test their components performing interoperability tests. The SIP Express Router (SER), one of the fastest existing SIP Proxies, can be used as a reference implementation and to proof interoperability with other SIP components [11]. The major focal point of IMS Playground is to put Application Server aside. Varieties of platforms enable rapid development of innovative services. Figure 7: Fokus IMS Testbed The Open IMS playground is deployed as an open technology test field with the target to develop prototype and validate existing and emerging NGN/IMS standard components. It extends the IMS architecture and protocols appropriately to be used on top of new access networks as well as to provide new seamless multimedia applications. It is important to stress that all components have been developed by FOKUS as reference implementations, such as an own open source IMS core system (to be publicly released in 2006 based on the famous SIP Express Router), IMS Clients and application servers (SIPSee), and HSS. The IMS playground is used on the one hand as the technology basis for own industry projects performed for national and international vendors and network operators as well as for more mid term academic R&D projects in the European IST context. In addition, the playground is used by others as well, i.e. FOKUS is providing consultancy and support services around the IMS playground. Users of the "Open IMS playground", e.g. vendors, are testing their components performing interoperability and benchmarking tests. Application developers are developing new IMS applications based on various programming platforms provided, i.e. IN/CAMEL, OSA/Parlay, JAIN, SIP Servlets, etc., and gain a proof of concept implementation.. The different platform options, each with their strengths and weaknesses, can be selected and used according to the customers' needs. Figure 7 displays the Open IMS playground partner components. Conclusion In this paper, we have presented the architecture of secure access and authentication of IP Multimedia Services based of SIP and HTTP communication using GBA (Generic Bootstrapping Architecture) as recommended by 3GPP and TISPAN as a part of Secure Service Provisioning (SSP) Framework of IMS at Fokus Fraunhofer IMS and 3Gb Testbed. REFERENCES [1] Third Generation Partnership Project; Technical Specification Group Services and System Aspects; TS 23.228 IP Multimedia Subsystem (IMS), Stage 2 / 3GPP2 X.S0013-002-0 v1.0, www.3gpp.org. [2] ETSI TISPAN (Telecommunications and Internet converged Services and Protocols for Advanced Networking) WG http://portal.etsi.org/tispan/TISPAN_ToR.asp. [3] Third Generation Partnership Project; Technical Specification Group Services and System Aspects; Generic Authentication Architecture (GAA); Generic Bootstrapping Architecture (GBA) (Release 7), 3GPP TS 33.220 V7 (2005). [4] Third Generation Partnership Project; Technical Specification Group Services and System Aspects; Generic Authentication Architecture (GAA); Access to Network Application Functions using Hypertext Transfer Protocol over Transport Layer Security (HTTPS) (Release 7), 3GPP TS 33.222 V7 (2005). [5] Third Generation & Beyond (3Gb) Testbed, www.fokus.fraunhofer.de/national_host & IP Multimedia System (IMS) Playground www.fokus.fraunhofer.de/ims. [6] M. Sher, T. Magedanz, "Secure Service Provisioning Framework (SSPF) for IP Multimedia System and Next Generation Mobile Networks" 3rd International Workshop in Wireless Security Technologies, London, U.K. (April 2005), IWWST'05 Proceeding (101-106), ISSN 1746-904X. [7] Third Generation Partnership Project; Technical Specification Group Services and System Aspects; 3G Security; Security Architecture (Release 6); 3GPP, TS 33.102 V6 (2004). [8] M. Sher, T. Magedanz: &quot;Network Access Security Management (NASM) Model for Next Generation Mobile Telecommunication Networks&quot;, IEEE/IFIP MATA'2005, 2 nd International Workshop on Mobility Aware Technologies and Applications - Service Delivery Platforms for Next Generation Networks, Montreal, Canada, October 17-19, 2005, Proceeding Springer-Verlag LNCS 3744-0263, Berlin 23 Heidelberg 2005, pp. 263-272. http://www.congresbcu.com/mata2005 [9] Third Generation Partnership Project; Technical Specification Group Services and System Aspects; Generic Authentication Architecture (GAA); Early Implementation of HTTPS Connection between a Universal Integrated Circuit Card (UICC) and Network Application Function (NAF) (Release 7), 3GPP TR 33.918 V7 (2005). [10] Third Generation Partnership Project; Technical Specification Group Core Network and Terminals; Universal Subscriber Identity Module (USIM) Application Toolkit (USAT) (Release 7), 3GPP TS 31.111 V7 (2005). [11] K. Knttel, T.Magedanz, D. Witszek: "The IMS Playground @ Fokus an Open Testbed for Next Generation Network Multimedia Services", 1 st Int. IFIP Conference on Testbeds and Research Infrastructures for the Development of Networks and Communities (Tridentcom), Trento, Italian, February 23 - 25, 2005, Proceedings pp. 2 11, IBSN 0-7695 -2219-x, IEEE Computer Society Press, Los Alamitos, California. Acronyms 3GPP Third Generation Partnership Project 3GPP2 Third Generation Partnership Project 2 AAA Authentication, Authorisation, and Accounting AKA Authentication and Key Agreement AP Authentication Proxy AS Application Server AuC Authentication Centre AV Authentication Function BGA Generic Bootstrapping Architecture BSF Bootstrapping Server Function B-TID Bootstrapping Transaction Identifier CAMEL Customized Applications for Mobile Enhanced Logic CGI Common Gateway Interface CK Cipher Key CORBA Common Object Request Broker Architecture CPL Call Programming Language CSCFs Call State Control Functions DNS Domain Name Server FMC Fixed Mobile Convergence FQDN Fully Qualified Domain Name GAA Generic Authentication Architecture GPRS General Packet Radio System GUSS GBA User Security Settings HSS Home Subscriber Server HTTP Hyper Text Transfer Protocol HTTPS HTTP Secure ( HTTP over TLS) ICSCF Interrogating Call State Control Function IETF Internet Engineering Task Force IK Integrity Key IM IP Multimedia IMPI IP Multimedia Private Identity IMS IP Multimedia Subsystem IN Intelligent Network IP Internet Protocol ISIM IM Service Identity Module Ks Session Key ME Mobile Equipment MG Media Gate MRF Media Resource Function NAF Network Authentication Function NGN Next Generation Network OMA Open Mobile Alliance OSA Open Service Access PCSCF Proxy Call State Control Function PDP Packet Data Protocol PoC PPT over Cellular PTT Push To Talk QoS Quality of Service RES Response RTP Real-time Transport Protocol SCSCF Serving Call State Control Function SDP Service Delivery Platform SER SIP Express Router SIP Session Initiation Protocol SSP Secure Service Provisioning TCP Transmission Control Protocol TISPAN Telecoms & Internet converged Services & Protocols for Advanced Networks TLS Transport Layer Security UE User Equipment UICC Universal Integrated Circuit Card UMTS Universal Mobile Telecommunication Standard USIM Universal Subscriber Identity Module WLAN Wireless Local Area Network 24

TLS Tunnel end points;Generic Authentication Architecture;GLMS/XDMS;General bootstrapping architecture;Transport Layer Security;Network authentication function;Signalling protocols;Generic Bootstrapping Architecture;Authentication Proxy;GBA;Diameter proxy;Transport layer security;IP Multimedia System;Authentication proxy;TLS;IP multimedia subsystem;IMS platform;Fokus IMS Testbed;NAF;AP;Security and Privacy

175

Secure Hierarchical In-Network Aggregation in Sensor Networks

In-network aggregation is an essential primitive for performing queries on sensor network data. However, most aggregation algorithms assume that all intermediate nodes are trusted. In contrast, the standard threat model in sensor network security assumes that an attacker may control a fraction of the nodes, which may misbehave in an arbitrary (Byzantine) manner. We present the first algorithm for provably secure hierarchical in-network data aggregation. Our algorithm is guaranteed to detect any manipulation of the aggregate by the adversary beyond what is achievable through direct injection of data values at compromised nodes. In other words, the adversary can never gain any advantage from misrepresenting intermediate aggregation computations. Our algorithm incurs only O(log2n) node congestion, supports arbitrary tree-based aggregator topologies and retains its resistance against aggregation manipulation in the presence of arbitrary numbers of malicious nodes. The main algorithm is based on performing the SUM aggregation securely by first forcing the adversary to commit to its choice of intermediate aggregation results, and then having the sensor nodes independently verify that their contributions to the aggregate are correctly incorporated. We show how to reduce secure MEDIAN , COUNT , and AVERAGE to this primitive.

INTRODUCTION Wireless sensor networks are increasingly deployed in security-critical applications such as factory monitoring, environmental monitoring , burglar alarms and fire alarms. The sensor nodes for these applications are typically deployed in unsecured locations and are not made tamper-proof due to cost considerations. Hence, an adversary could undetectably take control of one or more sensor nodes and launch active attacks to subvert correct network operations. Such environments pose a particularly challenging set of constraints for the protocol designer: sensor network protocols must be highly energy efficient while being able to function securely in the presence of possible malicious nodes within the network. In this paper we focus on the particular problem of securely and efficiently performing aggregate queries (such as MEDIAN , SUM and AVERAGE ) on sensor networks. In-network data aggregation is an efficient primitive for reducing the total message complexity of aggregate sensor queries. For example, in-network aggregation of the SUM function is performed by having each intermediate node forward a single message containing the sum of the sensor readings of all the nodes downstream from it, rather than forwarding each downstream message one-by-one to the base station. The energy savings of performing in-network aggregation have been shown to be significant and are crucial for energy-constrained sensor networks [9, 11, 20]. Unfortunately, most in-network aggregation schemes assume that all sensor nodes are trusted [12, 20]. An adversary controlling just a few aggregator nodes could potentially cause the sensor network to return arbitrary results, thus completely subverting the function of the network to the adversary's own purposes. Despite the importance of the problem and a significant amount of work on the area, the known approaches to secure aggregation either require strong assumptions about network topology or adversary capabilities, or are only able to provide limited probabilistic security properties. For example, Hu and Evans [8] propose a secure aggregation scheme under the assumption that at most a single node is malicious. Przydatek et al. [17] propose Secure Information Aggregation (SIA), which provides a statistical security property under the assumption of a single-aggregator model. In the single-aggregator model, sensor nodes send their data to a single aggregator node, which computes the aggregate and sends it to the base station. This form of aggregation reduces communications only on the link between the aggregator and the base station, and is not scalable to large multihop sensor deployments. Most of the algorithms in SIA (in particular, MEDIAN , SUM and AVERAGE ) cannot be directly adapted to a hierarchical aggregation model since 278 they involve sorting all of the input values; the final aggregator in the hierarchy thus needs to access all the data values of the sensor nodes. In this paper, we present the first provably secure sensor network data aggregation protocol for general networks and multiple adver-sarial nodes. The algorithm limits the adversary's ability to manipulate the aggregation result with the tightest bound possible for general algorithms with no knowledge of the distribution of sensor data values. Specifically, an adversary can gain no additional influence over the final result by manipulating the results of the in-network aggregate computation as opposed to simply reporting false data readings for the compromised nodes under its control. Furthermore, unlike prior schemes, our algorithm is designed for general hierarchical aggregator topologies and multiple malicious sensor nodes. Our metric for communication cost is congestion, which is the maximum communication load on any node in the network. Let n be the number of nodes in the network, and be the maximum degree of any node in the aggregation tree. Our algorithm induces only O (log 2 n ) node congestion in the aggregation tree. RELATED WORK Researchers have investigated resilient aggregation algorithms to provide increased likelihood of accurate results in environments prone to message loss or node failures. This class of algorithms includes work by Gupta et al. [7], Nath et al. [15], Chen et al. [3] and Manjhi et al. [14]. A number of aggregation algorithms have been proposed to ensure secrecy of the data against intermediate aggregators. Such algorithms have been proposed by Girao et al. [5], Castelluccia et al. [2], and Cam et al. [1]. Hu and Evans [8] propose securing in-network aggregation against a single Byzantine adversary by requiring aggregator nodes to forward their inputs to their parent nodes in the aggregation tree. Jadia and Mathuria [10] extend the Hu and Evans approach by incorporating privacy, but also considered only a single malicious node. Several secure aggregation algorithms have been proposed for the single-aggregator model. Przydatek et al. [17] proposed Secure Information Aggregation (SIA) for this topology. Also for the single-aggregator case, Du et al. [4] propose using multiple witness nodes as additional aggregators to verify the integrity of the aggregator's result. Mahimkar and Rappaport [13] also propose an aggregation-verification scheme for the single-aggregator model using a threshold signature scheme to ensure that at least t of the nodes agree with the aggregation result. Yang et al. [19] describe a probabilistic aggregation algorithm which subdivides an aggregation tree into subtrees, each of which reports their aggregates directly to the base station. Outliers among the subtrees are then probed for inconsistencies. Wagner [18] addressed the issue of measuring and bounding malicious nodes' contribution to the final aggregation result. The paper measures how much damage an attacker can inflict by taking control of a number of nodes and using them solely to inject erroneous data values. PROBLEM MODEL In general, the goal of secure aggregation is to compute aggregate functions (such as SUM , COUNT or AVERAGE ) of the sensed data values residing on sensor nodes, while assuming that a portion of the sensor nodes are controlled by an adversary which is attempting to skew the final result. In this section, we present the formal parameters of the problem. 3.1 Network Assumptions We assume a general multihop network with a set S = {s 1 ,...,s n } of n sensor nodes and a single (untrusted) base station R, which is able to communicate with the querier which resides outside of the network. The querier knows the total number of sensor nodes n, and that all n nodes are alive and reachable. We assume the aggregation is performed over an aggregation tree which is the directed tree formed by the union of all the paths from the sensor nodes to the base station (one such tree is shown in Figure 1(a)). These paths may be arbitrarily chosen and are not necessarily shortest paths. The optimisation of the aggregation tree structure is out of the scope of this paper--our algorithm takes the structure of the aggregation tree as given. One method for constructing an aggregation tree is described in TaG [11]. 3.2 Security Infrastructure We assume that each sensor node has a unique identifier s and shares a unique secret symmetric key K s with the querier. We further assume the existence of a broadcast authentication primitive where any node can authenticate a message from the querier. This broadcast authentication could, for example, be performed using TESLA [16]. We assume the sensor nodes have the ability to perform symmetric-key encryption and decryption as well as computations of a collision-resistant cryptographic hash function H. 3.3 Attacker Model We assume that the attacker is in complete control of an arbitrary number of sensor nodes, including knowledge of all their secret keys. The attacker has a network-wide presence and can record and inject messages at will. The sole goal of the attacker is to launch what Przydatek et al. [17] call a stealthy attack, i.e., to cause the querier to accept a false aggregate that is higher or lower than the true aggregate value. We do not consider denial-of-service (DoS) attacks where the goal of the adversary is to prevent the querier from getting any aggregation result at all. While such attacks can disrupt the normal operation of the sensor network, they are not as potentially hazardous in security-critical applications as the ability to cause the operator of the network to accept arbitrary data. Furthermore, any maliciously induced extended loss of service is a detectable anomaly which will (eventually) expose the adversary's presence if subsequent protocols or manual intervention do not succeed in resolving the problem. 3.4 Problem Definition and Metrics Each sensor node s i has a data value a i . We assume that the data value is a non-negative bounded real value a i [0,r] for some maximum allowed data value r. The objective of the aggregation process is to compute some function f over all the data values, i.e., f (a 1 ,...,a n ). Note that for the SUM aggregate, the case where data values are in a range [r 1 ,r 2 ] (where r 1 ,r 2 can be negative) is reducible to this case by setting r = r 2 - r 1 and add nr 1 to the aggregation result. Definition 1 A direct data injection attack occurs when an attacker modifies the data readings reported by the nodes under its direct control, under the constraint that only legal readings in [0,r] are reported. Wagner [18] performed a quantitative study measuring the effect of direct data injection on various aggregates, and concludes that the aggregates addressed in this paper (truncated SUM and AV ERAGE , COUNT and QUANTILE ) can be resilient under such attacks . 279 Without domain knowledge about what constitutes an anomalous sensor reading, it is impossible to detect a direct data injection attack, since they are indistinguishable from legitimate sensor readings [17, 19]. Hence, if a secure aggregation scheme does not make assumptions on the distribution of data values, it cannot limit the adversary's capability to perform direct data injection. We can thus define an optimal level of aggregation security as follows. Definition 2 An aggregation algorithm is optimally secure if, by tampering with the aggregation process, an adversary is unable to induce the querier to accept any aggregation result which is not already achievable by direct data injection. As a metric for communication overhead, we consider node congestion , which is the worst case communication load on any single sensor node during the algorithm. Congestion is a commonly used metric in ad-hoc networks since it measures how quickly the heaviest-loaded nodes will exhaust their batteries [6, 12]. Since the heaviest-loaded nodes are typically the nodes which are most essential to the connectivity of the network (e.g., the nodes closest to the base station), their failure may cause the network to partition even though other sensor nodes in the network may still have high battery levels. A lower communication load on the heaviest-loaded nodes is thus desirable even if the trade-off is a larger amount of communication in the network as a whole. For a lower bound on congestion, consider an unsecured aggregation protocol where each node sends just a single message to its parent in the aggregation tree. This is the minimum number of messages that ensures that each sensor node contributes to the aggregation result. There is (1) congestion on each edge on the aggregation tree, thus resulting in (d) congestion on the node(s) with highest degree d in the aggregation tree. The parameter d is dependent on the shape of the given aggregation tree and can be as large as (n) for a single-aggregator topology or as small as (1) for a balanced aggregation tree. Since we are taking the aggregation tree topology as an input, we have no control over d. Hence, it is often more informative to consider per-edge congestion, which can be independent of the structure of the aggregation tree. Consider the simplest solution where we omit aggregation altogether and simply send all data values (encrypted and authenticated ) directly to the base station, which then forwards it to the querier. This provides perfect data integrity, but induces O (n) congestion at the nodes and edges nearest the base station. For an algorithm to be practical, it must cause only sublinear edge congestion. Our goal is to design an optimally secure aggregation algorithm with only sublinear edge congestion. THE SUM ALGORITHM In this section we describe our algorithm for the SUM aggregate, where the aggregation function f is addition. Specifically, we wish to compute a 1 + + a n , where a i is the data value at node i. We defer analysis of the algorithm properties to Section 5, and discuss the application of the algorithm to other aggregates such as COUNT , AVERAGE and MEDIAN in Section 6. We build on the aggregate-commit-prove framework described by Przydatek et al. [17] but extend their single aggregator model to a fully distributed setting. Our algorithm involves computing a cryptographic commitment structure (similar to a hash tree) over the data values of the sensor nodes as well as the aggregation process . This forces the adversary to choose a fixed aggregation topology and set of aggregation results. The individual sensor nodes then independently audit the commitment structure to verify that their respective contributions have been added to the aggregate. If the adversary attempts to discard or reduce the contribution of a legitimate sensor node, this necessarily induces an inconsistency in the commitment structure which can be detected by the affected node. This basic approach provides us with a lower bound for the SUM aggregate. To provide an upper-bound for SUM , we can re-use the same lower-bounding approach, but on a complementary aggregate called the COMPLEMENT aggregate. Where SUM is defined as a i , COMPLEMENT is defined as (r - a i ) where r is the upper bound on allowable data values. When the final aggregates are computed, the querier enforces the constraint that SUM + COM PLEMENT = nr. Hence any adversary that wishes to increase SUM must also decrease COMPLEMENT , and vice-versa, otherwise the discrepancy will be detected. Hence, by enforcing a lower-bound on COMPLEMENT , we are also enforcing an upper-bound on SUM . The overall algorithm has three main phases: query dissemination , aggregation-commit, and result-checking. Query dissemination. The base station broadcasts the query to the network. An aggregation tree, or a directed spanning tree over the network topology with the base station at the root, is formed as the query is sent to all the nodes, if one is not already present in the network. Aggregation commit. In this phase, the sensor nodes iteratively construct a commitment structure resembling a hash tree. First, the leaf nodes in the aggregation tree send their data values to their parents in the aggregation tree. Each internal sensor node in the aggregation tree performs an aggregation operation whenever it has heard from all its child sensor nodes. Whenever a sensor node s performs an aggregation operation, s creates a commitment to the set of inputs used to compute the aggregate by computing a hash over all the inputs (including the commitments that were computed by the children of s). Both the aggregation result and the commitment are then passed on to the parent of s. After the final commitment values are reported to the base station (and thus also to the querier), the adversary cannot subsequently claim a different aggregation structure or result. We describe an optimisation to ensure that the constructed commitment trees are perfectly balanced, thus requiring low congestion overhead in the next phase. Result-checking. The result-checking phase is a novel distributed verification process. In prior work, algorithms have relied on the querier to issue probes into the commitment structure to verify its integrity [17, 19]. This induces congestion nearest the base station, and moreover, such algorithms yield at best probabilistic security properties. We show that if the verification step is instead fully distributed , it is possible to achieve provably optimal security while maintaining sublinear edge congestion. The result-checking phase proceeds as follows. Once the querier has received the final commitment values, it disseminates them to the rest of the network in an authenticated broadcast. At the same time, sensor nodes disseminate information that will allow their peers to verify that their respective data values have been incorporated into the aggregate. Each sensor node is responsible for checking that its own contribution was added into the aggregate. If a sensor node determines that its data value was indeed added towards the final sum, it sends an authentication code up the aggregation tree towards to the base station. Authentication codes are aggregated along the way with the XOR function for communication efficiency. When the querier has received the XOR of all the authentication codes, it can then verify that all the sensor nodes have confirmed that the aggregation structure is consistent with their data values. If so, then it accepts the aggregation result. We now describe the details of each of the three phases in turn. 280 (a) Example network graph. Arrows: Aggregation tree. R: Base station. Q: Querier. G 0 = 1,a G ,r - a G ,G F 1 = 2,v F 1 ,v F 1 ,H[N||2||v F 1 ||v F 1 ||F 0 ||G 0 ] C 1 = 4,v C 1 ,v C 1 ,H[N||4||v C 1 ||v C 1 ||C 0 ||E 0 ||F 1 ] A 1 = 9,v A 1 ,v A 1 ,H[N||9||v A 1 ||v A 1 ||A 0 ||B 1 ||C 1 ||D 0 ] R = 12,v R ,v R ,H[N||12||v R ||v R ||H 0 ||A 1 ||I 0 ] (b) Naive commitment tree, showing derivations of some of the vertices. For each sensor node X , X 0 is its leaf vertex, while X 1 is the internal vertex representing the aggregate computation at X (if any). On the right we list the labels of the vertices on the path of node G to the root. Figure 1: Aggregation and naive commitment tree in network context 4.1 Query Dissemination First, an aggregation tree is established if one is not already present. Various algorithms for selecting the structure of an aggregation tree may be used. For completeness, we describe one such process, while noting that our algorithm is directly applicable to any aggregation tree structure. The Tiny Aggregation Service (TaG) [11] uses a broadcast from the base station where each node chooses as its parent in the aggregation tree, the node from which it first heard the tree-formation message. To initiate a query in the aggregation tree, the base station originates a query request message which is distributed following the aggregation tree. The query request message contains an attached nonce N to prevent replay of messages belonging to a prior query, and the entire request message is sent using an authenticated broadcast . 4.2 Aggregation-Commit Phase The goal of the aggregation-commit phase is to iteratively construct a series of cryptographic commitments to data values and to intermediate in-network aggregation operations. This commitment is then passed on to the querier. The querier then rebroadcasts the commitment to the sensor network using an authenticated broadcast so that the rest of the sensor network is able to verify that their respective data values have been incorporated into the aggregate. 4.2.1 Aggregation-Commit: Naive Approach We first describe a naive approach that yields the desired security properties but has suboptimal congestion overhead when sensor nodes perform their respective verifications. In the naive approach, when each sensor node performs an aggregation operation, it computes a cryptographic hash of all its inputs (including its own data value). The hash value is then passed on to the parent in the aggregation tree along with the aggregation result. Figure 1(b) shows a commitment tree which consists of a series of hashes of data values and intermediate results, culminating in a set of final commitment values which is passed on by the base station to the querier along with the aggregation results. Conceptually, a commitment tree is a hash tree with some additional aggregate accounting information attached to the nodes. A definition follows. Recall that N is the query nonce that is disseminated with each query. Definition 3 A commitment tree is a tree where each vertex has an associated label representing the data that is passed on to its parent. The labels have the following format: count, value, complement, commitment Where count is the number of leaf vertices in the subtree rooted at this vertex; value is the SUM aggregate computed over all the leaves in the subtree; complement is the aggregate over the COMPLEMENT of the data values; and commitment is a cryptographic commitment. The labels are defined inductively as follows: There is one leaf vertex u s for each sensor node s, which we call the leaf vertex of s. The label of u s consists of count=1, value =a s where a s is the data value of s, complement=r - a s where r is the upper bound on allowable data values, and commitment is the node's unique ID. Internal vertices represent aggregation operations, and have labels that are defined based on their children. Suppose an internal vertex has child vertices with the following labels: u 1 ,u 2 ,...,u q , where u i = c i ,v i ,v i ,h i . Then the vertex has label c ,v,v,h , with c = c i , v = v i , v = v i and h = H[N||c||v||v||u 1 ||u 2 ||||u q ]. For brevity, in the remainder of the paper we will often omit references to labels and instead refer directly to the count, value, complement or commitment of a vertex. While there exists a natural mapping between vertices in a commitment tree and sensor nodes in the aggregation tree, a vertex is a logical element in a graph while a sensor node is a physical device . To prevent confusion, we will always refer to the vertices in the commitment tree; the term nodes always refers to the physical sensor node device. Since we assume that our hash function provides collision resistance , it is computationally infeasible for an adversary to change any of the contents of the commitment tree once the final commitment values have reached the root. With knowledge of the root commitment value, a node s may verify the aggregation steps between its leaf vertex u s and the root of the commitment tree. To do so, s needs the labels of all its off-path vertices. Definition 4 The set of off-path vertices for a vertex u in a tree is the set of all the siblings of each of the vertices on the path from u to the root of the tree that u is in (the path is inclusive of u). 281 Figure 2: Off-path vertices for u are highlighted in bold. The path from u to the root of its tree is shaded grey. Figure 2 shows a pictorial depiction of the off-path vertices for a vertex u in a tree. For a more concrete example, the set of off-path commitment tree vertices for G 0 in Figure 1 is {F 0 , E 0 , C 0 , B 1 , A 0 , D 0 , H 0 , I 0 }. To allow sensor node G to verify its contribution to the aggregate, the sensor network delivers labels of each off-path vertex to G 0 . Sensor node G then recomputes the sequence of computations and hashes and verifies that they lead to the correct root commitment value. Consider the congestion on the naive scheme. Let h be the height of the aggregation tree and be the maximum degree of any node inside the tree. Each leaf vertex has O (h) off-path vertices, and it needs to receive all their labels to verify its contribution to the aggregate , thus leading to O (h) congestion at the leaves of the commitment tree. For an aggregation tree constructed with TaG, the height h of the aggregation tree depends on the diameter (in number of hops) of the network, which in turn depends on the node density and total number of nodes n in the network. In a 2-dimensional deployment area with a constant node density, the best bound on the diameter of the network is O (n) if the network is regularly shaped. In irregular topologies the diameter of the network may be (n). 4.2.2 Aggregation-Commit: Improved Approach We present an optimization to improve the congestion cost. The main observation is that, since the aggregation trees are a sub-graph of the network topology, they may be arbitrarily unbalanced. Hence, if we decouple the structure of the commitment tree from the structure of the aggregation tree, then the commitment tree could be perfectly balanced. In the naive commitment tree, each sensor node always computes the aggregate sum of all its inputs. This can be considered a strategy of greedy aggregation. Consider instead the benefit of delayed aggregation at node C 1 in Figure 1(b). Suppose that C, instead of greedily computing the aggregate sum over its own reading (C 0 ) and both its child nodes E 0 and F 1 , instead computes the sum only over C 0 and E 0 , and passes F 1 directly to A along with C 1 = C 0 + E 0 . In such a commitment tree, F 1 becomes a child of A 1 (instead of C 1 ), thus reducing the depth of the commitment tree by 1. Delayed aggregation thus trades off increased communication during the aggregation phase in return for a more balanced commitment tree, which results in lower verification overhead in the result-checking phase. Greenwald and Khanna [6] used a form of delayed aggregation in their quantile summary algorithm. Our strategy for delayed aggregation is as follows: we perform an aggregation operation (along with the associated commit operation ) if and only if it results in a complete, binary commitment tree. We now describe our delayed aggregation algorithm for producing balanced commitment trees. In the naive commitment tree, each sensor node passes to its parent a single message containing the label of the root vertex of its commitment subtree T s . In the delayed aggregation algorithm, each sensor node now passes on the labels of the root vertices of a set of commitment subtrees F = {T 1 ,...,T q }. We call this set a commitment forest, and we enforce the condition that the trees in the forest must be complete binary trees, and no two trees have the same height. These constraints are enforced by continually combining equal-height trees into complete binary trees of greater height. Definition 5 A commitment forest is a set of complete binary commitment trees such that there is at most one commitment tree of any given height. A commitment forest has at most n leaf vertices (one for each sensor node included in the forest, up to a maximum of n). Since all the trees are complete binary trees, the tallest tree in any commitment forest has height at most log n. Since there are no two trees of the same height, any commitment forest has at most log n trees. In the following discussion, we will for brevity make reference to "communicating a vertex" to another sensor node, or "communicating a commitment forest" to another sensor node. The actual data communicated is the label of the vertex and the labels of the roots of the trees in the commitment forest, respectively. The commitment forest is built as follows. Leaf sensor nodes in the aggregation tree originate a single-vertex commitment forest, which they then communicate to their parent sensor nodes. Each internal sensor node s originates a similar single-vertex commitment forest. In addition, s also receives commitment forests from each of its children. Sensor node s keeps track of which root vertices were received from which of its children. It then combines all the forests to form a new forest as follows. Suppose s wishes to combine q commitment forests F 1 ,...,F q . Note that since all commitment trees are complete binary trees, tree heights can be determined by inspecting the count field of the root vertex. We let the intermediate result be F = F 1 F q , and repeat the following until no two trees are the same height in F: Let h be the smallest height such that more than one tree in F has height h. Find two commitment trees T 1 and T 2 of height h in F, and merge them into a tree of height h +1 by creating a new vertex that is the parent of both the roots of T 1 and T 2 according to the inductive rule in Definition 3. Figure 3 shows an example of the process for node A based on the topology in Figure 1. The algorithm terminates in O (qlogn) steps since each step reduces the number of trees in the forest by one, and there are at most q log n + 1 trees in the forest. Hence, each sensor node creates at most q log n + 1 = O(logn) vertices in the commitment forest. When F is a valid commitment forest, s sends the root vertices of each tree in F to its parent sensor node in the aggregation tree. The sensor node s also keeps track of every vertex that it created, as well as all the inputs that it received (i.e., the labels of the root vertices of the commitment forests that were sent to s by its children). This takes O (d logn) memory per sensor node. Consider the communication costs of the entire process of creating the final commitment forest. Since there are at most log n commitment trees in each of the forests presented by any sensor node to its parent, the per-node communication cost for constructing the final forest is O (logn). This is greater than the O(1) congestion cost of constructing the naive commitment tree. However, no path in the forest is longer than log n hops. This will eventually enable us to prove a bound of O (log 2 n ) edge congestion for the result-checking phase in Section 5.2. Once the querier has received the final commitment forest from the base station, it checks that none of the SUM or COMPLEMENT aggregates of the roots of the trees in the forest are negative. If 282 A 0 = 1,a A ,r - a A ,A D 0 = 1,a D ,r - a D ,D K 0 = 1,a K ,r - a K ,K C 2 = 4,v C 2 ,v C 2 ,H[N||4||v C 2 ||v C 2 ||F 1 ||C 1 ] B 1 = 2,v B 1 ,v B 1 ,H[N||2||v B 1 ||v B 1 ||B 0 ||J 0 ] (a) Inputs: A generates A 0 , and receives D 0 from D, C 2 from C, and (B 1 ,K 0 ) from B. Each dashed-line box shows the commitment forest received from a given sensor node. The solid-line box shows the vertex labels, each solid-line box below shows the labels of the new vertices. v A 1 = a A + a D v A 1 = r - a A + r - a D A 1 = 2,v A 1 ,v A 1 ,H[N||2||v A 1 ||v A 1 ||A 0 ||D 0 ] (b) First merge: Vertex A 1 created v A 2 = v A 1 + v B 1 v A 2 = v A 1 + v B 1 A 2 = 4,v A 2 ,v A 2 ,H[N||4||v A 2 ||v A 2 ||A 1 ||B 1 ] (c) Second merge: Vertex A 2 created v A 3 = v A 2 + v C 2 v A 3 = v A 2 + v C 2 A 3 = 8,v A 3 ,v A 3 ,H[N||8||v A 3 ||v A 3 ||A 2 ||C 2 ] (d) Final merge: Vertex A 3 created. A 3 and K 0 are sent to the parent of A in the aggregation tree. Figure 3: Process of node A (from Figure 1) deriving its commitment forest from the commitment forests received from its children. any aggregates are negative, the querier rejects the result and raises an alarm: a negative aggregate is a sure sign of tampering since all the data values (and their complements) are non-negative. Otherwise , the querier then computes the final pair of aggregates SUM and COMPLEMENT . The querier verifies that SUM + COMPLEMENT = nr where r is the upper bound on the range of allowable data values on each node. If this verifies correctly, the querier then initiates the result-checking phase. 4.3 Result-checking phase The purpose of the result-checking phase is to enable each sensor node s to independently verify that its data value a s was added into the SUM aggregate, and the complement (r - a s ) of its data value was added into the COMPLEMENT aggregate. The verification is performed by inspecting the inputs and aggregation operations in the commitment forest on the path from the leaf vertex of s to the root of its tree; if all the operations are consistent, then the root aggregate value must have increased by a s due to the incorporation of the data value. If each legitimate node performs this verification, then it ensures that the SUM aggregate is at least the sum of all the data values of the legitimate nodes. Similarly, the COMPLEMENT aggregate is at least the sum of all the complements of the data values of the legitimate nodes. Since the querier enforces SUM + COMPLEMENT = nr, these two inequalities form lower and upper bounds on an adversary's ability to manipulate the final result. In Section 5 we shall show that they are in fact the tightest bounds possible. A high level overview of the process is as follows. First, the aggregation results from the aggregation-commit phase are sent using authenticated broadcast to every sensor node in the network. Each sensor node then individually verifies that its contributions to the respective SUM and COMPLEMENT aggregates were indeed counted. If so, it sends an authentication code to the base station. The authentication code is also aggregated for communication effi-283 Figure 4: Dissemination of off-path values: t sends the label of u 1 to u 2 and vice-versa; each node then forwards it to all the vertices in their subtrees. ciency. When the querier has received all the authentication codes, it is then able to verify that all sensor nodes have checked that their contribution to the aggregate has been correctly counted. For simplicity, we describe each step of the process with reference to the commitment tree visualised as an overlay network over the actual aggregation tree. Hence, we will refer to vertices in the commitment tree sending information to each other; in the physical world, it is the sensor node that created the vertex is the physical entity that is responsible for performing communications and computations on behalf of the vertex. Each edge in the commitment tree may involve multiple hops in the aggregation tree; the routing on the aggregation tree is straightforward. Dissemination of final commitment values. After the querier has received the labels of the roots of the final commitment forest, the querier sends each of these labels to the entire sensor network using authenticated broadcast. Dissemination of off-path values. To enable verification, each leaf vertex must receive all its off-path values. Each internal vertex t in the commitment forest has two children u 1 and u 2 . To disseminate off-path values, t sends the label of u 1 to u 2 , and vice-versa (t also attaches relevant information tagging u 1 as the right child and u 2 as the left child). Vertex t also sends any labels (and left/right tags) received from its parent to both its children. See Figure 4 for an illustration of the process. The correctness of this algorithm in delivering all the necessary off-path vertex labels to each vertex is proven in Theorem 14 in Section 5.2. Once a vertex has received all the labels of its off-path vertices, it can proceed to the verification step. Verification of inclusion. When the leaf vertex u s of a sensor node s has received all the labels of its off-path vertices, it may then verify that no aggregation result-tampering has occurred on the path between u s and the root of its commitment tree. For each vertex t on the path from u s to the root of its commitment tree, u s derives the label of t (via the computations in Definition 3). It is able to do so since the off-path labels provide all the necessary data to perform the label computation. During the computation, u s inspects the off-path labels: for each node t on the path from u s to the root, u s checks that the input values fed into the aggregation operation at t are never negative. Negative values should never occur since the data and complement values are non-negative; hence if a negative input is encountered, the verification fails. Once u s has derived the label of the root of its commitment tree, it compares the derived label against the label with the same count that was disseminated by the querier. If the labels are identical, then u s proceeds to the next step. Otherwise, the verification fails and u s may either immediately raise an alarm (for example, using broadcast), or it may simply do nothing and allow the aggregate algorithm to fail due to the absence of its confirmation message in the subsequent steps. Collection of confirmations. After each sensor node s has suc-cessfully performed the verification step for its leaf vertex u s , it sends an authentication code to the querier. The authentication code for sensor node s is MAC K s (N||OK) where OK is a unique message identifier and K s is the key that s shares with the querier. The collation of the authentication codes proceeds as follows (note that we are referring to the aggregation tree at this point, not the commitment tree). Leaf sensor nodes in the aggregation tree first send their authentication codes to their parents in the aggregation tree. Once an internal sensor node has received authentication codes from all its children, it computes the XOR of its own authentication code with all the received codes, and forwards it to its parent. At the end of the process, the querier will receive a single authentication code from the base station that consists of the XOR of all the authentication codes received in the network. Verification of confirmations. Since the querier knows the key K s for each sensor node s, it verifies that every sensor node has released its authentication code by computing the XOR of the authentication codes for all the sensor nodes in the network, i.e., MAC K 1 (N||OK) MAC K n (N||OK). The querier then compares the computed code with the received code. If the two codes match, then the querier accepts the aggregation result. Otherwise, the querier rejects the result. A rejection may indicate the presence of the adversary in some unknown nodes in the network, or it may be due to natural factors such as node death or message loss. The querier may either retry the query or attempt to determine the cause of the rejection. For example, it could directly request the leaf values of every sensor node: if rejections due to natural causes are sufficiently rare, the high cost of this direct query is incurred infre-quently and can be amortised over the other successful queries. ANALYSIS OF SUM In this section we prove the properties of the SUM algorithm. In Section 5.1 we prove the security properties of the algorithm, and in Section 5.2 we prove bounds on the congestion of the algorithm. 5.1 Security Properties We assume that the adversary is able to freely choose any arbitrary topology and set of labels for the final commitment forest. We then show that any such forest which passes all the verification tests must report an aggregate result that is (optimally) close to the actual result. First, we define the notion of an inconsistency, or evidence of tampering, at a given node in the commitment forest. Definition 6 Let t = c t ,v t ,v t ,H t be an internal vertex in a commitment forest. Let its two children be u 1 = c 1 ,v 1 ,v 1 ,H 1 and u 2 = c 2 ,v 2 ,v 2 ,H 2 . There is an inconsistency at vertex t in a commitment tree if either (1) v t = v 1 + v 2 or v t = v 1 + v 2 or (2) any of {v 1 ,v 2 ,v 1 ,v 2 } is negative. Informally, an inconsistency occurs at t if the sums don't add up at t, or if any of the inputs to t are negative. Intuitively, if there are no inconsistencies on a path from a vertex to the root of the commitment tree, then the aggregate value along the path should be non-decreasing towards the root. Definition 7 Call a leaf-vertex u accounted-for if there is no inconsistency at any vertex on the path from the leaf-vertex u to the root of its commitment tree, including at the root vertex. Lemma 8 Suppose there is a set of accounted-for leaf-vertices with distinct labels u 1 ,...,u m and committed data values v 1 ,...,v m in 284 the commitment forest. Then the total of the aggregation values at the roots of the commitment trees in the forest is at least m i =1 v i . Lemma 8 can be rigorously proven using induction on the height of the subtrees in the forest (see Appendix A). Here we present a more intuitive argument. P ROOF . (Sketch) We show the result for m = 2; a similar reasoning applies for arbitrary m. Case 1: Suppose u 1 and u 2 are in different trees. Then, since there is no inconsistency on any vertex on the path from u 1 to the root of its tree, the root of the tree containing u 1 must have an aggregation value of at least v 1 . By a similar reasoning, the root of the tree containing u 2 must have an aggregation value of at least v 2 . Hence the total aggregation value of the two trees containing u 1 and u 2 is at least v 1 + v 2 . Case 2: Now suppose u 1 and u 2 are in the same tree. Since they have distinct labels, they must be distinct vertices, and they must have a lowest common ancestor t in the commitment tree. The vertices between u 1 and t (including u 1 ) must have aggregation value at least v 1 since there are no inconsistencies on the path from u 1 to t, so the aggregation value could not have decreased. Similarly, the vertices between u 2 and t (including u 2 ) must have aggregation value at least v 2 . Hence, one of the children of t has aggregation value at least v 1 and the other has aggregation value at least v 2 . Since there was no inconsistency at t, vertex t must have aggregation value at least v 1 +v 2 . Since there are no inconsistencies on the path from t to the root of the commitment tree, the root also must have aggregation value at least v 1 + v 2 . Negative root aggregate values are detected by the querier at the end of the aggregate-commit phase, so the total sum of the aggregate values of the roots of all the trees is thus at least v 1 + v 2 . The following is a restatement of Lemma 8 for the COMPLE MENTARY SUM aggregate; its proof follows an identical structure and is thus omitted. Lemma 9 Suppose there is a set of accounted-for leaf vertices with distinct labels u 1 ,...,u m with committed complement values v 1 ,...,v m in the commitment forest. Then the total COMPLEMENT aggregation value of the roots of the commitment trees in the forest is at least m i =1 v i . Lemma 10 A legitimate sensor node will only release its confirmation MAC if it is accounted-for. P ROOF . By construction, each sensor node s only releases its confirmation MAC if (1) s receives an authenticated message from the querier containing the query nonce N and the root labels of all the trees in the final commitment forest and (2) s receives all labels of its off-path vertices (the sibling vertices to the vertices on the path from the leaf vertex corresponding to s to the root of the commitment tree containing the leaf vertex in the commitment forest), and (3) s is able to recompute the root commitment value that it received from the base station and correctly authenticated, and (4) s verified that all the computations on the path from its leaf vertex u s to the root of its commitment tree are correct, i.e., there are no inconsistencies on the path from u s to the root of the commitment tree containing u s . Since the hash function is collision-resistant, it is computationally infeasible for an adversary to provide s with false labels that also happen to compute to the correct root commitment value. Hence, it must be that s was accounted-for in the commitment forest. Lemma 11 The querier can only receive the correct final XOR check value if all the legitimate sensor nodes replied with their confirmation MACs. P ROOF . To compute the correct final XOR check value, the adversary needs to know the XOR of all the legitimate sensor nodes that did not release their MAC. Since we assume that each of the distinct MACs are unforgeable (and not correlated with each other), the adversary has no information about this XOR value. Hence, the only way to produce the correct XOR check value is for all the legitimate sensor nodes to have released their relevant MACs. Theorem 12 Let the final SUM aggregate received by the querier be S. If the querier accepts S, then S L S (S L + r) where S L is the sum of the data values of all the legitimate nodes, is the total number of malicious nodes, and r is the upper bound on the range of allowable values on each node. P ROOF . Suppose the querier accepts the SUM result S. Let the COMPLEMENT SUM received by the querier be S. The querier accepts S if and only if it receives the correct final XOR check value in the result-checking phase, and S + S = nr. Since the querier received the correct XOR check value, we know that each legitimate sensor node must have released its confirmation MAC (Lemma 11), and so the leaf vertices of each legitimate sensor node must be accounted-for (Lemma 10). The set of labels of the leaf vertices of the legitimate nodes is distinct since the labels contain the (unique) node ID of each legitimate node. Since all the leaf vertices of the legitimate sensor nodes are distinct and accounted-for, by Theorem 8, S S L where S L is the sum of the data values of all the legitimate nodes. Furthermore, by Theorem 9, S S L , where S L is the sum of the complements of the data values of all the legitimate nodes. Let L be the set of legitimate sensor nodes, with |L| = l. Observe that S L = i L r - a i = lr - S L = (n - )r - S L = nr - (S L + r). We have that S + S = nr and S nr - (S L + r). Substituting, S = nr - S S L + r. Hence, S L S (S L + r). Note that nowhere was it assumed that the malicious nodes were constrained to reporting data values between [0,r]: in fact it is possible to have malicious nodes with data values above r or below 0 without risking detection if S L S (S L + r). Theorem 13 The SUM algorithm is optimally secure. P ROOF . Let the sum of the data values of all the legitimate nodes be S L . Consider an adversary with malicious nodes which only performs direct data injection attacks. Recall that in a direct data injection attack, an adversary only causes the nodes under its control to each report a data value within the legal range [0,r]. The lowest result the adversary can induce is by setting all its malicious nodes to have data value 0; in this case the computed aggregate is S L . The highest result the adversary can induce is by setting all nodes under its control to yield the highest value r. In this case the computed aggregate is S L + r. Clearly any aggregation value between these two extremes is also achievable by direct data injection. The bound proven in Theorem 12 falls exactly on the range of possible results achievable by direct data injection, hence the algorithm is optimal by Definition 2. The optimal security property holds regardless of the number or fraction of malicious nodes; this is significant since the security property holds in general, and not just for a subclass of attacker multiplicities. For example, we do not assume that the attacker is limited to some fraction of the nodes in the network. 5.2 Congestion Complexity We now consider the congestion induced by the secure SUM algorithm . Recall that node congestion is defined as the communication load on the most heavily loaded sensor node in the network, 285 and edge congestion is the heaviest communication load on a given link in the network. We only need to consider the case where the adversary is not performing an attack. If the adversary attempts to send more messages than the proven congestion bound, legitimate nodes can easily detect this locally and either raise an alarm or refuse to respond with their confirmation values, thus exposing the presence of the adversary. Recall that when we refer to a vertex sending and receiving information, we are referring to the commitment tree overlay network that lies over the actual physical aggregation tree. Theorem 14 Each vertex u receives the labels of its off-path vertices and no others. P ROOF . Since, when the vertices are disseminating their labels in the result-checking phase, every vertex always forwards any labels received from its parents to both its children, it is clear that when a label is forwarded to a vertex u , it is eventually forwarded to the entire subtree rooted at u . By definition, every off-path vertex u 1 of u has a parent p which is a node on the path between u and the root of its commitment tree. By construction, p sends the label of u 1 to its sibling u 2 which is on the path to u (i.e., either u 2 is an ancestor of u, or u 2 = u). Hence, the label u 1 is eventually forwarded to u. Every vertex u 1 that is not an off-path vertex has a sibling u 2 which is not on the path between u and the root of its commitment tree. Hence, u is not in the subtree rooted at u 2 . Since the label of u 1 is only forwarded to the subtree rooted at its sibling and nowhere else, the label of u 1 never reaches u. Theorem 15 The SUM algorithm induces O (log 2 n ) edge congestion (and hence O (log 2 n ) node congestion) in the aggregation tree. P ROOF . Every step in the algorithm except the label dissemination step involves either broadcast or convergecast of messages that are at most O (logn) size. The label-dissemination step is the dominating factor. Consider an arbitrary edge in the commitment-tree between parent vertex x and child vertex y. In the label dissemination step, messages are only sent from parent to child in the commitment tree. Hence the edge xy carries exactly the labels that y receives. From Theorem 14, y receives O (logn) labels, hence the total number of labels passing through xy is O (logn). Hence, the edge congestion in the commitment tree is O (logn). Now consider an arbitrary aggregation tree edge with parent node u and child node v. The child node v presents (i.e., sends) at most log n commitment-tree vertices to its parent u, and hence the edge uv is responsible for carrying traffic on behalf of at most log n commitment-tree edges -- these are the edges incident on the commitment tree vertices that v presented to u. Note that v may not be responsible for creating all the vertices that it presents to u, but v is nonetheless responsible for forwarding the messages down to the sensor nodes which created those vertices. Since each edge in the commitment tree has O (logn) congestion, and each edge in the aggregation tree carries traffic for at most log n commitment-tree edges, the edge congestion in the aggregation tree is O (log 2 n ). The node-congestion bound of O (log 2 n ) follows from the O(log 2 n ) edge congestion and the definition of as the greatest degree in the aggregation tree. OTHER AGGREGATION FUNCTIONS In this section we briefly discuss how to use the SUM algorithm as a primitive for the COUNT , AVERAGE and QUANTILE aggregates . The C OUNT Aggregate. The query COUNT is generally used to determine the total number of nodes in the network with some property; without loss of generality it can be considered a SUM aggregation where all the nodes have value either 1 (the node has the property) or 0 (otherwise). More formally, each sensor node s has a data value a s {0,1}, and we wish to compute f (a 1 ,...,a n ) = a 1 +a 2 ++a n . Since count is a special case of SUM , we can use the basic algorithm for SUM without modification. The A VERAGE Aggregate. The AVERAGE aggregate can be computed by first computing the SUM of data values over the nodes of interest, and then the COUNT of the number of nodes of interest, and then dividing the SUM by the COUNT . The -Q UANTILE Aggregate. In the QUANTILE aggregate, we wish to find the value that is in the n-th position in the sorted list of data values. For example, the median is a special case where = 0.5. Without loss of generality we can assume that all the data values are distinct; ties can be broken using unique node IDs. If we wished to verify the correctness of a proposed -quantile q, we can perform a COUNT computation where each node s presents a value a s = 1 if its data value a s q and presents a s = 0 otherwise. If q is the -quantile, then the computed sum should be equal to n. Hence, we can use any insecure approximate -quantile aggregation scheme to compute a proposed -quantile, and then securely test to see if the result truly is within the approximation bounds of the -quantile algorithm. CONCLUSION In-network data aggregation is an important primitive for sensor network operation. The strong standard threat model of multiple Byzantine nodes in sensor networks requires the use of aggregation techniques that are robust against malicious result-tampering by covert adversaries. We present the first optimally secure aggregation scheme for arbitrary aggregator topologies and multiple malicious nodes. This contribution significantly improves on prior work which requires strict limitations on aggregator topology or malicious node multiplicity , or which only yields a probabilistic security bound. Our algorithm is based on a novel method of distributing the verification of aggregation results onto the sensor nodes, and combining this with a unique technique for balancing commitment trees to achieve sublinear congestion bounds. The algorithm induces O (log 2 n ) node congestion (where is the maximum degree in the aggregation tree) and provides the strongest security bound that can be proven for any secure aggregation scheme without making assumptions about the distribution of data values. REFERENCES [1] H. Cam, S. Ozdemir, P. Nair, D. Muthuavinashiappan, and H. O. Sanli. Energy-efficient secure pattern based data aggregation for wireless sensor networks. Computer Communications, 29:446455, 2006. [2] C. Castelluccia, E. Mykletun, and G. Tsudik. Efficient aggregation of encrypted data in wireless sensor networks. In Proceedings of The Second Annual International Conference on Mobile and Ubiquitous Systems, 2005. [3] J.-Y. Chen, G. Pandurangan, and D. Xu. Robust computation of aggregates in wireless sensor networks: distributed randomized algorithms and analysis. In Proceedings of the Fourth International Symposium on Information Processing in Sensor Networks, 2005. [4] W. Du, J. Deng, Y. Han, and P. K. Varshney. A witness-based approach for data fusion assurance in wireless sensor 286 networks. In Proceedings of the IEEE Global Telecommunications Conference, 2003. [5] J. Girao, M. Schneider, and D. Westhoff. CDA: Concealed data aggregation in wireless sensor networks. In Proceedings of the ACM Workshop on Wireless Security, 2004. [6] M. B. Greenwald and S. Khanna. Power-conserving computation of order-statistics over sensor networks. In Proceedings of the twenty-third ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems, 2004. [7] I. Gupta, R. van Renesse, and K. P. Birman. Scalable fault-tolerant aggregation in large process groups. In Proceedings of the International Conference on Dependable Systems and Networks, 2001. [8] L. Hu and D. Evans. Secure aggregation for wireless networks. In Workshop on Security and Assurance in Ad hoc Networks, 2003. [9] C. Intanagonwiwat, D. Estrin, R. Govindan, and J. Heidemann. Impact of network density on data aggregation in wireless sensor networks. In Proceedings of the 22nd International Conference on Distributed Computing Systems, 2002. [10] P. Jadia and A. Mathuria. Efficient secure aggregation in sensor networks. In Proceedings of the 11th International Conference on High Performance Computing, 2004. [11] S. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong. TAG: a tiny aggregation service for ad-hoc sensor networks. SIGOPS Oper. Syst. Rev., 36(SI):131146, 2002. [12] S. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong. The design of an acquisitional query processor for sensor networks. In Proceedings of the 2003 ACM International Conference on Management of Data, 2003. [13] A. Mahimkar and T. Rappaport. SecureDAV: A secure data aggregation and verification protocol for sensor networks. In Proceedings of the IEEE Global Telecommunications Conference, 2004. [14] A. Manjhi, S. Nath, and P. B. Gibbons. Tributaries and deltas: efficient and robust aggregation in sensor network streams. In Proceedings of the ACM International Conference on Management of Data, 2005. [15] S. Nath, P. B. Gibbons, S. Seshan, and Z. R. Anderson. Synopsis diffusion for robust aggregation in sensor networks. In Proceedings of the 2nd International Conference on Embedded Networked Sensor Systems, 2004. [16] A. Perrig, R. Szewczyk, J. D. Tygar, V. Wen, and D. E. Culler. SPINS: Security protocols for sensor networks. Wirel. Netw., 8(5):521534, 2002. [17] B. Przydatek, D. Song, and A. Perrig. SIA: Secure information aggregation in sensor networks. In Proceedings of the 1st International Conference on Embedded Networked Sensor Systems, 2003. [18] D. Wagner. Resilient aggregation in sensor networks. In Proceedings of the 2nd ACM Workshop on Security of Ad-hoc and Sensor Networks, 2004. [19] Y. Yang, X. Wang, S. Zhu, and G. Cao. SDAP: A secure hop-by-hop data aggregation protocol for sensor networks. In Proceedings of the ACM International Symposium on Mobile Ad Hoc Networking and Computing, 2006. [20] Y. Yao and J. Gehrke. The COUGAR approach to in-network query processing in sensor networks. SIGMOD Rec., 31(3):918, 2002. APPENDIX A. PROOF OF LEMMA 8 We first prove the following: Lemma 16 Let F be a collection of commitment trees of height at most h. Suppose there is a set U of accounted-for leaf-vertices with distinct labels u 1 ,...,u m and committed values v 1 ,...,v m in F. Let the set of trees that contain at least one member of U be T F . Define val (X) for any forest X to be the total of the aggregation values at the roots of the trees in X . Then val (T F ) m i =1 v i . P ROOF . Proof: By induction on h. Base case: h = 0. Then all the trees are singleton-trees. The total aggregation value of all the singleton-trees that contain at least one member of U is exactly m i =1 v i . Induction step: Assume the theorem holds for h, and consider an arbitrary collection F of commitment trees with at most height h + 1 where the premise holds. If there are no trees of height h + 1 then we are done. Otherwise, let the set R be all the root vertices of the trees of height h + 1. Consider F = F\R, i.e., remove all the vertices in R from F. The result is a collection of trees with height at most h. Let T F be the set of trees in F containing at least one member of U . The induction hypothesis holds for F , so val (T F ) m i =1 v i . We now show that replacing the vertices from R cannot produce an T F such that val (T F ) &lt; val(T F ). Each vertex r from R is the root of two subtrees of height h in F. We have three cases: Case 1: Neither subtree contains any members of U . Then the new tree contains no members of U , and so is not a member of T F . Case 2: One subtree t 1 contains members of U . Since all the members of U are accounted-for, this implies that there is no inconsistency at r. Hence, the subtree without a member of U must have a non-negative aggregate value. We know that r performs the aggregate sum correctly over its inputs, so it must have aggregate value at least equal to the aggregate value of t 1 . Case 3: Both subtrees contain members of U . Since all the members of U are accounted-for, this implies that there is no inconsistency at r. The aggregate result of r is exactly the sum of the aggregate values of the two subtrees. In case 2 and 3, the aggregate values of the roots of the trees of height h +1 that were in T F , was no less than the sum of the aggregate values of their constituent subtrees in T F . Hence, val (T F ) val (T F ) m i =1 v i . Let the commitment forest in Lemma 8 be F. Let the set of trees in F that contain at least one of the accounted-for leaf-vertices be T . By the above lemma, val (T) m i =1 v i . We know that there are no root labels with negative aggregation values in the commitment forest, otherwise the querier would have rejected the result. Hence, val (F) val(T) m i =1 v i . 287

algorithm;Secure aggregation;commitment forest;in-network data aggregation;commitment tree;Sensor Networks;secure hierarchical data aggregation protocol;sensor network;aggregation commit;result checking;query dissemination;congestion complexity;Data aggregation

176

SensorBus: A Middleware Model for Wireless Sensor Networks

The use of middleware eases the development of distributed applications by abstracting the intricacies (communication and coordination among software components) of the distributed network environment. In wireless sensor networks, this is even trickier because of their specific issues such as addressing, mobility, number of sensors and energy-limited nodes. This paper describes SensorBus, a message-oriented middleware (MOM) model for wireless sensor networks based on the publish-subscribe paradigm and that allows the free exchange of the communication mechanism among sensor nodes allowing as result the capability of using more than one communication mechanism to address the requirements of larger number of applications. We intend to provide a platform which addresses the main characteristics of wireless sensor networks and also allows the development of energy-efficient applications. SensorBus incorporates constraint and query languages which will aid the development of interactive applications. It intends with the utilization of filters reduces data movement minimizing the energy consumption of nodes.

INTRODUCTION Recent advances in wireless networking technology, low-power digital circuits, sensing materials and Micro Electro-Mechanical Systems (MEMS) opened up the possibility of building small sensor devices capable of data processing, remote sensing and wireless communication. When several small sensors are scattered and linked over an area we may call this arrangement a "Sensor Network". These networks can be used for collecting and analyzing data from the physical environment. More specifically, sensor networks are comprised of hundreds or even thousands of heterogeneous sensor nodes exchanging information to perform distributed sensing and collaborative data processing [1]. From a functional perspective sensor networks behave like distributed systems with many different types of sensor nodes. Given the diversity of node functionality and the size of these networks it is important for a user to be able to program and manage the distributed applications that perform the information gathering. A programmer may develop these applications using operating system primitives. This kind of procedure, however, brings another level of complexity to the programmer, in which he not only has to deal with low-level primitives but he will also have to treat issues concerning communication and coordination among software components distributed over the network. A much friendlier approach is the utilization of a middleware in order to provide higher-level primitives to hide issues concerning the distributed environment. Traditional middleware is not suited to this task because of the characteristics of wireless networks. For example, conventional middleware designed for wired networks raises exceptions when they do not find a specific component, but this situation is much more like the standard than the exception in wireless environments. The lower bandwidth available for wireless networks requires optimizing the transport of data and this is not considered in conventional middleware. The coordination primitives of these middleware products do not take into account the frequent disconnections that happen in wireless networks. Another problem is the size and computing requirements of these middleware products; they are often too large and too heavy to be running in a device with so few resources. Finally, the transparency level provided is not sufficient enough because the application running in such devices needs information about the execution context to better adapt itself. A series of new middleware environments were proposed to deal with the requirements imposed by the wireless environment [2]. Middleware products based on computing reflection are designed to be light (concerning the computing power required to run) and easily configurable. Middleware based on tuple space were proposed to address the problem of frequent disconnections, and present a more natural way to deal with asynchronous communication. Context-aware middleware includes the ability of an application to access its information context (context-awareness ). These proposals addressed adequately the issues brought by the mobile networks, but are not well suited to support the specific requirements of the target applications used or to be used in wireless sensor networks because they are designed to support traditional client-server applications used in regular (wired) environments. Wireless sensor networks are very similar to conventional wireless networks; including energy-limited nodes, low bandwidth and communication channels more prone to errors. However, communication in wireless sensor nets differs from the end-to-end connections often necessary in usual networks [1]. In other words, the function of the network is to report information considering the phenomenon to the observer who is not necessarily aware that the sensor infrastructure is being used as a means of communication. In addition, energy is much more limited in sensor networks than in other types of wireless nets due to the nature of the sensor devices and the difficulty of reloading batteries in hostile regions. Some works have shown that the execution of 3000 instructions costs the same amount of energy necessary to send 1-bit of data over 100 meters via radio [3]. Those studies indicate that we must prioritize computing over communications. The communication issues are addressed in the several routing protocols proposed for wireless sensor nets. The communication model allows other ways of addressing the sensor nodes besides single addressing. The sensor nodes can be addressed by their own attributes or by attributes extracted from the physical environment (attribute-based naming). The sharp limitation of energy demands that sensor nodes actively take part in the processing and dissemination of information in order to save as much energy as possible. Although the majority of the protocols reviewed are efficient in saving energy, they differ in addressing capabilities. Some of them utilize single addressing [4] while others utilize attribute-based naming [5]. Thus, each type of application requires an adaptation of the communication mechanism to address specific application issues. Trying to overcome these problems this paper proposes SensorBus, a message oriented middleware for sensor networks allowing the free exchanging of the communication mechanism among sensor nodes. We propose a platform that provides facilities for the development of energy-efficient applications and that also addresses the key characteristics of sensor networks. This type of middleware should be suited to perform environmental monitoring where single addressing is demanded (small areas) as well as where attribute-based naming is necessary (large areas). The remainder of this paper is organized as follows: Section 2 describes the type of sensor networks considered in this research; Section 3 presents the target application and explains its requirements; Section 4 broaches the abstractions and mechanisms needed to address the requirements listed on the previous section; Section 5 describes the components of the SensorBus architecture; Section 6 presents the communication architecture and explains the steps needed to develop an application using SensorBus; Section 7 broaches implementing and coding issue; Section 8 presents the related works, and finally Section 9 concludes the paper. ASSUMPTIONS Most of the algorithms catalogued in the sensor networks literature are hypothetical [1], i.e., they were proposed as an experiment and were not tested in real networks (although many of them were deployed in testbed environments). This research is no different. When we speak about wireless sensor networks, we are referring to the projected and experimental designs and deployments discussed in the literature and not to actual instances of wireless sensor networks deployed in the field. Differently from the real settings, the testbed environments are built and organized focusing on the network features one wants to observe and test. This organization involves three main components: infrastructure, network protocols and applications [6]. The infrastructure is formed of sensor nodes and their topology the way they were scattered over a determined region. The network protocol is responsible for the creation and maintenance of communication links between sensor nodes and the applications. The applications extract information about a determined phenomenon through the sensor nodes. The following topics introduce in more details the assumptions made considering those aspects. 2.1 Applications The way in which the applications gather data from the sensor nodes depends on the network design. In the literature, we found that there are four data transfer modes between sensor nodes and applications: continuous, event-oriented, query-oriented and hybrid [6]. In the continuous model, sensor nodes transfer their data continuously at a predefined rate. In the event-oriented model, nodes transfer data only when an event of interest occurs. In the query-oriented model the application is responsible for deciding which events are of interest, thus requesting data about a phenomenon. Lastly, in the hybrid model, the three approaches may coexist together. In this research, we adopt a hybrid approach in the way that it utilizes the query and the event-oriented model as will be shown in the target application presented in Section 3. 2.2 Network Protocol The performance of the network protocol is influenced by the communications model adopted, the packet data transfer mode and the network mobility. In order to evaluate how a network protocol behaves it is important to take into account these aspects. Communication in sensor networks is classified in two major categories [6]: application and infrastructure. Application communication consists of the transfer of data obtained by the sensor nodes to the observer. This kind of communication is of 3 two types: cooperative and non-cooperative. In cooperative mode sensor nodes exchange data among themselves before transmitting the data gathered. In non-cooperative mode, however, sensor nodes do not exchange any kind of information; each one is solely responsible for reporting its collected data. The infrastructure data refers to the information needed to set, maintain and optimize the running network. As the network protocol must support both categories, the SensorBus architecture will not address those issues. The packet data transfer is a routing issue concerning the network protocol. This routing is divided into three types: flooding, unicast and multicast [6]. In the flooding approach, the sensor node broadcasts its information to neighboring nodes that, in turn, broadcast this information to their neighboring nodes until the information reaches the destination node. Alternatively, the sensor node may transmit its data directly to the observer using unicast multi-hop routing and also might use a cluster-head through one-to -one unicast. Lastly, the multicast approach casts information to predefined groups of nodes. The routing protocol is responsible for treating packet data transfer relieving SensorBus of these issues. Regarding mobility, sensor networks are divided into static and dynamic [6]. In static nets there is no movement by sensor nodes, the observers or the phenomenon to be studied. Conversely, in dynamic networks the nodes, observers and the phenomenon might well change their locations. This kind of network is further classified by the mobility of its components in dynamic nets with mobile observer, dynamic nets with mobile sensors and dynamic nets with mobile phenomena respectively. In the first, the observer is mobile in relation to the sensors and phenomena; in the second; the sensors are moving with respect to each other and the observer; and in the later; the phenomenon itself is in motion. The routing protocol is also responsible for treating mobility issues, relieving SensorBus of these concerns. 2.3 Infrastructure As for the infrastructure, the issues to take into consideration are location, access point and sensor node's computing power. The nodes have well-know locations and are to be scattered over a well-defined area. We will assume that all information is transmitted and received by means of a unique access point called the sink node. Despite the fact that, for this model, we will consider all nodes as being the same, there is nothing to prevent one node from having more memory, more energy or more computing power available. ENVIRONMENTAL MONITORING APPLICATIONS Environmental Monitoring Applications are used to evaluate qualitatively and quantitatively the natural resources of a determined area. These applications collect data, analyze and follow continuously and systematically environmental variables in order to identify current patterns and predict future trends [7]. Environmental monitoring provides information about the factors influencing conservation, preservation, degradation and environmental recovery. One might consider it a tool of evaluation and control. Wireless sensor networks can be used for performing environmental monitoring in indoor locations such as a building or a house or outdoors locations such as forests, lakes, deserts, rivers, etc. Internal monitoring might be described as tracking the variables in an indoor location. For example, one might deploy an infrared camera to track motion in a room that is supposed to be secure; if motion is detected an internal device might trigger an alarm. Sometimes in order to detect and identify an event, information from more than one sensor might be required. These results are processed and compared with the signature of the event of interest. In outdoor monitoring, there may be thousands of sensors scattered over an area and when an event of interest occurs such as temperature change, moisture change or CO2 increase the sensor might trigger the management events module which in turn sends the observer a signal to notify him or her of the event. Wireless sensors might be useful in a way that can save money in deploying a sensor infrastructure such as described in [8] where the authors were able to decrease the number of sensors needed to monitor forest fires in comparison with a wired model. In summary, the value of a wireless sensor network relies in its ability to provide information over a large area in reply to the questions put to users. The query mode is the most common approach used. Another approach is the mode in which sensors may remain waiting for some event to happen. By observing these aspects we draw the first requirement (R1) of our middleware model: The system must be able to function in two modes: query-driven and event-oriented. Depending on the application would be more convenient to access a specific node or a specific property. For example, in internal environmental monitoring, if one wants to know the temperature of a determined room you will have to access the information collected by a specific sensor, thus requiring unique node addressing and identification. On the other hand, in external environmental monitoring, sensor nodes do not need to be uniquely identified, as in this kind of application the purpose is to collect the value of a certain variable in a given area. From that observation we extract the second requirement (R2): The system must be able to address uniquely the sensor nodes and also by attribute (property to be observed). In some applications the mobility of sensor nodes must be taken into account. For example, sensors scattered over a forest for collecting dampness and temperature data are to be static, i.e. they must not change its geographical location, while that placing sensors in a river's surface for collecting data about its contamination levels characterizes a mobile environment. Thus, the third requirement (R3) of our middleware model is taking into consideration mobility issues. The sensing coverage area of a given wireless node is smaller than its radio coverage. Besides, sensors operate in noisy environments. To achieve a trustworthy sensory resolution a high density of sensors is required. In some applications the size of the coverage area leads to a great number of sensor nodes. A simple application in the field of environmental monitoring such as surveillance of oceans and forests requires from hundreds to thousands of nodes. In other applications, like internal environmental monitoring, the amount of nodes is limited by the size of the area. Therefore, the fourth requirement (R4) is to take into account the size of the network. 4 In external environmental monitoring, the nodes are spread in a hostile region, where it is not possible to access them for maintenance. The lifetime of each sensor node depends exclusively on the little available energy for the node. To conserve energy, the speed of the CPU and the bandwidth of the RF channel (Radio Frequency) must be limited. This requirement adds some restrictions in CPU performance, memory size, RF bandwidth and in battery size. In applications where the sensors are not spread in a hostile region it is possible to access them for maintenance and the battery lifetime of each sensor does not become a critical aspect. Finally, the fifth requirement (R5) is to take into consideration the limited energy resources of each sensor node. MECHANISMS AND ABSTRACTIONS This middleware model is comprised of three mechanisms and one abstraction. The publish-subscribe paradigm is employed as well as constraints and query languages and application filters to meet R1, R2 and R5 requirements. The design patterns abstraction is used to meet R2, R3 and R4 requirements. 4.1 Publish-Subscribe Paradigm The SensorBus is a Message Oriented Middleware (MOM) that employs the publish-subscribe paradigm. In this approach, a component that generates events (producer) publishes the types of events that will be available to other components (consumers) [9]. The consumer interested in a determined event "subscribes" to this event, receiving from this moment on notifications about the event "subscribed" to. These notifications are sent asynchronously from producers to all interested consumers. The MOM performs the functions of collecting producer's messages, filtering and transforming such messages (when necessary) and routing them to the appropriate consumers. The publish-subscribe communication is anonymous, asynchronous and multicast. Data are sent and received by asynchronous broadcast messages, based in subject, independent from identity and location of producers and consumers. This kind of communication enlists desirable properties for sensor networks; for example, this model saves energy while a given node does not need to be waiting for a synchronous response to proceed as it is in networks that implements end-to-end connections, increasing the lifetime of the network. Furthermore, as it also implements multicast, a group of sensor might be formed regarding a specific application. As a consequence, the adoption of the publish-subscribe paradigm meets the R1 requirement, concerning the need for events and the R2 requirement pertaining to attribute addressing. In addition it also meets the R5 requirement related to energy saving. 4.2 Constraint and Query Languages Constraint and Query languages are used to filter collecting data by specifying restrictions in the values and preferences of the attributes. A statement in these languages is a string that represents an expression. The constraint language only includes constants (values) and operations over values. Values and operations with integer, float, boolean and strings are allowed. The language admits several types of expressions. The expressions can be comparative: == (equality), != (inequality), &gt;, &gt;=, &lt;, &lt;=. For instance, Temperature &lt; 36.6 means to consider data where the attribute Temperature is less than 36.6 degrees Celsius. The expressions can be boolean: AND, OR, NOT. For example, Temperature &gt;= 26.6 AND Temperature &lt; = 36.6 implies to consider data where the value of the attribute Temperature is between 26.6 and 36.6 degrees Celsius. The expressions can be numerical with the mathematical operators + (addition), - (subtraction), * (multiplication) and / (division). The query language has its syntax based on a subgroup of the conditional expression syntax SQL92 [10]. It is an extension of the constraints language with new functions. This new language embodies identifiers that can hold a constant value. A mapping between identifiers and values is required. In the evaluation of an expression, the occurrence of an identifier is replaced by its associated value. The addition of new operators (between, like, in, is, escape) allows submitting queries similar to those used in databases compliant with SQL92. For example, queries of the type -- Temperature between 26.6 and 36.6 -- are possible. The constraint and query languages are intended to ease the work programming of online applications. This type of application access the information sent in real-time by the sensor nodes. Thus, it completes the attendance of the R1 requirement on the way of operation for query. 4.3 Application Filters Filters are application specific software modules that deal with diffusion and data processing [11]. Filters are provided before deploying a sensor network. Each filter is specified using a list of attributes to make possible the matching with the incoming data. Filters are used to make internal aggregation of data, collaborative signals processing, caching and tasks that control the data flow in the sensor network [11]. In SensorBus, filters will be used to limit the data flow in the network. A filter can be designed to restrict the range of values of a determined attribute, for example the application requires that the attribute Temperature has values ranging between 20 and 30 degrees Celsius, the values outside this particular range are of no interest. The filtering process discards the unnecessary data reducing the flow between the nodes. This decrease reduces the consumption of energy in sensor nodes. Thus, it completes the attendance of the R5 requirement about the energy saving of the sensor nodes. 4.4 Design Patterns Design patterns are descriptions of objects and communicating classes that are customized to solve a general design problem within a particular context [12]. It describes commonly recurring design architectures extracted from the experience of one or more domain specialists. A design pattern names, abstracts, and identifies the key aspects of a common design structure that make it useful for creating a reusable object-oriented design [12]. We make use of design patterns in SensorBus project and the types we have utilized are as follows: 5 The Observer pattern: Defines a one-to-many dependency between objects so that when one object changes state, all of its dependents are notified and updated automatically. We utilize this pattern to implement the publish-subscribe mechanism. The Interpreter pattern: Defines a representation for its grammar along with an interpreter that uses the representation to interpret sentences in the language. We make use of this pattern to implement the constraint and query language. The Facade pattern: Defines a unified (higher-level) interface to a set of interfaces in a subsystem that makes the subsystem easier to use. We use this pattern to implement the middleware high-level primitives which will be available to developers. The Mediator pattern: Defines an object that encapsulates how a set of objects interact. Mediator promotes loose coupling by keeping objects from referring to each other explicitly, and it lets you vary their interaction independently. The Adapter pattern: Converts the interface of a class into another interface clients expect. Adapter lets classes work together that couldn't otherwise because of incompatible interfaces. The Router pattern: Decouples multiple sources of input from multiple sources of output to route data correctly without blocking. The design patterns Mediator, Adapter e Router are utilized to implement the middleware message bus. The exchangeable communication mechanism was written using these patterns. This mechanism allows the utilization of any routing protocol designed for sensor networks meeting as a result the requirements R2, R3 and R4. SENSORBUS ARCHITECTURE SensorBus is comprised of the following elements: an application service, a message service and a context service as shown in Figure 1. Figure 1. Middleware architecture. The following sections present each one of the services mentioned by the means of UML (Unified Modeling Language) component diagrams [13]. 5.1 Application Service The application service provides Application Programming Interface (API) which simplifies application development. This service is comprised of three components as shown in Figure 2: Figure 2. Application Service Architecture. DataBus: component providing a set of operations relating to bus communication for consumers and producers. These operations include: Announcement of data item (producer); to find a data item (consumer); Announcement of data change (consumer); Exclude data item (producer). Filter: component providing a set of operations relating to data filtering. Language: component that implements the commands and the constraint and query language interpreter. 5.2 Message Service Message service is responsible for providing communication and coordination for the distributed components, abstracting the developer from these issues. This service also comprises three components as is shown in figure 3: Figure 3. Message Service Architecture. Channel: Component designed to deal with the specific transport implementations. Each instance of Channel represents a simple system channel. The component Channel maintains the global state information about the availability of channels and is also responsible for exchanging channel's messages to the transport implementation and vice versa. Transport: The communication among the nodes is made through a specific transport implementation such as sockets. Each transport implementation communicates through a channel with a message exchange server called Sinker. All transport implementations have a common interface which is called ITransport. Sinker: Component responsible for routing messages among instances of transport implementation, each instance corresponding to an instance of Channel. Channel IChannel Sinker Transport ITransport Filter DataBus IDataBus Language ILanguage NOS - JVM/KVM Application/User Application Service Message Context Service 6 5.3 Context Service Inherently, an application running on a wireless sensor network needs to capture information from the execution context, for example, battery level, memory availability, bandwidth, location, application specific information such as temperature and pressure, etc. The middleware gets this information by interacting with several heterogeneous sensors; for example, the level of energy remaining on batteries can be obtained by executing an operating system primitive, location can be acquired from various communications technology such as GPS, infrared and RF. This work does not take into consideration how the context sensing is executed; it is assumed that each sensor provides an interface so the middleware can use it to get the value of the resource of interest. The context service manages the heterogeneous sensors that collect information from the environment. For each resource the middleware manages, there is an adapter that interacts with the physical sensor, processes its information thus obtaining the information demanded by the application. Only resource adapters that are necessary to the running application will be loaded to avoid unnecessary spending of the node's scarce computing power. Figure 4 shows an energy adapter interacting with an energy sensor (an operating system primitive, in this example). Figure 4. Context Service Architecture. MIDDLEWARE ARCHITECTURE EXAMPLE Figure 5 shows the sensor network communication architecture. Each node in the field has the ability to collect data and send it to the next sink node. The sink node can be a mobile node acting as a data source or a fixed host computer (a PC). The user node connects with the sink node through a conventional wireless LAN (e.g. IEEE 802.11x). Figure 5. Communication Architecture. The services and components of SensorBus are distributed in three distinct types of sensor nodes. The components DataBus, Language, Channel and Transport are in the user node. The Sinker component is in the sink node. The sensor nodes contains the Channel and Transport components while filter component and context service will only be loaded if the application requires energy management and other resources such as memory and bandwidth. The development of an application using SensorBus consists in coding the parts for the producer and consumer. The consumer code runs in the user machine while the producer code runs in sensor nodes. The minimum steps required for the use of SensorBus are as follows: 1. Create a new DataBus instance. A new transport implementation is created by identifying a specific Sinker; 2. Instantiate a producer or a consumer; 3. Instantiate a "Channel" entity; 4. Register the just created producer or consumer for the channel; and 5. The producer generates data items and places them into Channel while the consumer finds and "crunches" those data. SensorBus offers other functions that might be implemented, such as listing the available channels, adding new channels and stop receiving new channels. The producer sensor code has to be implemented before setup of the network. If it is not possible to retrieve the sensor for maintenance, the attributes of the data sent will always be the same. To overcome this obstacle the constraint and query languages are used to add new queries that had not been initially foreseen. These queries are sent by the interested consumer (client) in the form of messages. Figure 6. Middleware architecture example. WLAN Sink node User node Sensor field Sensor nodes EnergySensor EnergyAdapter EnergyEvent IEvent IAdapter Data &lt;&lt;user&gt;&gt; &lt;&lt;application&gt;&gt; Language Filter Channel Transport Sinker Battery Battery Temperature Sensor IAdapter Application Service Message Service Context Service Temperature Adapter 7 Filters are as well implemented in nodes. Soon after a producer is instantiated, a filter is also instantiated and registered for a new channel. Figure 6 shows the components that may be active in a given moment. Although most of the components are the same for a given application, different settings may occur on the context service. The figure shows only two adapters running at the same time and interacting with its associated sensors (temperature and battery). Distinct sensors can be used depending on the physical measurement to be taken and the type of computing resource to be managed. IMPLEMENTATION ISSUES The testbed setup for SensorBus evaluation consists of Intel-based equipment equipped with 802.11b cards. The sink node is a centrino-based Dell Lattitude notebook, the sensor nodes are deployed in handheld computers HP iPAQ running Linux operating system on Intel XScale processor. The sensor nodes are placed at various locations in the Electrical Engineering Department Building (about 40m60m) at Federal University of Par. Linksys Wireless LAN cards are used working in the DCF mode with a channel bandwidth of 11Mbps. In the building, there is interference from IEEE 802.11 access points (AP) and other electronic devices. 7.1 Working Prototype For our working prototype, we have chosen the Java platform as our implementation technology because of its broad installed base and to ensure compatibility with most hardware platforms. The KVM (Kilobyte Virtual Machine) [14] is being used due to its freely available source-code and its designed targeting towards small limited-resources devices similar to the sensor nodes of this work. The SensorBus API and its constraint and query language are being coded as Java classes. Due to issues regarding efficiency, the code that will run in sensor nodes is being implemented as native code. An object serialization mechanism was implemented because KVM does not support this facility. Serialization mechanism converts an object and its state into a byte stream allowing this object to be moved over a network or persisted in a local file system. Object recovery is performed through another mechanism called deserialization. Other Java technologies as J2SE Java 2 Standard Edition utilize this kind of facility to support the encoding of objects into a stream of bytes while protecting private and transient data. Serialization is used in distributed programming through sockets or Remote Method Invocation (RMI). We have coded a semiautomatic serialization in order to store the state of the objects. To achieve this, we had to define a series of new interfaces and classes. One of the most critical problems with serialization is security because when an object is converted in a byte stream any attacker equipped with properly sniffer software can intercept and access it; in this case even private attributes can be accessed without any special technique. To tackle this issue, secure protocols as HTTPS (Secure HyperText Transport Protocol) or serialization encryption can be used. 7.2 Simulation Having implemented this working prototype, this research now intends to do performance evaluation by using the very well known tool NS Network Simulator [15]. We plan to integrate the SensorBus middleware with NS in a way that the sensor nodes will plug into NS in order to provide real data for feeding the simulator model. To do so, an execution environment will be added to the simulator. This environment will run as a sole UNIX process and will be plugged to the NS protocol stack through a sensor agent. The sensor agent is actually a NS agent responsible for connecting an execution environment instance to the NS protocol stack. The communication takes place through a pair of UDP (User Datagram Protocol) sockets. Incoming packets are encapsulated in NS packets and transmitted through the simulated sensor network. Parameters that need to be known to the protocol stack are placed in the header of the NS packet while the rest of the information is added to the payload of the NS packet. Similarly, outgoing packets are retrieved from the NS packets and sent to the execution environment to be processed. It will be necessary to provide a mechanism to synchronize the execution and simulation environment since they run in distinct times. Simulations will be performed using a NS Directed Diffusion transport implementation [5] for wireless sensor networks. RELATED WORKS In [16] an overall description of the challenges involving middleware for wireless sensor networks is presented focusing on the restraint aspects of these systems. Cougar [17] and SINA [18], Sensor Information Networking Architecture, provide a distributed database interface for wireless sensor networks that use a query language to allow applications to run monitoring functions. Cougar manages the power by distributing the query among the sensor nodes to minimize energy required in data gathering. SINA adds low-level mechanisms to build hierarchical clustering of sensors aiming at efficient data aggregation and also provides protocols which limit the rebroadcast of similar information to neighbor's nodes. AutoSec [19], Automatic Service Composition, manages resources of the sensor networks by providing access control for applications to ensure quality of service. This approach is very similar to conventional middleware technology but the techniques to collect resource information are suitable for wireless sensor networks. DSWare [20] provides service abstraction similar to AutoSec, but instead of having a service provided by only one sensor node, the service is supplied by a group of neighbor's nodes. Smart Messages Project [21] proposes a distributed computing model based on migration of executing units. Smart messages are migratory units containing data and code. The goal of Smart Messages Project is to develop a computing model and systems architecture to Networks Embedded Systems (NES). EnviroTrack [22] is a middleware for object-based distributed systems that lifts the abstraction level for programming 8 environmental monitoring applications. It contains mechanisms that abstract groups of sensors into logical objects. Impala [23] exploits mobile code techniques to alter the middleware's functionality running on a sensor node. The key to energy-efficient management in Impala is that applications are as modular and concise as possible so little changes demands fewer energy resources. MiLAN [24] was developed to allow dynamic network setup to meet the applications performance requirements. The applications represent its requests by the means of specialized graphics which incorporates changes due to applications needs. In [25], an adaptative middleware is proposed to explore the commitment between resource spending and quality during information collecting. The main goal is to decrease the transmissions among sensor nodes without compromising the overall result. Every one of those middleware proposals is designed to make efficient use of wireless sensor networks; they do not support free exchange of the transport mechanism. More specifically, most of those approaches are not capable of altering the routing protocol to meet different application requirements. CONCLUDING REMARKS As was demonstrated, application development is closely related to wireless sensor network design. Each communication mechanism provided by a determined routing protocol is application specific, e.g. it is designed to meet some application specific requirement. We suggest that the utility of the middleware for wireless sensor networks is supported by decoupling the communication mechanism from the programming interfaces and also by capability of using more than one communication mechanism to address the requirements of larger number of applications. We have shown that SensorBus, a sensor network middleware that we are developing to meet these goals, can aid the development of different types of sensor network applications. REFERENCES [1] P. Rentala, R. Musunuri, S. Gandham and U. Saxena, Survey on Sensor Networks, Technical Report, University of Texas, Dept. of Computer Science, 2002. [2] G. -C. Roman, A. L. Murphy, and G. P. Picco, Software Engineering for Mobility: A Roadmap. In The Future of Software Engineering 22 nd Int. Conf. On Software Engineering (ICSE2000), pages 243-258. ACM Press, May 2000. [3] J. Pottie and W. J. Kaiser, Embedding the internet wireless integrated network sensors, Communications of the ACM, vol. 43, no. 5, pp. 51-58, May 2000. [4] W. Heinzelman, A. Chandrakasan and H. Balakrishnan, Energy-efficient communication protocol for wireless micro sensor networks. Proceedings of the 33 rd Annual Hawaii International Conference on System Sciences, Pages 3005-3014 , 2000. [5] C. Intanagonwiwat, R. Govindan, and D. Estrin, Directed diffusion: A scalable and robust communication paradigm for sensor networks. In Proceedings of the ACM/IEEE International Conference on Mobile Computing and Networking, pages 56-67, Boston, MA, USA, Aug. 2000. [6] T. Sameer, N. B. Abu-Ghazaleh and Heinzelman W. A Taxonomy of Wireless Micro-Sensor Network Models. Mobile Computing and Communications Review, Volume 1, Number 2, 2003. [7] Guia de Chefe Brazilian Institute of Environment (IBAMA) http://www2.ibama.gov.br/unidades/guiadechefe/guia/t-1corpo .htm. December, 2004. [8] B. C. Arrue, A. Ollero e J. R. M. de DIOS, An intelligent system for false alarm reduction in infrared forest-fire detection, IEEE Intelligent Systems, vol. 15, pp. 64-73, 2000. [9] G. Couloris, J. Dollimore, e T. Kindberg Distributed Systems: Concepts and Design. Third edition. Addison-Wesley , 2001. [10] SQL92 Database Language SQL July 30, 1992. http://www.cs.cmu.edu/afs/andrew.cmu.edu/usr/shadow/www /sql/sql1992.txt [11] J. Heidemann, F. Silva, C. Intanagonwiwat, R. Govindan, D. Estrin and D. Ganesan. Building efficient wireless sensor networks with low-level naming. In Proceedings of the Symposium on Operating Systems Principles, pages 146-159, Chateau Lake Louise, Banff, Alberta, Canada, October 2001. [12] E. Gamma, R. Helm, R. Johnson e J. Vlissides, Design Patterns. Addison-Wesley, 1995. [13] J. Rumbaugh, I. Jacobson and G. Booch. The Unified Modeling Language Reference Manual. Addison Wesley, 1998. [14] KVM The K Virtual Machine Specification. http://java.sun.com/products/kvm/ , August 2004. [15] UCB/LBNL/VINT Network Simulator NS (Version 2). http//www.isi.edu./nsnam/ns/, August 2004. [16] K. Rmer, O. Kasten and F. Mattern. Middleware Challenges for Wireless Sensor Networks. Mobile Computing and Communications Review, volume 6, number 2, 2002. [17] P. Bonnet, J. Gehrke and P. Seshadri. Querying the Physycal World. IEEE Personal Communication, 7:10-15, October 2000. [18] C. Srisathapornphat, C. Jaikaeo and C. Shen. Sensor Information Networking Architecture, International Workshop on Pervasive Computing (IWPC00), Toronto Canada, August 2000. [19] Q. Han and N. Venkatasubramanian. AutoSec: An integrated middleware framework for dynamic service brokering. IEEE Distributed Systems Online, 2(7), 2001. 9 [20] S. Li, S. Son, and J. Stankovic. Event detection services using data service middleware in distributed sensor networks. In Proceedings of the 2 nd International Workshop on Information Processing in Sensor Networks, April 2003. [21] Smart Messages project. March, 2003. http://discolab.rutgers.edu/sm. [22] T. Abdelzaher, B. Blum, Q. Cao, D. Evans, J. George, S. George, T. He, L. Luo, S. Son, R. Stoleru, J. Stankovic and A. Wood. EnviroTrack: Towards an Environmental Computing Paradigm for Distributed Sensor Networks. Technical report, Department of Computer Science, University of Virginia, 2003. [23] T. Liu and M. Martonosi. Impala: A middleware system for managing autonomic, parallel sensor systems. In ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP03), June 2003 [24] A. Murphy and W. Heinzelman, MiLan: Middleware linking applications and networks, Technical Report TR-795, University of Rochester, 2002. [25] X. Yu, K. Niyogi, S. Mehrotra and N. Venkatasubramanian, Adaptive middleware for distributed sensor networks, IEEE Distributed Systems Online, May 2003.

message service;publish-subscribe paradigm;message-oriented middleware model;environmental monitoring applications;application filters;context service;Middleware;constraint and query languages;design pattern;wireless sensor networks;application service;wireless sensor network

177

Seven Cardinal Properties of Sensor Network Broadcast Authentication

We investigate the design space of sensor network broadcast authentication . We show that prior approaches can be organized based on a taxonomy of seven fundamental proprieties, such that each approach can satisfy at most six of the seven proprieties. An empirical study of the design space reveals possibilities of new approaches, which we present in the following two new authentication protocols : RPT and LEA. Based on this taxonomy, we offer guidance in selecting the most appropriate protocol based on an application's desired proprieties. Finally, we pose the open challenge for the research community to devise a protocol simultaneously providing all seven properties.

INTRODUCTION Due to the nature of wireless communication in sensor networks, attackers can easily inject malicious data messages or alter the content of legitimate messages during multihop forwarding. Sensor network applications thus need to rely on authentication mechanisms to ensure that data from a valid source was not altered in transit. Authentication is thus arguably the most important security primitive in sensor network communication. Source authentication ensures a receiver that the message originates from the claimed sender, and data authentication ensures that the data from that sender was unchanged (thus also providing message integrity). When we use the term authentication we mean both source and data authentication. Broadcast authentication is a challenging problem. Furthermore, it is of central importance as broadcasts are used in many applications . For example, routing tree construction, network query, software updates, time synchronization, and network management all rely on broadcast. Without an efficient broadcast authentication algorithm , the base station would have to resort to per-node unicast messages, which does not scale to large networks. The practical-ity of many secure sensor network applications thus hinges on the presence of an efficient algorithm for broadcast authentication. In point-to-point authentication, authentication can be achieved through purely symmetric means: the sender and receiver would share a secret key used to compute a cryptographic message authentication code (MAC) over each message [15, 23]. When a message with a valid MAC is received, the receiver can be assured that the message originated from the sender. Researchers showed that MACs can be efficiently implemented on resource-constrained sensor network nodes [31], and find that computing a MAC function requires on the order of 1ms on the computation-constrained Berkeley mote platform [11, 14]. Authentication of broadcast messages in sensor networks is much harder than point-to-point authentication [1]. The symmetric approach used in point-to-point authentication is not secure in broadcast settings, where receivers are mutually untrusted. If all nodes share one secret key, any compromised receiver can forge messages from the sender. In fact, authenticated broadcast requires an asymmetric mechanism [1]. The traditional approach for asymmetric mechanisms is to use digital signatures, for example the RSA signature [34]. Unfortunately, asymmetric cryptographic mechanisms have high computation, communication, and storage overhead, making their usage on resource-constrained devices impractical for many applications . The property we need is asymmetry, and many approaches had been suggested for sensor network broadcast authentication. However , objectively comparing such approaches and selecting the most appropriate one for a given application is a non-trivial process, especially for an engineer not specialized in security. The goal of this work is to provide guidance for sensor network broadcast authentication by presenting a systematic investigation of the design space. We arrive at a taxonomy of seven fundamental properties, and present protocols that satisfy all but one property. The list of the desired properties are: 147 1. Resistance against node compromise, 2. Low computation overhead, 3. Low communication overhead, 4. Robustness to packet loss, 5. Immediate authentication, 6. Messages sent at irregular times, 7. High message entropy. If we remove any one of the above requirements, a viable protocol exists. Table 1 gives an overview of the seven approaches for addressing each case. We show that existing protocols, or small modifications thereof, make up for five of the seven possible cases. We also introduce novel approaches for addressing the final two cases: the RPT protocol to authenticate messages sent at regular times, and the LEA protocol to authenticate low-entropy messages. Finally, we pose the open challenge to the research community to design a broadcast authentication mechanism that satisfies all seven properties. Outline. The paper is organized as follows. We introduce the taxonomy of seven properties and discuss how current approaches can be organized based on our taxonomy in Section 2. Section 3 describes the TESLA broadcast authentication protocol and presents several extensions to increase its efficiency and robustness to DoS attacks. In Section 3.3, we introduce RPT, a novel protocol that authenticates synchronous messages. In Section 4, we introduce LEA, a novel protocol for efficient network broadcast authentication for low-entropy messages. Implementation and evaluation is discussed in Section 5. Finally, we present related work in Section 6 and our conclusions and future work in Section 7. TAXONOMY OF EXISTING PROTOCOLS In this section, we discuss the seven properties of broadcast authentication and describe possible approaches if we were to leave out one of the seven requirements. Node Compromise. Since sensor nodes are not equipped with tamper-proof or tamper-resistant hardware, any physical attacker would be able to physically compromise a node and obtain its cryptographic keys [5]. Since it is unlikely that tamper-proof hardware will be deployed on sensor motes in the near future, secure sensor network protocols need to be resilient against compromised nodes. However, if the nodes are deployed in a physically secured area (such as an attended army base), or if the application itself is resilient against malicious nodes, node compromise might not be an issue. If we assume no compromised nodes, all parties could maintain a network-wide key that is used to generate and verify a single Message Authentication Code (MAC) per message. If instead one can assume a low number of compromised nodes, a simple approach exists which uses a different key for each receiver and adds one MAC per receiver to each message. Unfortunately, this approach does not scale to large networks since a 10-byte MACs per receiver would result in prohibitively large messages. To trade off communication overhead with security, researchers propose a multi-MAC approach [3]. In their scheme, the sender chooses some number of random MAC keys, and distributes a subset of keys to each node. Every message carries one MAC with each key (assuming 10 bytes per MAC), 1 which adds a substantial overhead. If an attacker compromises a node, it can only forge a subset of MACs, thus with high probability, other nodes will be able to detect the forgery with their subset of keys. A variant of this approach was used to prevent malicious injection of messages in sensor networks [36, 37]. Computation Overhead. Sensor nodes have limited computation resources, so an ideal protocol would have low computation overhead for both sender and receiver. However, there exist scenarios where computation might not be a particularly critical issue. For example, it is conceivable that certain applications would only require authenticated broadcasts for a small number of packets. In such a case, the application engineer might be willing to allow for a small number of intensive computations. If we admit a high computation overhead, we can use digital signatures . RSA today requires at least a 1024-bit modulus to achieve a reasonable level of security, and a 2048-bit modulus for a high level of security [18]. ECC can offer the same level of security using 160-bit keys and 224-bit keys, respectively. Recent advancement in ECC signature schemes on embedded processors can perform signature verification using 160-bit ECC keys in about 1 second [10]. Although this represents a dramatic improvement over earlier public key cryptographic schemes [2, 4, 21], signature verification is still 3 orders of magnitude slower than MAC verification , while signature generation is 4 orders of magnitude slower. While we expect future sensor nodes to have more powerful processors , the energy constraints dictated by the limited battery resources will always favor the use of more efficient symmetric cryptographic primitives. Communication Overhead. Energy is an extremely scarce resource on sensor nodes, and as a result, heavily influences the design of sensor network protocols. In particular, radio communication consumes the most amount of energy, and thus protocols with high communication overhead are avoided if possible. However, in some settings (e.g., powered nodes) energy consumption is not an issue. Thus an authentication protocol that requires high communication overhead would be acceptable. If we admit a high communication overhead, we can leverage efficient one-time signatures that are fast to compute, but require on the order of 100200 bytes per signature. Examples include the Merkle-Winternitz (MW) signature which requires 230 bytes per signature [25, 26, 35] (we describe the MW signature in detail in Section 4.1), or the HORS signature, which requires around 100 bytes per signature [33]. The MW signature requires around 200 one-way function computations to verify a signature (which corresponds to roughly 200 ms computation time on a sensor node), while the HORS signature only requires 11 one-way function computations . The disadvantage of the HORS signature is that the public key is about 10 Kbytes, 2 whereas the public key for the MW signature is only 10 bytes. Signature generation is very efficient for both mechanisms, and can be reduced to a single hash function computation assuming a lookup table for the cryptographic values. We leverage the MW signature to construct the LEA broadcast authentication mechanism, which we present in Section 4. Message Reliability. Our fourth property is message reliability. Reliable message delivery is the property of a network such that valid messages are not dropped. Ultimately, message reliability is an applications issue - some applications require message reliability , while others do not. 1 An 80-bit MAC value achieves security comparable to a 1024-bit RSA signature [18]. 2 This is prohibitively large, since each public key of a one-time signature can be used to authenticate only a single message. 148 Desired property Approach if property is relaxed Resistance to node compromise Network-wide key Low computation overhead Digital signatures Low communication overhead One-time signatures Robustness to packet loss HORS + chaining of public keys Immediate authentication TESLA Messages sent at irregular times RPT, described in Section 3.3 High message entropy LEA, described in Section 4.2 Table 1: Overview of desired properties of broadcast authentication and approaches. The left column presents the desired property, and the right column presents the approach that achieves all properties but relaxes the property in its left column. The text describes each approach in more detail. If we have perfect message reliability, we can achieve efficient and immediate authentication by using the HORS signature in a special construction that combines multiple public keys [28]. In this construction, a public key is still 10 Kbytes, but a single public key can be used to authenticate almost arbitrarily many messages, as the public values are incrementally updated as signed messages are sent. The communication and computation costs are the same as for the HORS signature: 1 ms for signature generation, 11 ms for signature verification, and 100 bytes for the signature. Note that in such a scheme, an attacker can start forging HORS signatures if many packets are dropped. Authentication Delay. Depending on the application, authentication delay may influence the design of the sensor network protocol . For time-critical messages such as fire alarms, the receiver would most likely need to authenticate the message immediately. However, authentication delay is typically acceptable for non-time-critical messages. If we admit an authentication delay and assume that the receivers are loosely time synchronized with the sender, the TESLA broadcast authentication protocol only adds a 10 byte MAC and an optional 10 byte key to each message [31]. We review the TESLA protocol in detail in Section 3.1. To achieve a low computation overhead in the case of infrequent messages sent at unpredictable times, we need to extend the TESLA protocol to enable fast authentication of the keys in the one-way key chain. In Section 3.2 we present a more efficient key chain construction that enables efficient authentication in this case. Simultaneously, our approach protects TESLA against denial-of-service attacks by sending bogus key chain values. Synchronous Messages. Some applications send synchronous messages at regular and predictable times. For example, a key revocation list might be sent to the entire network everyday at noon. We extend the TESLA protocol to provide efficient and immediate authentication for synchronous messages sent at regular and predictable times. We name the protocol RPT (Regular-Predictable Tesla), and we present its details in Section 3.3. Message Entropy. So far, all schemes we describe authenticate unpredictable messages with high entropy. However, in practice, many protocols might only communicate with low-entropy messages . For example, in many applications, there are only a handful of valid commands that a base station can send to a sensor node. Therefore, these command packets could be considered as low-entropy messages. If we can assure a low upper bound on message entropy, we can leverage one-time signatures in constructions that provide message recovery, where the message is not hashed but directly encoded in the signature. We describe our new LEA protocol in Section 4. For messages with merely a single bit of entropy, we could employ the following optimization using two hash chains. One hash chain would correspond to messages of '1', while another would correspond to messages of '0'. The sender first sends the last value of both chains to the receivers in an authenticated manner (e.g., using one-time signatures or digital signatures). Next, whenever the sender wishes to send a '0', it would reveal the next value in the hash chain corresponding to '0'. The same is done for the hash chain corresponding to '1'. The receiver needs to keep state of the most recent value it received for each hash chain. Consequently, the receiver can easily verify the authenticity of new values by hashing them and comparing them against the most recent value of each hash chain. BROADCAST AUTHENTICATION WITH THE TESLA PROTOCOL In this section, we first present a brief overview of the TESLA protocol [29], the recommended broadcast authentication protocol if immediate authentication is not required. We improve the TESLA broadcast authentication protocol to provide efficient authentication for infrequent messages sent at unpredictable times (Section 3.2). In Section 3.3, we describe RPT, further modification of TESLA that provides immediate authentication for synchronous messages sent at regular and predictable times. 3.1 TESLA Overview The TESLA protocol provides efficient broadcast authentication over the Internet which can scale to millions of users, tolerate packet loss, and support real time applications [30]. Currently, TESLA is in the process of being standardized in the MSEC working group of the IETF for multicast authentication. TESLA has been adapted for broadcast authentication in sensor networks, the resulting protocol is called the TESLA broadcast authentication protocol [30, 31]. TESLA is used to secure routing information [17], data aggregation messages [12, 32], etc. We now overview the TESLA protocol, a detailed description is available in our earlier paper [31]. Broadcast authentication requires a source of asymmetry, such that the receivers can only verify the authentication information, but not generate valid authentication information. TESLA uses time for asymmetry. TESLA assumes that receivers are all loosely time synchronized with the sender up to some time synchronization error , all parties agree on the current time. Recent research in sensor network time synchronization protocols has made significant progress, resulting in time synchronization accuracy in the range of s [6, 7], which is much more accurate than the loose time synchronization required by TESLA. By using only symmetric cryptographic primitives, TESLA is very efficient and provides practical solutions for resource-constrained sensor networks. Figure 1 shows an example of TESLA authentication, and here is a sketch of the basic approach: 149 M j M j+1 M j+2 M j+3 M j+4 M j+5 M j+6 K i -1 K i K i +1 K i +2 F (Ki) F (Ki+1) F (Ki+2) F (Ki+3) Interval i - 1 Interval i Interval i + 1 Interval i + 2 time Figure 1: At the top of the figure is the one-way key chain (using the one-way function F). Time advances left-to-right. At the bottom of the figure, we can see the messages that the sender sends in each time interval. For each message, the sender uses the current time interval key to compute the MAC of the message. The sender splits up the time into time intervals of uniform duration. Next, the sender forms a one-way chain of self-authenticating keys, by selecting key K N of interval N at random , and by repeatedly applying a one-way hash function F to derive earlier keys. A cryptographic hash function, such as SHA-1 [27], offers the required properties. The sender assigns keys sequentially to time intervals (one key per time interval). The one-way chain is used in the reverse order of generation, so any key of a time interval can be used to derive keys of previous time intervals. For example, assuming a disclosure delay of 2 time intervals, key K i will be used to compute MACs of broadcast messages sent during time interval i, but disclosed during time interval i + 2. The sender defines a disclosure delay for keys, usually on the order of a few time intervals. The sender publishes the keys after the disclosure time. The sender attaches a MAC to each message, computed over the data, using the key for the current time interval. Along with the message, the sender also sends the most recent key that it can disclose. In the example of Figure 1, the sender uses key K i +1 to compute the MAC of message M j +3 , and publishes key K i -1 assuming a key disclosure delay of two time intervals. Each receiver that receives the message performs the following operation. It knows the schedule for disclosing keys and, since the clocks are loosely synchronized, can check that the key used to compute the MAC is still secret by determining that the sender could not have yet reached the time interval for disclosing it. If the MAC key is still secret, then the receiver buffers the message. In the example of Figure 1, when the receiver gets message M j +3 , it needs to verify that the sender did not yet publish key K i +1 , by using the loose time synchronization and the maximum time synchronization error . If the receiver is certain that the sender did not yet reach interval i + 3, it knows that key K i +1 is still secret, and it can buffer the packet for later verification. Each receiver also checks that the disclosed key is correct (using self-authentication and previously released keys) and then checks the correctness of the MAC of buffered messages that were sent in the time interval of the disclosed key. Assuming the receiver knows the authentic key K i -2 , it can verify the authenticity of key K i -1 by checking that F (K i -1 ) equals K i -2 . If K i -1 is authentic, the receiver can verify the authenticity of buffered packets sent during time interval i - 1, since they were authenticated using key K i -1 to compute the MAC. One-way chains have the property that if intermediate keys are lost, they can be recomputed using later keys. So, even if some disclosed keys are lost due to packet loss or jamming attacks, a receiver can recover the key from keys disclosed later and check the authenticity of earlier messages. Along with each message M i , the sender broadcasts the TESLA authentication information. The broadcast channel may be lossy, but the sender would need to retransmit with an updated MAC key. Despite loss, each receiver can authenticate all the messages it receives . 3.2 Reducing Verification Overhead of TESLA Even though TESLA provides a viable solution for broadcast authentication in sensor networks, many challenges still remain. We describe the remaining challenges below and propose extensions and new approaches to address these challenges. Some applications broadcast messages infrequently at unpredictable times and the receivers may need to authenticate messages immediately. For example, a fire alarm event is infrequent and needs to be quickly distributed and authenticated. Unfortunately, when messages are infrequent, due to the one-way chain approach to verify the authenticity of keys, a receiver may need to compute a long chain of hash values in order to authenticate the key which could take several tens of seconds for verification. Such verification delays the message authentication significantly and may consume significant computation and energy resources. This approach also introduces a Denial-of-Service (DoS) attack: an attacker sends a bogus key to a receiver, and the receiver spends several thousands of one-way function computations (and several seconds) to finally notice that the sent key was incorrect. One approach is to periodically release TESLA keys and hence the work for verification of an infrequent message would be distributed over time. However, this approach wastes energy for periodic broadcast of TESLA keys. In the same vein, a sender can publish several keys in a packet to reduce the effect of DoS attacks by requiring a receiver to perform a small number of one-way function computations to incrementally authenticate each key of the one-way chain. An advantage of this approach is that it makes the DoS attack described above less attractive to an attacker, as a receiver would need to follow the one-way chain for a short interval only to detect a bogus key. Another approach to counteract the slow and expensive verification problem is to use a Merkle hash tree [24] instead of a one-way chain to authenticate TESLA keys. This approach has been suggested in another context [13]. For N keys, the tree has height d = log 2 (N) and along with each message, the sender sends d values to verify the key. Despite the logarithmic communication cost, this is still too large for most sensor networks: consider a network where we switch to a different hash tree every day, and we need a 150 k 2 k 5 k 8 k 11 k 14 k 17 k 20 k 23 k 1 k 4 k 7 k 10 k 13 k 16 k 19 k 22 k 0 k 3 k 6 k 9 k 12 k 15 k 18 k 21 F v 0 -7 = F(v 0 -3 || v 4 -7 ) v 0 -3 v 4 -7 v 01 v 23 v 45 v 67 v 0 v 1 v 2 v 3 v 4 v 5 v 6 v 7 Figure 2: Hash tree constructed over one-way chains of TESLA keys. key resolution of 1 second. The 86,400 keys that we need in one day require a tree of height 17. Assuming a hash output of 10 bytes, the sender would need to consequently add 170 bytes to each message for authentication (17 nodes at 10 bytes each). This is far too much for most sensor networks, where nodes typically communicate with messages shorter than 100 bytes. Splitting the load up into two messages is not a viable approach, because of the usually high packet loss rates in sensor networks. The receiver would only need to compute O (log(N)) operations for verification, 17 hash function computations in our example which requires around 17ms on current sensor nodes. To reduce the bandwidth overhead, we design a different approach that achieves lower message size at the cost of higher verification computation. our approach is to combine one-way chains with hash trees. Consider the structure that Figure 2 shows. We construct a hash tree over short one-way chains. If each one-way chain has a length of k, the verification cost is expected to be k /2+ log (N/k) (it is at most k + log(N/k)), and the communication cost is log (N/k). For a given upper bound on the verification time, we can thus minimize the communication overhead. Consider an upper bound on the verification time of approximately 500ms. We can set k = 2 9 = 512, thus the hash tree will have 8 levels, requiring 80 bytes per packet, making this an attractive approach for many applications . An alternative approach would be to construct a hash tree over the one-way key chain, where the every k'th key will be a leaf node of the hash tree (for example, in Figure 2, the value k 0 would be derived from the previous leaf node k 0 = F(v 1 )). The advantage of this approach is that a sender would not need to send the hash tree values along with a message, as a value can be authenticated by following the one-way chain to the last known value. However, if the sender did not send out any message during an extended time period, that authentication would be computationally expensive and thus the sender can choose to also send the hash tree nodes along for fast verification. This approach would also prevent DoS attacks since the verification is very efficient. M i M i K i -1 K i K i +1 F (Ki) F (Ki+1) Interval i - 1 Interval i time T i -1 T i T i +1 Figure 3: This figure shows authentication of one message in the RPT protocol. Message M i = MAC K i (M i ) , and message M i = M i ,K i . 3.3 RPT: Authenticating Messages Sent at Regular and Predictable Times As described in our taxonomy in Section 2, one additional property in the design space of broadcast authentication is to authenticate asynchronous messages sent at irregular and unpredictable times. All protocols described so far can achieve this property. However, if we were to remove this requirement, new possible approaches exist that can only authenticate messages sent at regular and predictable times, yet satisfy all of the other cardinal properties defined in our taxonomy. In this section, we introduce our design of one such protocol called RPT, a modification of the TESLA protocol. In practice, many protocols send synchronous messages at regular and predictable times. The plaintext of these messages are often known by the sender a priori. In particular, messages containing meta-data are especially well-suited for this type of communication . For example, a base-station often performs key update or time re-synchronization at a preset time of day. In these examples, the sender knows exactly what message needs to be sent at a particular time, but the protocol dictates that such messages cannot be sent until a pre-specified time. Consider an application that broadcasts a message every day at noon to all nodes. If we use standard TESLA with one key per 151 day, it would take one day to authenticate the message, since the receivers would need to wait for the disclosed key one day later. On the other hand, if we use many keys, for example, one key per second, it would require 86 ,400 keys per day (not using the optimization we presented in the previous section), and a sensor node would require an expected time of 43 seconds to verify the authenticity of the key. Hence, if messages are sent at very regular time intervals, we can streamline TESLA to immediately authenticate these messages. The RPT protocol (Regular-Predictable TESLA) achieves immediate authentication for messages sent at regular and predictable times. Consider a message that needs to be sent at times T i = T 0 + i D. The sender creates a one-way key chain, and assigns one key to each time interval of duration D. We assume that the sender knows the content of the message M i to be broadcast at time T i by time T i -, where is the maximum network broadcast propagation delay plus the maximum time synchronization error. At time T i - , the sender broadcasts message MAC K i (M i ) , and at time T i the sender broadcasts M i ,K i . As soon as the receiver receives the first message, it needs to verify the safety condition that key K i is still secret, given its current time and the maximum time synchronization error. When receiving the second message, the receiver first verifies the key K i . If the key is correct it verifies the MAC, and if the MAC is correct it is assured that M i is authentic. Note that this approach does not exhibit any authentication delay, as the receiver can immediately authenticate M i immediately after reception. At first glance, it may appear that RPT is susceptible to a denial-of -broadcast attack, where an attacker sends a large number of forged MACs around the time the legitimate is sent out. This problem had been studied and addressed in previous work [16]. However , it is not easy to evaluate how well this works in practice. BROADCAST AUTHENTICATION WITH ONE-TIME SIGNATURES Another way to achieve asymmetric authentication is through the use of one-time signatures. A one-time signature is much faster to generate and verify than general purpose signatures, but the private key associated with the signature can be used to sign only a single message, otherwise the security degrades and an attacker could forge signatures. Unlike TESLA, time synchronization is not necessary and authentication is immediate. Moreover, one-time signatures achieve non-repudiation in addition to authentication, which enables a node to buffer a message and retransmit it later. The receiver of the retransmitted message can still authenticate the message . One-time signatures are advantageous in applications with infrequent messages at unpredictable times, as they do not add computation to the receiver based upon the time at which the message is received. This makes them resilient to many forms of DoS attacks . We now present an overview of one-time signatures, and then present our LEA broadcast authentication protocol for authentication of low-entropy messages in Section 4.2. 4.1 One-Time Signatures Overview The Merkle-Winternitz signature was first drafted by Merkle [25, 26], and was later also used by Even, Goldreich, and Micali [8], and more recently also by Rohatgi for efficient stream authentication [35]. We briefly describe the basic principle of the Merkle-Winternitz signature. A Merkle-Winternitz signature relies on efficient one-way functions to construct a DAG (directed acyclic graph) to encode a signature . Each edge between two vertices (v 1 v 2 ) in the graph represents an application of the one-way function, where the value of the end node is the result of the one-way function applied to the beginning node (v 2 = F(v 1 ), where F represents the one-way function ). End nodes with multiple incoming edges take on the value of the hash of the concatenation of predecessor nodes. The initial values of the graph represent the private key, and the final value represents the public key. To achieve a secure one-time signature, the property of the signature encoding is that an attacker would have to invert at least one one-way function to sign any other value (i.e., forge a signature). We now discuss an example of a signature graph and signature encoding. Figure 4(a) depicts the one-time signature. A one-way hash chain of length 4 can be used to encode the values 0 - 3. For this signature chain, we will use the convention that the 1st value s 3 in the chain encodes the value 3, the second 2, etc. The signer derives the value s 3 from a randomly generated private key K priv by using a Pseudo-Random Function (PRF), e.g., s 3 = PRF K priv (0). 3 To prevent signature forgery (as we will explain later), the sender also creates a checksum chain c 0 ...c 3 , deriving value c 0 also from the private key, e.g., c 0 = PRF K priv (1), and again using the one-way function to derive the other values, e.g., c 1 = F(c 0 ). The application of the one-way function on s 0 and c 3 forms the public key: K pub = F(s 0 || c 3 ). To sign value i, where 0 i 3, the signer uses values s i and c i as the signature. To verify the signature s i and c i , the receiver follows the one-way chains and recomputes the public key as follows, with F 0 (x) = x: K pub = F(F i (s i ) || F 3 -i (c i )) A signature is correct if the recomputed value matches the public key For example, consider a signature on value 2: s 2 and c 2 . To verify, the receiver checks that K pub = F(F(F(s 2 )) || F(c 2 )). An attacker who wishes to forge a signature is forced to invert at least one one-way function (since the indices of the checksum chain run in direction opposite to the signature chain). Assuming the one-way function is secure, an attacker cannot invert the function to forge a signature, hence, the signature is secure. In practice, we can use a secure cryptographic hash function for our one-way function, but for increased efficiency we use a block cipher in hash mode, for example the commonly used Matyas-Meyer-Oseas mode [22]. Using two chains achieves a secure one-time signature, but does not scale well to sign a large number of bits. If we use two chains, a signature on 32 bits would require a chain 2 32 values long, which has a very high overhead to generate and verify. Instead, if more than one chain is used, each chain can encode some number of bits of the signature. For example, one could encode an 8 bit number by using four chains of length 4 to encode two bits in each chain. The public key is derived from the last value on all the chains. However, in this scheme, we would still need an additional 4 chains of length 4 to encode the values in the opposite direction to prevent forgeries. The Merkle-Winternitz signature reduces the number of checksum chains, in that the redundant checksum chains do not encode the actual value, but instead encode the sum of the values on the signature chains. As explained in detail by Merkle [25,26], the checksum chain encodes the sum of all values in the signature chains. Assuming k signature chains that sign m bits each, the maximum sum would be k (2 m -1), thus the checksum chains would encode 3 We use a block cipher to implement the PRF efficiently. A block cipher is a good PRF as long as we do not use the PRF to compute more than O (2 n ) operations with the same key, where n is the blocksize in bits. Since we only perform a few operations, the block cipher is a secure and efficient PRF. 152 K priv K pub s 0 s 1 s 2 s 3 c 3 c 2 c 1 c 0 (a) Simple one-time signature to sign 2 bits. F K priv K pub s 0 ,0 s 0 ,1 s 0 ,2 s 0 ,3 s 1 ,0 s 1 ,1 s 1 ,2 s 1 ,3 s 2 ,0 s 2 ,1 s 2 ,2 s 2 ,3 s 3 ,0 s 3 ,1 s 3 ,2 s 3 ,3 c 0 ,3 c 0 ,2 c 0 ,1 c 0 ,0 c 1 ,3 c 1 ,2 c 1 ,1 c 1 ,0 (b) Merkle-Winternitz one-time signature. This construction can sign 8 bits. Figure 4: This figure illustrates the Merkle-Winternitz one-time signature. log 2 k (2 m - 1) bits, providing for a significant savings. This approach still ensures that an attacker would have to invert at least one one-way function to forge a signature. Using signature chains with 4 values, a signature on n bits will then require n /2 signature chains. Since each chain encodes up to the value 3, the checksum chain at most needs to encode the value (n/2) 3 as the total sum; thus, the checksum chains need to sign log 2 (n/2 3) bits. If we also use checksum chains with 4 values, each checksum chain can again sign 2 bits and we need log 2 (n/2 3 )/2 checksum chains. Figure 4(b) shows an example of such a signature for signing 8 bits. Since the four signature chains can at most encode the number 3, the total sum is at most 4 3 = 12. Thus we only need 2 additional checksum chains to encode the 4 bits. Again, the indices in the checksum chain run opposite to the indices in the signature chain, to ensure that an attacker would have to invert at least one one-way function to forge a different signature. For the specific case of signing 80 bits, researchers suggest using chains of length 16 to encode 4 bits per chain [35]. Thus, we need 20 = 80/4 signature chains, and the checksum chains would need to encode at most values 0 ...300(= 20 15), which will require 9 bits, which again requires 3 checksum chains (where the third chain only requires 2 values to sign a single bit). 4 We now compute the computation overhead of signature verification . On average, signature verification requires following half the signature chains, which requires 8 one-way function computations . In the case of signing 80 bits with 20 signature chains, this will result in 160 one-way function computations. On average, the checksum chains require 16 one-way function computations, adding up to a total of 176 computations. 4.2 LEA: Authentication of Low-Entropy Messages If messages have high entropy, the one-time signature is still quite large in size. For example, if messages have 80 bits or more of entropy, the signer can hash the message before signing it. Using 4 We could also use 2 signature chains with 18 values each, as 18 2 = 324, saving one checksum chain. the construction we discussed in Section 4.1, signing an 80-bit hash value would yield a 230 bytes signature (or 184 bytes if we assume 8 byte long hash chain values). Unfortunately, this is still too large for current sensor networks. However, for messages with lower entropy, one-time signatures can be very effective. We thus present the LEA (Low-Entropy Authentication) protocol. The LEA protocol is based on Merkle-Winternitz one-time signatures, and periodically pre-distributes onetime public keys to receivers, and the sender uses the corresponding private keys to sign messages. The Merkle-Winternitz one-time signature is efficient for signing small numbers of bits. For example, assuming chains of length 16, to sign a message of n bits, we would need n /4 signature chains. Thus we need to encode log 2 (n/415) bits in the checksum chains, hence requiring log 2 (n/4 15)/4 additional checksum chains. For signing 8 bits, the signature would require 2 signature chains and 2 additional checksum chains to encode the sum ranging from 0 ...30, which would require 32 bytes assuming 8 byte values. Since communication cost is a premium, we could use a single checksum chain of length 30 to encode the checksum, thus saving 8 bytes. Hence, the total size of the authentication information would be 24 bytes. Since the size of the signature depends on the number of bits being signed, this method is preferable for situations where the message is a simple time critical command, such as an alarm, or a preset command. For example, to sign 128 different commands, we would only need one signature chain with 16 values, one signature chain with 8 values, and one checksum chain with 22 values. Assuming 8 byte values, the total signature length is 24 bytes. In some applications it may be possible to use a lossy compression algorithm to compress and quantize the data for the signature. This would allow the message to contain uncompressed data, but the attacher would only be able to change the message to a small degree. This could be helpful in commands which set the sensitivity of a motion sensor and the administrator is willing to allow a small error in the sensitivity which is actually received on the device. One of the main challenges of using one-time signatures is to dis-153 tribute one authentic public key for each signature to the receivers. Without an authentic public key an attacker could inject it's own public key and one-time signatures. This problem is easier than the original problem of general broadcast authentication because the public keys can be distributed far ahead of time at a predictable time. There are several methods by which this may be achieved. The simplest would be to distribute a set of k public keys to each receiver at bootstrap and these keys would be usable for the first k messages. If the lifetime of the devices compared to k is small, then the devices will not have to be re-bootstrapped. In general, the number of total messages is unknown. Thus, we design a mechanism to efficiently replenish authentic public keys after their use. We leverage the RPT protocol for this purpose. Nodes store a number of authentic public keys. The sender uses up one one-time signature (or one private key) per message it broadcasts . With this approach, all receivers can immediately authenticate the message. Periodically, the sender sends a RPT message at a regular time with new one-time public keys to replenish the used-up public keys at receivers. Since each public key is only 10 bytes long, this is an efficient approach. 4.3 Chaining Merkle-Winternitz Public Keys The above scheme illustrates an effective way to use TESLA in conjunction with Merkle-Winternitz signatures to provide fast and efficient authentication. The only drawback of using the Merkle-Winternitz one-time signature is that the public key can only be used once. Therefore when a TESLA authenticated message is sent at the beginning of the day authenticating k Merkle-Winternitz public keys, the sender and receiver are limited to only being able to authenticate k messages that day. The tradeoff is that choosing a large k uses up receiver memory resources. To circumvent this problem, rather than sending a fixed number of messages per interval, the public keys can be chained together in such a way that if more messages are needed they can be sent to the receiver and authenticated immediately. In this approach, the sender generates a large number of public and private keys for one-time signatures, labeling the public keys P 0 ,P 1 ,...,P n . These public keys are then combined, such that verification of one signature will automatically authenticate the public key of the next signature: V 0 = P 0 V 1 = H(P 1 || V 0 ) ... V i = H(P i || V i -1 ) ... V n = H(P n || V n -1 ) In this approach, the sender only needs to send the value V n authenticated with TESLA. The sender subsequently uses the private key that corresponds to the public key P n to sign a message, and sends value V n -1 along with the message. From the signature, the receiver can compute the public key P n , and together with the value V n -1 the receiver can authenticate the public key and V n -1 based on the trusted value V n . Now that the receiver trusts value V n -1 , the next public key P n -1 can be authenticated in the same way. This approach has the drawback that the message to be authenticated also needs to carry the value V n -1 increasing the message size by 810 bytes, and that message loss prevents later messages to be authenticated. We propose to use a hybrid approach: send k public keys authenticated with RPT each day, along with one value V n . If the sender needs to send more than k authenticated messages, it can then use the chained public keys after the first k messages. 0uA 500uA 1000uA 1500uA 2000uA 22 23 24 25 26 97 bits 60 bits 32 bits Power consumed (uA) vs. chain lengths Figure 5: The power consumption for an MSP430 sensor node receiving and validating Merkle-Winternitz signatures for varying signature chain lengths. IMPLEMENTATION AND PERFORMANCE EVALUATION Figure 5 illustrates the amount of energy required for using a Merkle-Winternitz signature for signing 32 bits, 60 bits, and 97 bits. In this example, the sensor is an 16-bit TI MSP430 processor running at 1 MHz, which can compute an 8-byte hash in approximately 5ms using RC5. This processor uses up 0.28 A per ms, and 3.8 A per byte received. Shown are the overall power consumption for five different chain lengths, 2 2 , 2 3 , 2 4 , 2 5 , 2 6 , and 2 7 . Table 2 shows the power consumption, validation times, and communication overhead for signing 60 bits with varying length chains. We implemented the PRF using the Helix stream cipher [9]. Unlike RC5, this cipher is not patented. It also features an efficient MAC construction which we use in our implementation of TESLA. The PRF is computed by using the input to the PRF as the key in encryption mode, and using the keystream as the output of the PRF. In this implementation, it takes about 8 ms to compute an 8-byte PRF. Since the signature generation requires comparable amount of computation as verification, generation of a 64-bit signature takes about 1.2 seconds and verification takes about 1 second in our un-optimized implementation. However, in this scheme, the public keys are generated in advance, so the sender must compute twice as many hashes because it must recompute the hashes when he wishes to actually compute a signature instead of simply generating the public key. This still makes it feasible for a sensor-node to act as the base station in our implementation, but generating a large amount of public keys becomes costly. The implementation is about 4k in size, 2k for the Helix assembly code, and 2k for the Merkle-Winternitz code (with code for both generation and validation ). RELATED WORK The TESLA protocol is a viable mechanism for broadcast authentication in sensor networks [31]. Unfortunately, this approach introduces an authentication delay and thus does not provide immediate authentication of messages which is necessary in applications 154 2 2 2 3 2 4 2 5 2 6 2 7 Power-cons (A) 1126.7 823.1 707.2 717.5 858.3 1163.2 Auth-time (ms) 332.5 442.5 680.0 1042.5 1762.5 2960.0 Overhead (bytes) 272 184 136 112 96 88 Table 2: Efficiency for signing a 60 bit value using Merkle-Winternitz one-time signature. with real-time requirements. Moreover, the TESLA approach has some denial-of-service vulnerabilities, which we address in this paper . Liu and Ning subsequently improved the efficiency of bootstrapping new clients, using multiple levels of one-way key chains [20]. This work also discussed the DoS attack explained in Section 3.2. Liu et al. also outlines a potential approach to authenticate commitment messages with Merkle hash trees [19]. Several researchers have investigated the use of asymmetric cryptographic techniques in sensor networks. Unfortunately, the overhead is too high to warrant use of such techniques for per-packet broadcast authentication. Such schemes were discussed in Section 2 in the context of protocols with high computation overhead. CONCLUSION We have studied viable and efficient solutions for efficient broadcast authentication in sensor networks. This problem is challenging due to the highly constrained nature of the devices and the unpredictable nature of communication in many environments. Since the authentication of broadcast messages is one of the most important security properties in sensor networks, we need to study viable approaches for a variety of settings. We establish a set of properties of broadcast authentication: security against compromised nodes, low computation and communication cost, immediate authentication (with no receiver delay), authentication of unpredictable messages with high entropy, and robustness to packet loss. We present a viable protocol for each case where we relax one property, and pose the open challenge to find a protocol that satisfies all properties REFERENCES [1] D. Boneh, G. Durfee, and M. Franklin. Lower bounds for multicast message authentication. In Advances in Cryptology -- EUROCRYPT '01, pages 434450, 2001. [2] M. Brown, D. Cheung, D. Hankerson, J. Lopez Hernandez, M. Kirkup, and A. Menezes. PGP in constrained wireless devices. In Proceedings of USENIX Security Symposium, August 2000. [3] R. Canetti, J. Garay, G. Itkis, D. Micciancio, M. Naor, and B. Pinkas. Multicast security: A taxonomy and some efficient constructions. In INFOCOMM'99, pages 708716, March 1999. [4] J. Deng, R. Han, and S. Mishra. A performance evaluation of intrusion-tolerant routing in wireless sensor networks. In Proceedings of IEEE Workshop on Information Processing in Sensor Networks (IPSN), April 2003. [5] J. Deng, C. Hartung, R. Han, and S. Mishra. A practical study of transitory master key establishment for wireless sensor networks. In Proceedings of the First IEEE/CreateNet Conference on Security and Privacy for Emerging Areas in Communication Networks (SecureComm), 2005. [6] Jeremy Elson, Lewis Girod, and Deborah Estrin. Fine-grained network time synchronization using reference broadcasts. In Proceedings of Symposium on Operating Systems Design and Implementation (OSDI), December 2002. [7] Jeremy Elson and Kay Romer. Wireless sensor networks: A new regime for time synchronization. In Proceedings of Workshop on Hot Topics In Networks (HotNets-I), October 2002. [8] S. Even, O. Goldreich, and S. Micali. On-line/off-line digital signatures. In Advances in Cryptology -- CRYPTO '89, volume 435, pages 263277, 1990. [9] Niels Ferguson, Doug Whiting, Bruce Schneier, John Kelsey, Stefan Lucks, and Tadayoshi Kohno. Helix: Fast encryption and authentication in a single cryptographic primitive. In Proceedings of the International Workshop on Fast Software Encryption (FSE 2003), 2003. [10] V. Gupta, M. Millard, S. Fung, Y. Zhu, N. Gura, H. Eberle, and S. C. Shantz. Sizzle: A standards-based end-to-end security architecture for the embedded internet. In Proceedings of the Third IEEE International Conference on Pervasive Computing and Communication (PerCom), 2005. [11] Jason Hill, Robert Szewczyk, Alec Woo, Seth Hollar, David E. Culler, and Kristofer S. J. Pister. System architecture directions for networked sensors. In Proceedings of Architectural Support for Programming Languages and Operating Systems (ASPLOS IX), pages 93104, 2000. [12] Lingxuan Hu and David Evans. Secure aggregation for wireless networks. In Workshop on Security and Assurance in Ad hoc Networks, January 2003. [13] Yih-Chun Hu, Adrian Perrig, and David B. Johnson. Packet leashes: A defense against wormhole attacks in wireless networks. In Proceedings of IEEE INFOCOM, April 2003. [14] J. M. Kahn, R. H. Katz, and K. S. Pister. Mobile networking for smart dust. In Proceedings of ACM/IEEE Conference on Mobile Computing and Networking (MobiCom), August 1999. [15] C. Karlof, N. Sastry, and D. Wagner. TinySec: A link layer security architecture for wireless sensor networks. In ACM SenSys, November 2004. [16] Chris Karlof, Naveen Sastry, Yaping Li, Adrian Perrig, and J. D. Tygar. Distillation codes and applications to dos resistant multicast authentication. In Proceedings of the Symposium on Network and Distributed Systems Security (NDSS), November 2004. [17] Chris Karlof and David Wagner. Secure routing in wireless sensor networks: Attacks and countermeasures. In Proceedings of First IEEE International Workshop on Sensor Network Protocols and Applications, May 2003. [18] A. Lenstra and E. Verheul. Selecting cryptographic key sizes. Journal of Cryptology, 14(4):255293, 2001. [19] D. Liu, P. Ning, S. Zhu, and S. Jajodia. Practical broadcast authentication in sensor networks. In Proceedings of The 2nd Annual International Conference on Mobile and Ubiquitous Systems: Networking and Services , November 2005. [20] Donggang Liu and Peng Ning. Efficient distribution of key chain commitments for broadcast authentication in 155 distributed sensor networks. In Proceedings of Network and Distributed System Security Symposium (NDSS), pages 263276, February 2003. [21] David Malan, Matt Welsh, and Michael Smith. A public-key infrastructure for key distribution in TinyOS based on elliptic curve cryptography. In Proceedings of IEEE International Conference on Sensor and Ad hoc Communications and Networks (SECON), October 2004. [22] S. Matyas, C. Meyer, and J. Oseas. Generating strong one-way functions with cryptographic algorithm. IBM Technical Disclosure Bulletin, 27:56585659, 1985. [23] A. Menezes, P. van Oorschot, and S. Vanstone. Handbook of Applied Cryptography. CRC Press, 1997. [24] R. Merkle. Protocols for public key cryptosystems. In Proceedings of the IEEE Symposium on Research in Security and Privacy, pages 122134, April 1980. [25] R. Merkle. A digital signature based on a conventional encryption function. In Advances in Cryptology -- CRYPTO '87, pages 369378, 1988. [26] R. Merkle. A certified digital signature. In Advances in Cryptology -- CRYPTO '89, pages 218238, 1990. [27] National Institute of Standards and Technology (NIST), Computer Systems Laboratory. Secure Hash Standard. Federal Information Processing Standards Publication (FIPS PUB) 180-2, February 2004. [28] A. Perrig. The BiBa one-time signature and broadcast authentication protocol. In Proceedings of ACM Conference on Computer and Communications Security (CCS), pages 2837, November 2001. [29] A. Perrig, R. Canetti, J. D. Tygar, and D. Song. Efficient authentication and signature of multicast streams over lossy channels. In Proceedings of the IEEE Symposium on Research in Security and Privacy, pages 5673, May 2000. [30] A. Perrig, R. Canetti, J. D. Tygar, and D. Song. The TESLA broadcast authentication protocol. RSA CryptoBytes, 5(Summer), 2002. [31] Adrian Perrig, Robert Szewczyk, Victor Wen, David Culler, and J. D. Tygar. SPINS: Security protocols for sensor networks. In Proceedings of ACM Conference on Mobile Computing and Networks (MobiCom), pages 189199, 2001. [32] Bartosz Przydatek, Dawn Song, and Adrian Perrig. SIA: Secure information aggregation in sensor networks. In Proceedings of the First ACM International Conference on Embedded Networked Sensor Systems (SenSys 2003), pages 255265, November 2003. [33] Leonid Reyzin and Natan Reyzin. Better than BiBa: Short one-time signatures with fast signing and verifying. In Proceedings of Conference on Information Security and Privacy (ACISP), July 2002. [34] R. Rivest, A. Shamir, and L. Adleman. A method for obtaining digital signatures and public-key cryptosystems. Communications of the ACM, 21(2):120126, February 1978. [35] P. Rohatgi. A compact and fast hybrid signature scheme for multicast packet. In Proceedings of the 6th ACM Conference on Computer and Communications Security, pages 93100. ACM Press, November 1999. [36] F. Ye, H. Luo, S. Lu, and L. Zhang. Statistical en-route filtering of injected false data in sensor networks. In Proceedings of IEEE INFOCOM, March 2004. [37] S. Zhu, S. Setia, S. Jajodia, and P. Ning. An interleaved hop-by-hop authentication scheme for filtering false data in sensor networks. In Proceedings of IEEE Symposium on Security and Privacy, pages 259271, May 2004. 156

Sensor Network;Broadcast Authentication;Taxonomy

178

Significance of gene ranking for classification of microarray samples

Many methods for classification and gene selection with microarray data have been developed. These methods usually give a ranking of genes. Evaluating the statistical significance of the gene ranking is important for understanding the results and for further biological investigations, but this question has not been well addressed for machine learning methods in existing works. Here, we address this problem by formulating it in the framework of hypothesis testing and propose a solution based on resampling. The proposed r-test methods convert gene ranking results into position p-values to evaluate the significance of genes. The methods are tested on three real microarray data sets and three simulation data sets with support vector machines as the method of classification and gene selection. The obtained position p-values help to determine the number of genes to be selected and enable scientists to analyze selection results by sophisticated multivariate methods under the same statistical inference paradigm as for simple hypothesis testing methods.

INTRODUCTION AN important application of DNA microarray technologies in functional genomics is to classify samples according to their gene expression profiles, e.g., to classify cancer versus normal samples or to classify different types or subtypes of cancer. Selecting genes that are informative for the classification is one key issue for understanding the biology behind the classification and an important step toward discovering those genes responsible for the distinction . For this purpose, researchers have applied a number of test statistics or discriminant criteria to find genes that are differentially expressed between the investigated classes [1], [2], [3], [4], [5], [6], [7]. This category of gene selection methods is usually referred to as the filtering method since the gene selection step usually plays the role of filtering the genes before doing classification with some other methods. Another category of methods is the so-called wrapper methods, which use the classification performance itself as the criterion for selecting the genes and genes are usually selected in a recursive fashion [8], [9], [10], [11], [12]. A representative method of this category is SVM-RFE based on support vector machines (SVM), which uses linear SVM to classify the samples and ranks the contribution of the genes in the classifier by their squared weights [10]. All these selection methods produce rankings of the genes. When a test statistic, such as the t-test, F-test, or bootstrap test, is used as the criterion, the ranking is attached by p-values derived from the null distribution of the test statistic, which reflects the probability of a gene showing the observed difference between the classes simply due to chance. Such p-values give biologists a clear understanding of the information that the genes probably contain. The availability of the p-value makes it possible to investigate the microarray data under the solid framework of statistical inference and many theoretical works have been built based on the extension of the concept of p-value, such as the false discovery rate (FDR) study [13]. Existing gene selection methods that come with p-values are of the filtering category and are all univariate methods. To consider possible combinatorial effects of genes, most wrapper methods adopt more sophisticated multivariate machine learning strategies such as SVMs and neural networks. These have been shown in many experiments to be more powerful in terms of classification accuracy. However, for gene selection, the gene rankings produced with these methods do not come with a measure of statistical significance. The ranking is only a relative order of genes according to their relevance to the classifier. There is no clear evaluation of a gene's contribution to the classification. For example, if a gene is ranked 50th according to its weight in the SVM classifier, it is only safe to say that this gene is perhaps more informative than the gene ranked at 51st. However, there is no way to describe how significant it is and there is no ground to compare the information it contains with a gene also ranked as 50th by the same method in another experiment . This nature of relative ranking makes it hard to interpret and further explore the gene selection results achieved with such advanced machine learning methods. For example, it is usually difficult to decide on the proper number of genes to be selected in a specific study with such machine learning methods. Most existing works usually select a subgroup of genes with some heuristi-cally decided numbers or thresholds [6], [8], [10]. The advanced estimation techniques, such as FDR, based on significance measures do not apply for such methods. Evaluating the statistical significance of the detected signal is the central idea in the paradigm of statistical inference from experimental data. There should be an equivalent study on those machine-learning-based multivariate gene selection methods which produce ranks according to their own criteria. Strategies such as permutation can be utilized to assess the significance of the classification accuracy, but they do not measure the significance of the selected genes directly. Surprisingly, this question has not been addressed by the statistics or bioinformatics community in existing literature. We therefore propose that the question be asked in this way: For an observed ranking of genes by a certain method, what is the probability that a gene is ranked at or above the observed position due to chance (by the same method) if the gene is, in fact, not informative to the classification? ("Being informative" is in the sense of the criteria defined or implied by the classification and ranking method. It may have different meanings for different methods.) We call this problem the significance of gene ranking or feature ranking. We raise this problem in this paper and describe our strategy toward a solution. The problem is discussed in the context of microarray classification of cancer samples, but the philosophy and methodology is not restricted to this scenario. THE SIGNIFICANCE OF RANKING PROBLEM Suppose a microarray data set contains m cases X fx i ; i 1; ; m g. Each case is characterized by a vector of the expression values of n genes x i x i1 ; x i2 ; ; x in T 2 R n ; i 1; ; m . Each gene is a vector of their expression values across the cases g j x 1j ; x 2j ; ; x mj T and we denote the set of all genes a s G fg j ; j 1; ; n g. E a c h c a s e h a s a l a b e l y i f 1; 1g, i 1; ; m indicating the class it belongs to among the studied two classes, e.g., normal versus cancer, or two subtypes of a cancer, etc. Among the n genes, usually some are informative to the classification and some are not, but we do not know which genes are informative and which are not. For the convenience of description, we denote the set of informative genes as I G and that of the uninformative genes as U G . To simplify the problem, we assume that I G \ U G and I G [ U G G: 1 The goal is to build a classifier that can predict the classes ^ y i of the cases from x i and, at the same time, to identify the genes that most likely belong to I G . The former task is called classification and the latter one is called gene selection. In the current study, we assume that there has already been a ranking method RM which produces a ranking position for each gene according to some criterion assessing the gene's relevance with the classification: r j rank g j j x i ; y i ; i 1; ; m f g ; j 1; ; n 2 and we do not distinguish the specific types of the RM. The ranking is obtained based on the samples, thus r j is a random variable. The significance-of-ranking problem is to calculate the following probability: p r j 4 P rank g j r j jg j 2 U G ; 3 i.e., given a gene is uninformative to the classification (according to RM's criterion), what is the probability that it is ranked at or above the observed ranking position by the ranking method? We call this probability the p-value of a gene's ranking position or, simply, position p-value. This significance-of-ranking problem is distinct from existing statistics for testing differentially expressed genes in several aspects. It applies to more complicated multivariate classification and gene selection methods. Even when it is applied on gene ranking methods based on univariate hypothesis tests like t-test, the position p-value is different with the t-test p-value by definition. The t-test p-value of a gene is calculated from the expression values of this gene in the two sample sets by comparing with the assumed null distribution model when the gene is not differentially expressed in the two classes. The position p-value of a single gene, however, is defined on its context, in the sense that its value depends not only on the expression of this gene in the samples, but also on other genes in the same data set. A gene with the same expression values may have different position p-values in different data sets. The null distributions of ranks of uninformative genes are different in different data sets and, therefore, the foremost challenge for solving the problem is that the null distribution has to be estimated from the specific data set under investigation. THE R-TEST SCHEME The significance-of-ranking problem is formulated as a hypothesis testing problem. The null hypothesis is that the gene is not informative or g j 2 U G , the alternative hypothesis is g j 2 I G (the gene is informative) since we have assumption (1) and the statistic to be used to test the hypothesis is the ranking position. As in standard hypothesis testing, the key to solving the problem is to obtain the distribution of the statistic under the null hypothesis, i.e., the distribution of the ranks of uninformative genes: P r jg 2 U G : 4 For the extreme case when I G (all the genes are uninformative) and the ranking method is not biased, it is obvious that the null distribution is uniform. In a real microarray data set, however, usually some genes are informative and some are not, thus the uniform null distribution is not applicable. The null rank distribution in a practical investigation depends on many factors, including the separability of the two classes, the underlying number of informative genes, the power of the ranking method, the sample size, etc. The characteristics of these factors are not well understood in either statistics or biology and, therefore, we have to estimate an empirical null distribution from the data set itself. We propose to tackle this problem in two steps. First, we identify a set of putative uninformative genes (PUGs) which are a subset of U G . This is possible in practice because, although we do not know U G , discovering a number of genes that are irrelevant to the classification is usually not hard in most microarray data sets. We denote the identified subset as ZHANG ET AL.: SIGNIFICANCE OF GENE RANKING FOR CLASSIFICATION OF MICROARRAY SAMPLES 313 U 0G . The next step is to estimate the null distribution of ranks with the ranking positions of these PUGs. From the original data set, we resample L new data sets and apply the ranking method on each of them, producing, for each gene L, ranks r l j ; l 1; ; L . In our implementa-tion , we randomly resample half of the cases in the original data set each time. Other resampling schemes such as bootstrapping can also be used to obtain similar results according to our experiments (data not shown). Since, usually, the size of U G is much larger than that of I G (i.e., most genes are uninformative), if a gene tends to always be ranked at the bottom in the L rankings, it is very likely that the gene is an uninformative one. Thus, we define r j as the average position of gene j in the L rankings, r j 1 L X L l 1 r l j ; j 1; ; n 5 and select the bottom k genes with the largest r j as the PUGs to form U 0G , where k is a preset number. We rewrite U 0G as U 0k G when we need to emphasize the role of k in this procedure. We assume that U 0k G is a random sample of U G and r l j , l 1; ; L for g j 2 U 0k G is a random sample from the underlying null distribution of the ranks of uninformative genes. Thus, we have k L observations of the null distribution of ranks from which we estimate the null distribution using a histogram. More sophisticated non-parametric methods can be adopted to fit the distribution if necessary. We denote the estimated null distribution as ^ P r jg 2 U G P histogram r l j jg j 2 U 0k G : 6 With this estimated null distribution, the calculation of the position p-value is straightforward: For gene i with ranking position r i , ^ p r i ^ P r r i jg 2 U G P histogram r l j r i jg j 2 U 0k G : 7 Applying this on all the genes, we convert the ranking list to a list of position p-values reflecting the significance of the genes' being informative to the classification. This whole procedure for estimating the p-value of a ranking is illustrated in Fig. 1. We name this scheme the r-test and call the position p-value thus calculated the r-test p-value. COMPENSATION FOR BIAS IN THE ESTIMATED PUGs One important problem with the r-test scheme is the selection of the PUGs. Ideally, the ranks used to select PUGs and the ranks used to estimate null distribution should be independent. However, this is impractical in that 314 IEEE/ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS, VOL. 3, NO. 3, JULY-SEPTEMBER 2006 Fig. 1. The diagram showing the principle of r-test (pr-test). (a) A number (L) of new data sets is resampled from the original data set. A ranking is generated for each new data set with the ranking method, resulting in a total of L rankings. (b) The genes are ordered by their average positions of the L rankings. The horizontal axis is genes by this order and the vertical axis is ranking position in the resample experiments. For each gene, its ranking positions in the L experiments are drawn in a box plot, with a short dash in the middle showing the median. (c) From the bottom (rightmost) of the ordered gene list, k genes are selected as putative uninformative genes or PUGs. The box-plots of the ranks of the k PUGs are illustrated in this enlarged image. The null distribution of ranks of uninformative genes is to be estimated from these ranks. (d) An example null distribution estimated from PUGs. For each gene on the microarray, its actual ranking is compared with the null distribution to calculate the position p-value of the gene being noninformative. For mr-test and tr-test, the average position in the L rankings is used in the calculation of the p*-value. In tr-test, the PUGs are not selected from the bottom of the ranking, but rather from the genes with the largest t-test p-values. there is actually only one data set available. In our strategy, the same ranks are used to estimate both PUGs and the distribution of their ranks. This is an unplanned test in the sense that the PUGs are defined after the ranks are observed [14]. The PUGs in U 0k G are not an unbiased estimate of U G . In the extreme case when k is small, uninformative genes that are ranked higher are underrepresented and the ranks of U 0k G might represent only a tail of the ranks of U G on the right. If this happens, it will cause an overoptimistic estimation of the r-test p-values and result in more genes being claimed significant. Therefore, we propose two modified strategies to compensate for the possible bias. 4.1 Modified r-Test with Average Ranks In (7), the position p-value is calculated by comparing the rank r i of gene g i obtained from the whole data set with the estimated null distribution. Intuitively, when the sample size is small, one single ranking based on a small sample set can have a large variance, especially when all or most of the genes are uninformative. We propose replacing the rank r i by the r i defined in (5), i.e., to use the average position of gene g i in the L resampling experiments as the estimate of the true rank, and to calculate the position p-value with this estimated rank rather than the single observation of the rank: p r i ^ P r r i jg 2 U G P histogram r l j r i jg j 2 U 0k G : 8 The estimated null distribution is the distribution of single ranks of putative uninformative genes, but the r i to be compared to it is the averaged rank and (8) is no longer a p-value in the strict sense. Therefore, we name it p*-value instead and call this modified r-test the mr-test for convenience. Ideally, if a gene is informative to the classification and the ranking method can consistently rank the gene according this information on both the whole data set and on the resampled subsets, we'll have r i r i ; for g i 2 I G ; 9 in which case the p*-value will be equivalent to the original r-test p-value for these genes. In practice, when the sample size is small and the signal in some informative genes are not so strong, we always have r i r i when r i is small; therefore, the estimated ranks move toward the right on the rank distribution comparing with the single-run ranks, which, as an effect, can be a compensation to the bias in the estimated null distribution. (For the genes ranked in the lower half of the list, the averaged rank will move leftward, but these genes are not of interest to us in this study since we assume only a minority of the genes can be informative.) 4.2 Independent Selection of PUGs The ultimate reason that may cause biased estimation of the null distribution is that the PUGs in the above r-test scheme are estimated from the same ranking information as that being used for the calculation of the test statistics. A solution is to select a group of PUGs that are an unbiased sample from the U G . This is a big challenge because estimating the rank position distribution of U G is the question itself. When the ranking method RM is a multivariate one such as SVM-based methods, the ranking of the genes will not directly depend on the differences of single genes between the classes. We therefore can use a univariate statistic such as the t-test to select a group of nondifferentially expressed genes as the PUGs since these genes will have a high probability of not being informative as they are basically the same in the two classes. This selection will be less correlated with the ranking by RM. Applying a threshold on the t-test p-value p t , we select the PUG set U 0 G as: U 0 G 4 g j jg j 2 G; p t g j 10 and estimate the null position distribution according the ranking of the U 0 G genes by RM in the resampled data: ^ P t r jg 2 U G P histogram r l j jg j 2 U 0 G : 11 The position p*-value of a gene ranked on average at r i is calculated as: ^ p r i ^ P t r r i jg 2 U G P histogram r l j r i jg j 2 U 0 G : 12 For the convenience of discussion, we call this strategy the tr-test and call the primary r-test defined by (7) the pr-test. We view the pr-test, mr-test, and tr-test as three specific methods under the general r-test scheme. It should be noted that if the ranking produced by RM is highly correlated with t-test ranking, the result of tr-test will be close to that of the original pr-test. On the other hand, since insignificant genes evaluated individually may not necessarily be uninformative when combined with certain other genes, the PUGs selected by (10) may include informative genes for RM. Therefore, the estimated null distribution may bias toward the left end in some situations, making the results overconservative. However, in the experiments described below, it is observed that the tr-test results are not sensitive to changes in the p-value cut-offs used for selecting PUGs (10), which is an implication that the method is not very biased. EXPERIMENTS WITH SVM ON REAL AND SIMULATED DATA 5.1 r-Test with SVM Gene Ranking Due to the good generalization ability of support vector machines (SVM) [15], they are regarded as one of the best multivariate algorithms for classifying microarray data [9], [10], [16]. In the experiments for r-test in this work, we adopted linear SVM as the ranking machine RM. The linear SVM is trained with all genes in the data set, producing the discriminate function f x w x b ; 13 where w P n i 1 i y i x i and i are the solutions of the following quadratic programming problem: L p 1 2 w k k 2 X n i 1 i y i x i w b X n i 1 i : 14 Following [10], the contribution of each gene in the classifier can be evaluated by ZHANG ET AL.: SIGNIFICANCE OF GENE RANKING FOR CLASSIFICATION OF MICROARRAY SAMPLES 315 DL p 1=2 @ 2 L p @w 2 i Dw i 2 w i 2 15 and, thus, the genes are ranked by w i 2 . There are other ways of assessing the relative contribution of the genes in a SVM classifier [17], but, since the scope of this paper is not to discuss the ranking method, we adopt the ranking criterion given in (15) here. The ranking only reflects the relative importance of the genes in the classifier, but cannot reveal how important each gene is. The r-test converts the ranking to position p-values (or p*-values) to evaluate the significance. 5.2 Data Sets Experiments were done on six microarray data sets: three real data sets and three simulated data sets. The leukemia data set [1] contains the expression of 7,129 genes (probe sets) of 72 cases, 47 of them are of the ALL class and 25 are of the AML class. The colon cancer data set [18] contains 2,000 genes of 62 cases, among which 40 are from colon cancers and 22 from normal tissues. These two data sets have been widely used as benchmark sets in many methodology studies. Another data set used in this study is a breast cancer data set [19] containing 12,625 genes (probe sets) of 85 cases. The data set is used to study the classification of two subclasses of breast cancer. Forty-two of the cases are of class 1 and 43 are of class 2. Simulated data sets were generated to investigate the properties of the methods in different situations. The first case is for an extreme situation where none of the genes are informative. The simulated data set contains 1,000 genes and 100 cases. The expression values of the genes are independently generated from normal distributions with randomized means and variations in a given range. The 100 cases are generated with the same model, but are assigned arbitrarily to two fake classes (50 cases in each class). So, the two classes are, in fact, not separable and all the genes are uninformative. We refer to this data set as the "fake-class" data set in the following description. Each of the other two simulated data sets also contains 1,000 genes and 100 cases of two classes (50 cases in each class). In one data set (we call it "simu-1"), 700 of the genes follow N(0, 1) for both classes and the 300 genes follow N(0.25, 1) for class 1 and N(-0.25, 1) for class 2. In the other data set (we call it "simu-2"), 700 of the genes follow N(0, 1) for both classes and the 300 genes follow N(0.5, 1) for class 1 and N(-0.5, 1) for class 2. With these two simulated data sets, we hope to mimic situations where there are weak and strong classification signals in the data. All the data sets except simu-1 and simu-2 were standardized to 0-mean and standard deviation 1 first across the cases and then across the genes. This is to prevent possible bias in the ranking affected by the scaling. In practical investigations, this step might not be needed or might need to be done in some other way according to the specific situation of the data and the specific ranking methods to be adopted. The six data sets used in our experiments represent different levels of separability of the investigated classes. For the leukemia data set, almost perfect classification accuracy has been achieved [1], [9], [10], so it represents a relatively easy classification task. For the colon cancer data set, the samples can still be well separated, but with some errors [10], [18]. The two subclasses studied in the breast cancer data set are hardly separable as observed in this data set, but it is believed that there could be some degree of separability [20], [21]. The fake-class simulation represents a situation where the two classes are completely nonseparable and the simu-1 and simu-2 simulation represents an ideal situation where separation is defined on a subset of the genes and the uninformative genes are i.i.d. To check the classification accuracy that can be achieved on these data sets, we randomly split them into independent training and test sets and applied linear SVM on them. These experiments were done 200 times for each data set and the classification accuracy obtained at different gene selection levels is summarized in Table 1. It can be seen that the accuracies are consistent with the reports in the literature and with the design of the simulations. (Note that the error rates reported here are independent test results based on only half of the samples for training, so they are larger than the cross-validation errors reported elsewhere. The scope of this paper is not to improve or discuss classification accuracy.) 316 IEEE/ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS, VOL. 3, NO. 3, JULY-SEPTEMBER 2006 TABLE 1 Separability of the Classes of the Six Data Sets 5.3 Number of Significant Genes According to the mr-Test and tr-Test We systematically experimented with the SVM-based pr-test, mr-test, and tr-test methods on the six data sets and studied the number of genes claimed as significantly informative with each method at various significant levels. The results of the pr-test are affected by different choices of the number k of selected PUGs (data not shown), indicating that the pr-test can be very biased unless we know the accurate number of informative genes. Therefore, we focus on the mr-test and tr-test in the following discussion. Table 2 shows the number of significant genes according to the mr-test at different p*-value levels, with different choices of ks on the six data sets. Comparing with the pr-test results, the mr-test results are less sensitive to changes in the number k. This is especially true when there are ideal classification signals, as in the simu-2 data, where we can see a more than 10-fold change of k causes only little variance in the estimated gene numbers. With p*-value levels from 0.001 to 0.1, the estimated significant gene numbers are all around the correct number (300). The claimed significant genes are all those true informative genes in the model when the estimated genes are less than 300. For the situations where the number of estimated informative genes is larger than 300, all the true informative genes are discovered. When the data are less ideal, we see that the results are stable within a smaller variation of k. More experiment results with larger variations in the choice of k are provided in the supplemental material, which can be found on the Computer Society Digital Library at http:// computer.org/tcbb/archives.htm. From Table 2, it can also be observed that the number of significant genes is not directly correlated with the classification accuracy. For example, the breast cancer data and fake-class data both look nonseparable according to the classification errors (Table 1), However, for the breast cancer data, more than 200 genes are identified as significantly informative among the 12,625 genes ( 1:6% ) at the p*-value = 0.01 level, but, for the fake-class data, this number is only about 0.4 percent of the 1,000 genes. Results of the tr-test with different t-test p-value cut-offs are shown in Table 3. It can be seen that different cut-offs result in different numbers of PUGs, but the variation in estimated position p*-values due to PUG number difference is even smaller than in the mr-test. This implies that the tr-test results are not biased by the selection of PUGs since, if the PUG selection was biased, different numbers of PUGs at t-test p-value cut-offs would have caused different degrees of bias and the results would have varied greatly. Comparing between Table 2 and Table 3, as well as the results in the supplemental material, which can be found on the Computer Society Digital Library at http://computer. org/tcbb/archives.htm, we observe that, for the mr-test, although there is a range of k for each data set in which the results are not very sensitive to variations of k, this range can be different with different data sets. On the other hand, for the tr-test, within the same ranges of cut-off t-test p-values , results on all the data sets show good consistency with regard to variations in the cut-off value. This makes the tr-test more applicable since users do not need to tune the parameter specifically to each data set. Comparing the number of genes selected by the tr-test and mr-test (Table 2 and Table 3), it is obvious that the tr-test is more stringent and selects much fewer genes than ZHANG ET AL.: SIGNIFICANCE OF GENE RANKING FOR CLASSIFICATION OF MICROARRAY SAMPLES 317 TABLE 2 The Number of Genes Selected at Various r-Test p*-Value Levels with SVM the mr-test on the real data sets. The differences are smaller on the simulated data. Similarly to the results of the mr-test, almost all the informative genes in simu-2 data can be recovered at p*-value levels from 0.001 to 0.05 and there are only a very few false-positive genes (e.g., the 307 genes selected at p*-value = 0.05 contains all the 300 true-positive genes and seven false-positive genes). This shows that the SVM method is good in both sensitivity and specificity in selecting the true informative genes for such ideal case, and both of the two r-test methods can detect the correct number of informative genes at a wide range of significance levels. For simu-1 data, not all the informative genes can be recovered in the experiments. This reflects the fact of the large overlap of the two distributions in this weak model. Many of original 300 "informative" genes are actually not statistically significant in the contexts of both univariate methods and multivariate methods. For the real data sets, there are no known answers for the "true" number of informative genes. The mr-test uses the tail in the ranking list to estimate the null distribution for assessing the significance of the genes on the top of the list, therefore there is a higher possibility of the p*-values being underestimated, although this has been partially compen-sated by using the average rank positions. Thus, the number of genes being claimed significant by mr-test might be overestimated. In this sense, the tr-test scheme provides a more unbiased estimation of the null distribution, which is supported by the decreased sensitivity to PUG numbers. With the tr-test, at the 0.05 p*-value level, we get about 410 significant genes from the 7,129 genes (5.75 percent) in the leukemia data. On the other two real data sets, the results tend to be too conservative: about 13 out of 2,000 genes (0.65 percent) in the colon cancer data and 50 out of 12,625 genes (0.4 percent) in the breast cancer data are claimed as significant at this level. It should be noted that the PUGs selected according to the t-test may contain informative genes for SVM, which considers the combined effects of genes. This will cause the number of genes called significant by the tr-test to be underestimated. This is especially true for data sets in which the major classification signal exists in the combinatorial effects of genes instead of differences in single genes. The correct answer may be somewhere between the two estimations of the tr-test and mr-test. When the signal is strong, the two estimations will be close as we see in the simu-2 data. In practice, one can choose which one to use according to whether the purpose is to discover more possibly informative genes or to discover a more manageable set of significant genes for follow-up investigations. DISCUSSION Statistical hypothesis testing is a fundamental framework for scientific inference from observations. Unfortunately, existing hypothesis testing methods are not sufficient to handle high-dimensional multivariate analysis problems arising from current high-throughput genomic and proteomic studies. Many new data mining techniques have been developed both in statistics and in the machine learning field. These methods are powerful in analyzing complicated high-dimensional data and helped greatly in functional genomics and proteomics studies. However, the analysis of the statistical significance of data mining results has not been paid enough attention. One reason might be that many methods are rooted in techniques aimed at solving problems in engineering and technological applications rather than in scientific discoveries. As an example, many machine-learning -based gene selection and classification methods may achieve very good performance in solving the specific classification problems, but the results are usually of a 318 IEEE/ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS, VOL. 3, NO. 3, JULY-SEPTEMBER 2006 TABLE 3 The Number of Genes Selected at Various Position p-Value Levels by tr-Test with SVM "black-box" type and judging the significance of the features being used for the classification was usually not deemed important. This fact compromises their further contribution in helping biologists to understand the mechanisms underlying the investigated disease classification. This paper raises the problem of the significance of gene rankings in microarray classification study and proposes a solution strategy called the r-test that converts the ranking of genes obtained with any method to position p-values (p*-values) that reflect the significance of the genes being informative. The concept of this question is important and the formulation and solution are challenging for several reasons as addressed in the paper. First of all, the definition of a gene being informative to the classification may not yet be completely clear for many classification methods. Even under the same criterion, there may not be a clear boundary between informative and uninformative genes. A biological status may be affected by several genes with different levels of contribution and it may affect the expression of many other genes. Differences between individuals and instrumental noises may make the genes that have no relation with the studied biological process show some relevancy in the limited samples. All these (and other) complexities make it hard to mathematically model microarray data. We propose the r-test methods based on intuitive reasoning under certain assumptions about the nature of the data. As shown in the experiments, the methods provide reasonable solutions, but the decision by the mr-test and tr-test method can be very different for some situations. Theoretically, rigorous methods are still to be developed. Under the proposed r-test framework, the key issue is the choice of putative uninformative genes or PUGs. Since the null distribution has to be estimated from the data themselves, avoiding bias in the estimation is the most challenging task. Besides the methods used by the mr-test and tr-test, we have also tried several other ways to tackle the problem, including selecting the PUGs according to the distribution of the ranks of all genes in the resample experiments, deciding the number of PUGs recursively according to the rank with an EM-like strategy, selection of an independent set of PUGs by fold-change, etc. Different resampling strategy has also been experimented with. Among these efforts, the reported mr-test and tr-test give the most satisfactory results. They both perform perfectly on ideal simulations. For practical cases, the mr-test has a tendency to be overoptimistic by claiming more significant genes and the tr-test has a tendency to be conservative by approving only a small number of significant genes. Note that both r-test schemes do not change the ranking itself; therefore, it is the role of the classification and gene selection method (the RM) to guarantee that the ranking itself is reasonable for the biological investigation. The r-test only helps to decide on the number of genes to be selected from the list at given significance levels. Since there is currently no theoretical solution to completely avoid estimation bias, one can make a choice between mr-test and tr-test results by balancing between the two opposite trends of possible biases according to the particular biological problem at hand. ACKNOWLEDGMENTS The authors would like to thank Drs. J.D. Iglehart and A.L. Richardson for providing them with their microarray data for the experiments. They would also like to thank the editor and reviewers for their valuable suggestions that contributed a lot to the work. They thank Dustin Schones for helping to improve their writing. This work is supported in part by NSFC projects 60275007, 60234020, and the National Basic Research Program (2004CB518605) of China. This work was performed while Chaolin Zhang was with the MOE Key Laboratory of Bioinformatics/Department of Automation, Tsinghua University, Beijing. Chaolin Zhang and Xuesong Lu contributed equally in this work and should be regarded as joint authors. The corresponding author is Xuegong Zhang. REFERENCES [1] T.R. Golub, D.K. Slonim, P. Tamayo, C. Huard, M. Gaasenbeek, J.P. Mesirov, H. Coller, M.L. Loh, J.R. Downing, M.A. Caligiuri, C.D. Bloomfield, and E.S. Lander, "Molecular Classification of Cancer: Class Discovery and Class Prediction by Gene Expression Monitoring," Science, vol. 286, no. 5439, pp. 531-537, 1999. [2] C.-A. Tsai, Y.-J. Chen, and J.J. Chen, "Testing for Differentially Expressed Genes with Microarray Data," Nuclear Acids Research, vol. 31, no. 9, p. e52, 2003. [3] M.S. Pepe, G. Longton, G.L. Anderson, and M. Schummer, "Selecting Differentially Expressed Genes from Microarray Experiments ," Biometrics, vol. 59, no. 1, pp. 133-142, 2003. [4] P. Broberg, "Statistical Methods for Ranking Differentially Expressed Genes," Genome Biology, vol. 4, no. 6, p. R41, 2003. [5] W. Pan, "A Comparative Review of Statistical Methods for Discovering Differentially Expressed Genes in Replicated Microarray Experiments," Bioinformatics, vol. 18, no. 4, pp. 546-554, 2002. [6] S. Ramaswamy, K.N. Ross, E.S. Lander, and T.R. Golub, "A Molecular Signature of Metastasis in Primary Solid Tumors," Nature Genetics, vol. 33, no. 1, pp. 49-54, 2003. [7] L.J. van 't Veer, H. Dai, M.J. van de Vijver, Y.D. He, A.A.M. Hart, M. Mao, H.L. Peterse, K. van der Kooy, M.J. Marton, A.T. Witteveen, G.J. Schreiber, R.M. Kerkhoven, C. Roberts, P.S. Linsley, R. Bernards, and S.H. Friend, "Gene Expression Profiling Predicts Clinical Outcome of Breast Cancer," Nature, vol. 415, no. 6871, pp. 530-536, 2002. [8] M. Xiong, X. Fang, and J. Zhao, "Biomarker Identification by Feature Wrappers," Genome Research, vol. 11, no. 11, pp. 1878-1887, 2001. [9] X. Zhang and H. Ke, "ALL/AML Cancer Classification by Gene Expression Data Using SVM and CSVM Approach," Proc. Conf. Genome Informatics, pp. 237-239, 2000. [10] I. Guyon, J. Weston, S. Barnhill, and V. Vapnik, "Gene Selection for Cancer Classification Using Support Vector Machines," Machine Learning, vol. 46, pp. 389-422, 2002. [11] C. Furlanello, M. Serafini, S. Merler, and G. Jurman, "Entropy-Based Gene Ranking without Selection Bias for the Predictive Classification of Microarray Data," BMC Bioinformatics, vol. 4, no. 1, p. 54, 2003. [12] H. Yu, J. Yang, W. Wang, and J. Han, "Discovering Compact and Highly Discriminative Features or Feature Combinations of Drug Activities Using Support Vector Machines," Proc. 2003 IEEE Bioinformatics Conf. (CSB '03), 2003. [13] J.D. Storey and R. Tibshirani, "Statistical Significance for Genome Wide Studies," Proc. Nat'l Academy of Science USA, vol. 100, no. 16, pp. 9440-9445, 2003. [14] R.R. Sokal and F.J. Rohlf, Biometry. San Francisco: Freeman, 1995. [15] V. Vapnik, The Nature of Statistical Learning Theory. Springer-Verlag , 1995. [16] S. Ramaswamy, P. Tamayo, R. Rifkin, S. Mukherjee, C.-H. Yeang, M. Angelo, C. Ladd, M. Reich, E. Latulippe, J.P. Mesirov, T. Poggio, W. Gerald, M. Loda, E.S. Lander, and T.R. Golub, "Multiclass Cancer Diagnosis Using Tumor Gene Expression Signatures," Proc. Nat'l Academy of Sciences, vol. 98, no. 26, pp. 15149-15154, 2001. ZHANG ET AL.: SIGNIFICANCE OF GENE RANKING FOR CLASSIFICATION OF MICROARRAY SAMPLES 319 [17] X. Zhang and W.H. Wong, "Recursive Sample Classification and Gene Selection Based on SVM: Method and Software Description ," technical report, Dept. of Biostatistics, Harvard School of Public Health, 2001. [18] U. Alon, N. Barkai, D.A. Notterman, K. Gish, S. Ybarra, D. Mack, and A.J. Levine, "Broad Patterns of Gene Expression Revealed by Clustering Analysis of Tumor and Normal Colon Tissues Probed by Oligonucleotide Arrays," Proc. Nat'l Academy of Sciences, vol. 96, no. 12, pp. 6745-6750, 1999. [19] Z.C. Wang, M. Lin, L.-J. Wei, C. Li, A. Miron, G. Lodeiro, L. Harris, S. Ramaswamy, D.M. Tanenbaum, M. Meyerson, J.D. Iglehart, and A. Richardson, "Loss of Heterozygosity and Its Correlation with Expression Profiles in Subclasses of Invasive Breast Cancers," Cancer Research, vol. 64, no. 1, pp. 64-71, 2004. [20] E. Huang, S.H. Cheng, H. Dressman, J. Pittman, M.H. Tsou, C.F. Horng, A. Bild, E.S. Iversen, M. Liao, and C.M. Chen, "Gene Expression Predictors of Breast Cancer Outcomes," The Lancet, vol. 361, no. 9369, pp. 1590-1596, 2003. [21] M. West, C. Blanchette, H. Dressman, E. Huang, S. Ishida, R. Spang, H. Zuzan, J.A. Olson Jr., J.R. Marks, and J.R. Nevins, "Predicting the Clinical Status of Human Breast Cancer by Using Gene Expression Profiles," Proc. Nat'l Academy of Sciences, vol. 98, no. 20, pp. 11462-11467, 2001. Chaolin Zhang received the BE degree from the Department of Automation at Tsinghua University, Beijing, China, in 2002. From 2002 to 2004, he worked as a graduate student on machine learning applications in microarray data analysis and literature mining at the MOE Key Laboratory of Bioinformatics, Tsinghua University . He is now a PhD student at Cold Spring Harbor Laboratory and the Department of Biomedical Engineering, the State University of New York at Stony Brook. Xuesong Lu received the BE degree from the Department of Automation, Tsinghua University, Beijing, China, in 2001. He is currently a PhD candidate in the Department of Automation and the MOE Key Laboratory of Bioinformatics at Tsinghua University, Beijing, China. His research interests include microarray data mining, gene network modeling, and literature mining. Xuegong Zhang received the PhD degree in pattern recognition and intelligent systems from Tsinghua University, Beijing, China, in 1994. He is currently a professor in the Department of Automation and the MOE Key Laboratory of Bioinformatics at Tsinghua University. His research interests include machine learning and pattern recognition, bioinformatics, computa-tional genomics, and systems biology. . For more information on this or any other computing topic, please visit our Digital Library at www.computer.org/publications/dlib. 320 IEEE/ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS, VOL. 3, NO. 3, JULY-SEPTEMBER 2006

Significance of gene ranking;gene selection;microarray data analysis;classification

179

Simplifying Flexible Isosurfaces Using Local Geometric Measures

The contour tree, an abstraction of a scalar field that encodes the nesting relationships of isosurfaces, can be used to accelerate isosurface extraction, to identify important isovalues for volume-rendering transfer functions, and to guide exploratory visualization through a flexible isosurface interface. Many real-world data sets produce unmanageably large contour trees which require meaningful simplification. We define local geometric measures for individual contours, such as surface area and contained volume, and provide an algorithm to compute these measures in a contour tree. We then use these geometric measures to simplify the contour trees, suppressing minor topological features of the data. We combine this with a flexible isosurface interface to allow users to explore individual contours of a dataset interactively.

Introduction Isosurfaces, slicing, and volume rendering are the three main techniques for visualizing three-dimensional scalar fields on a two-dimensional display. A recent survey [Brodlie and Wood 2001] describes the maturation of these techniques since the mid 1980s. For example, improved understanding of isosurfaces has produced robust definitions of watertight surfaces and efficient extraction methods . We believe that the same improved understanding and structuring leads to new interfaces that give the user better methods to select isosurfaces of interest and that provide a rich framework for data-guided exploration of scalar fields. Although key ideas in this paper apply to both isosurfaces and volume rendering, the immediate application is to isosurface rendering . An isosurface shows the surface for a fixed value (the isovalue ) of the scalar field and is the 3D analogue of equal-height contour lines on a topographic map. The contour tree represents the nesting relationships of connected components of isosurfaces, which we call contours, and is thus a topological abstraction of a scalar field. Since genus changes to surfaces do not affect the nesting relationship, they are not represented in the contour tree. Our contribution is to combine the flexible isosurface interface [Carr and Snoeyink 2003] with online contour tree simplification guided by geometric properties of contours to produce a tool for interactive exploration of large noisy experimentally-sampled data sets. An additional contribution is to draw attention to other potential applications of simplified contour trees, such as detail-preserving denoising, automated segmentation, and atlasing. Figure 1 shows a comparison between a conventional isosurface and a flexible isosurface extracted from the same data set after contour tree simplification. On the left, the outermost surface (the skull) occludes other surfaces, making it difficult to study structures inside the head. Moreover, the contour tree for this data set has over 1 million edges, making it impractical as a visual representation. Figure 2: The topographic map (2-d scalar field), surface rendering, and contour tree for a volcanic crater lake with a central island. A: a maximum on the crater edge; B: maximum of island in the lake; F: lake surface; C and D: saddle points. On the right is a flexible isosurface constructed using a simplified contour tree, laid out and coloured to emphasize the structure of the data set. Of particular interest is that there are no "magic numbers" embedded in the code. Instead, the surfaces shown were chosen directly from the simplified contour tree during exploration of this data set, with the level of simplification being adjusted as needed. The remainder of this paper is as follows. Section 2 reviews work on contour trees in visualization. Section 3 shows how to simplify the contour tree, and the effects on the data. Section 4 shows how to compute local geometric measures efficiently to guide simplification . Section 5 gives implementation details, and Section 6 reports results. Finally, Section 7 gives possible future extensions. Related Work Most of the relevant work deals with a topological structure called the contour tree that is becoming increasingly important in visualization . Section 2.1 reviews the contour tree and algorithms to compute it. Section 2.2 then reviews visualization tools that use the contour tree, while Section 2.3 reviews work on topological simplification and on efficient computation of geometric properties. 2.1 The Contour Tree For a scalar field f : IR 3 IR, the level set of an isovalue h is the set L (h) = {(x,y,z) | f (x,y,z) = h}. A contour is a connected component of a level set. As h increases, contours appear at local minima , join or split at saddles, and disappear at local maxima of f . Shrinking each contour to a single point gives the contour tree, which tracks this evolution. It is a tree because the domain IR 3 is simply-connected; in more general domains we obtain the Reeb graph [Reeb 1946], which is used in Morse theory [Matsumoto 2002; Milnor 1963] to study the topology of manifolds. Figure 2 shows a 2-dimensional scalar field describing a volcanic crater lake with a central island. The contour tree of this field is an abstract, but meaningful, depiction of the structure of all local maxima, minima, and saddle points, and gives clues to interesting contours. Individual contours are represented uniquely as points on the contour tree. For example, the isolines c 1 , c 2 , and c 3 are all at 2000m, but each has a unique location on the contour tree. The contour tree has been used for fast isosurface extraction [van Kreveld et al. 1997; Carr and Snoeyink 2003], to guide mesh simplification [Chiang and Lu 2003], to find important isovalues for transfer function construction [Takahashi et al. 2004b], to compute topological parameters of isosurfaces [Kettner et al. 2001], as an abstract representation of scalar fields [Bajaj et al. 1997], and to manipulate individual contours [Carr and Snoeyink 2003]. Algorithms to compute the contour tree efficiently in three or more dimensions have been given for simplicial meshes [van Kreveld et al. 1997; Tarasov and Vyalyi 1998; Carr et al. 2003; Chiang et al. 2002; Takahashi et al. 2004b] and for trilinear meshes [Pascucci and Cole-McLaughlin 2002]. Much of this work focusses on "clean" data from analytic functions or numerical simulation see for example [Bajaj et al. 1997; Takahashi et al. 2004b]. All of the topology in this data is assumed to be important and significant effort is expended on representing it accurately using trilinear interpolants [Pascucci and Cole-McLaughlin 2002] and topology-preserving simplifications [Chiang and Lu 2003]. In contrast, we are interested in noisy experimentally-acquired data such as medical datasets. We expect to discard small-scale topological features so that we can focus on large-scale features. We have therefore chosen to work with the well-known Marching Cubes cases [Lorenson and Cline 1987; Montani et al. 1994], and with approximate geometric properties. This paper does not turn on these choices, however, and can also be applied to trilinear interpolants and exact geometric properties. 2.2 Flexible Isosurfaces The contour spectrum [Bajaj et al. 1997] uses the contour tree to represent the topology of a field, alongside global measures of level sets such as surface area and enclosed volume. In contrast, the flexible isosurface interface [Carr and Snoeyink 2003] uses the contour tree actively instead of passively. The user selects an individual contour from the contour tree or from the isosurface display, then manipulates it. Operations include contour removal and contour evolution as the isovalue is changed, using the contour tree to track which contours to display. This interface depends on attaching isosurface seeds called path seeds to each edge of the contour tree so that individual contours can be extracted on demand. A major disadvantage of both these interfaces is that contour trees with more than a few tens of edges make poor visual abstractions. A principal contribution of this paper to simplify the contour tree while preserving the exploratory capabilities of the flexible isosurface . This requires that each point in a simplified contour tree represents an extractable contour. Moreover, extracted contours must evolve as smoothly as possible when the isovalue is adjusted. We satisfy these constraint with simplifications that have pre-dictable effects on the scalar field and geometric measures that iden-498 tify unimportant contour tree edges for simplification 2.3 Simplification and Geometric Measures The distinction between this paper and other work that simplifies contour trees or Reeb graphs is our emphasis on using tree structure for local exploration. [Takahashi et al. 2004a] simplify the contour tree by replacing three edges at a saddle point with a single new edge, based on the height of the edge. [Takahashi et al. 2004b] use the approximate volume of the region represented by the subtree that is discarded. Saddles are processed until only a few remain, then a transfer function is constructed that emphasizes the isovalues of those saddles. Our simplification algorithm extends this work to preserve local information such as isosurface seeds and to compute arbitrary geometric measures of importance. We also describe the effects of simplification on the scalar field. Since removing a leaf of the contour tree cancels out a local ex-tremum with a saddle, this form of simplification can be shown to be equivalent to topological persistence [Edelsbrunner et al. 2003; Edelsbrunner et al. 2002; Bremer et al. 2003] if the geometric measure used is height. For other measures, such as volume or hypervolume , the method described in this paper is necessary to define these properties, but thereafter, the process can optionally be described in terms of persistence. Moreover, work on persistence has focussed on the Morse complex , which is difficult to compute and segments data according to the gradient of the field. When the boundary of an object such as an organ is better described by a contour than by drainage, contour trees are more directly applicable than Morse complexes, and the additional overhead of working with the Morse complex is unnecessary . [Hilaga et al. 2001] have shown how to simplify the Reeb graph by progressive quantization of the isovalue to obtain a multi-resolution Reeb graph. This suffers from several drawbacks, in particular that it is strictly tied to a function value which is treated as height (or persistence). Extension to geometric measures of importance such as volume or hypervolume is therefore problematic. Moreover, the quantization used imposes serious restrictions on isosurface generation and the level of simplification, as well as generating artifacts related to the quantization. In particular, we note that this quantization process limits potential simplification to at most as many levels as there are bits in each input sample. Finally, this method is relatively slow: 15s is claimed for a 2-manifold input mesh with 10,000 vertices: extensions to 10,000,000+ sample volumetric data have not yet been published. Work also exists on computing geometric measures efficiently in large data sets. [Bentley 1979] defined problems to be decomposable if their solution could be assembled from the solutions of an arbitrary decomposition into subproblems. Decomposability has been used for a variety of problems, including computation of geometric properties of level sets [Bajaj et al. 1997] and extraction of isosurfaces [Lorenson and Cline 1987]. We use decomposability in Section 4 to compute local geometric measures. Contour Tree Simplification Given a contour tree and a scalar field, we apply graph simplification to the contour tree. This simplification can then be carried back to simplify the input data. Alternately, we can use the simplified contour tree to extract the reduced set of isosurfaces that would result if we had simplified the data. In this section, we describe the contour tree structure, the simplification operators, and the algorithms for simplification and isosurface extraction. 3.1 Contour Tree Structure A contour tree is the result of contracting every contour to a point. We use a simple tree structure in which every vertex is assigned a y-coordinate, and every edge is associated with the set of contours between its vertices. We store path seeds for generating individual contours, as in [Carr and Snoeyink 2003]. That is, we store a pointer to a monotone path that intersects all contours along the edge, which then serves as a seed to generate any given contour. In this section, we assume that each edge has a simplification value (weight) that indicates the edge's priority. Low priority edges are good candidates for simplification. 3.2 Basic Simplification Operations We simplify the contour tree with two operations: leaf pruning and vertex reduction. Leaf pruning removes a leaf of the tree, reducing the complexity of the tree, as shown in Figure 3, where vertex 80 is pruned from the tree on the left to produce the tree in the middle. Vertex reduction chooses a vertex with one neighbor above and one below, and deletes the vertex without changing the essential structure of the contour tree. This is also illustrated in Figure 3, where vertex 50 has been removed from the tree in the middle to produce the tree on the right. Since vertex reductions do not change the essential structure of the contour tree, we prefer them to leaf prunes. Also, pruning the only up- or down- edge at a saddle is prohibited to preserve the edge for a later vertex reduction. It is clear that these operations can simplify the tree to any desired size. We can also think of these operations as having well-defined effects on the underlying scalar field: pruning a leaf corresponds to levelling off a maximum or minimum, while vertex reduction requires no changes. As an example, in Figure 3 we show the result of leaf-pruning vertex 80 and edge 80 - 50 from the tree. Since 80 - 50 represents the left-hand maximum, pruning it flattens out the maximum, as shown in the middle terrain. Similarly, the right-hand image shows the results of reducing vertex 50 after the leaf prune. The edges incident to vertex 50 in the tree correspond to the regions above and below the contour through vertex 50. Removing vertex 50 merely combines these two regions into one. The fact that simplification operations can be interpreted as modifying the scalar field suggests that one way to assess the cost of an operation is to measure geometric properties of the change. We show how this can be done efficiently in Section 4. 3.3 Simplification Algorithm To simplify the contour tree, we apply the following rules: 1. Always perform vertex reduction where possible. 2. Always choose the least important leaf to prune. 3. Never prune the last up- or down- leaf at an interior vertex. We implement this with a priority queue to keep track of the leaves of the tree with their associated pruning cost. We assume that for each edge e of the tree, we know two costs: up (e) for pruning the edge from the bottom up: i.e. collapsing the edge to its upper vertex, and down (e) for the cost of pruning the edge from the 499 0 90 0 90 0 90 50 0 90 50 0 90 50 80 0 90 50 80 Leaf 80 is pruned Vertex 50 is reduced Figure 3: Leaf Pruning Levels Extrema; Vertex Reduction Leaves Scalar Field Unchanged top downwards. We add each leaf to the priority queue, with priority of up (e) for a lower leaf and down(e) for an upper leaf. We then repeatedly remove the lowest cost leaf edge from the priority queue and prune it. If this pruning causes a vertex to become reducible, we do so immediately. When a vertex is reduced, two edges e 1 and e 2 are merged into a simplified edge d. The cost of pruning d is based on the costs of the two reduced edges. Since up (d) is the cost of pruning d upwards , we set it to up (e 1 ), the cost of pruning the upper edge upwards . Similarly, we set down (d) to down(e 2 ), the cost of pruning the lower edge downwards. If d is a leaf edge, we add it to the priority queue. To simplify queue handling, we mark the reduced edges for lazy deletion. When a marked edge reaches the front of the priority queue, we discard it immediately. Similarly, when the edge removed from the queue is the last up- or down- edge at its interior vertex, we discard it, preserving it for a later vertex reduction. A few observations on this algorithm: First, any desired level of simplification of the tree can be achieved in a number of queue operations linear in t, the size of the original tree. Since at least half the nodes are leaves, this bound is tight. And if the contour tree is stored as nodes with circular linked lists of upwards and downwards edges, every operation except (de)queueing takes constant time. As a result, the asymptotic cost of this algorithm is dominated by the O (t log(t)) cost of maintaining the priority queue. Second, the simplified contour tree can still be used to extract isosurface contours. Vertex reductions build monotone paths corresponding to the simplified edges, while leaf prunes discard entire monotone paths. Thus, any edge in a simplified contour tree corresponds to a monotone path through the original contour tree. To generate the contour at a given isovalue on a simplified edge, we perform a binary search along the contour tree edges that make up the monotone path for that simplified edge. This search identifies the unique contour tree edge that spans the desired isovalue, and we use the path seed associated with that edge to generate the contour. Third, we extract contours from seeds as before. Instead of simplifying individual contours, we reduce the set of contours that can be extracted. Surface simplification of contours is a separate task. Finally, up (e) and down(e) actually need not be set except at leaves of the tree. As a leaf is pruned and vertex reduced, new values can be computed using information from the old nodes and edges. It is not hard to show by induction that any desired level of simplification of the tree can be achieved. And, since leaf pruning and vertex reduction are the only two operations, the net result can also be a meaningful simplification of the underlying scalar field, assuming that a reasonable geometric measure is used to guide the simplification . We therefore next discuss geometric measures. Local Geometric Measures [Bajaj et al. 1997] compute global geometric properties, and display them alongside the contour tree in the contour spectrum. [Pascucci and Cole-McLaughlin 2002] propagate topological indices called the Betti numbers along branches of the contour tree, based on previous work by [Pascucci 2001]. We bring these two ideas together to compute local geometric measures for individual contours. In 2D scalar fields, the geometric properties we could compute include the following contour properties: line length (perimeter), cross-sectional area (area of region enclosed by the contour), volume (of the region enclosed), and surface area (of the function over the region). In 3D scalar fields, there are analogous properties that include isosurface area, cross-sectional volume (the volume of the region enclosed by the isosurface), and hypervolume (the integral of the scalar field over the enclosed volume). Figure 4: Contours Sweeping Past a Saddle Point Consider a plane sweeping through the field in Figure 2 from high to low isovalues. At any isovalue h, the plane divides the field into regions above and below the plane. As the isovalue decreases, the region above the plane grows, sweeping past the vertices of the mesh one at a time. Geometric properties of this region can be written as functions of the isovalue h. Such properties are decomposable over the cells of the input data for each cell we compute a piecewise polynomial function, and sum them to obtain a piecewise polynomial function for the entire region. [Bajaj et al. 1997] compute these functions by sweeping through the isovalues, altering the function as each vertex is passed. Figure 4 illustrates this process, showing the contours immediately above and below a vertex s. As the plane sweeps past s, the function is unchanged in cells outside the neighbourhood of s, but changes inside the neighbourhood of s. This sweep computes global geometric properties for the region above the sweep plane. Reversing the direction of the sweep computes global geometric properties for the region below 500 the sweep plane. In Figure 2, the region above the sweep plane at 2000m consists of two connected components, one defined by contours c 1 and c 2 , the other by c 3 . To compute properties for these components, we sweep along an edge of the contour tree, representing a single contour sweeping through the data. This lets us compute functions for the central maximum at B. For the crater rim defined by contours c 1 and c 2 , we use inclusion/exclusion. We sweep one contour at a time, computing properties for the region inside the contour, including regions above and below the isovalue of the contour. The area of the crater rim can then be computed by subtracting the area inside contour c 2 from the area inside contour c 1 . We define local geometric measures to be geometric properties of regions bounded by a contour. We compute these measures in a manner similar to the global sweep of [Bajaj et al. 1997], but by sweeping contours along contour tree edges. 4.1 Local Geometric Measures To define local geometric measures attached to contour tree edges, we must be careful with terminology. Above and below do not apply to the region inside c 1 in Figure 2, since part of the region is above the contour and part is below. Nor do inside and outside, which lose their meaning for contours that intersect the boundary. We therefore define upstart and downstart regions of a contour. An upstart region is a region reachable from the contour by paths that initially ascend from the contour and never return to it. For contour c 1 , there is one upstart region (inside) and one downstart region (outside). At saddles such as D, there may be several upstart regions. Since each such region corresponds to an edge in the contour tree, we refer, for example, to the upstart region at D for arc CD. We now define upstart and downstart functions: functions computed for upstart or downstart regions. Note that the upstart and downstart functions do not have to be the same. For example, the length of a contour line is independent of sweep direction, so the upstart and downstart functions for contour length in 2D are identical . But the area enclosed by a contour depends on sweep direction, so the upstart and downstart functions will be different. Since upstart and downstart functions describe geometric properties local to a contour, we refer to them collectively as local geometric measures. These measures are piecewise polynomial since they are piecewise polynomial in each cell. Because we need to track connectivity for inclusion/exclusion, they are not strictly decomposable . Stated another way, in order to make them decomposable, we need to know the connectivity during the local sweep. We are fortunate that the contour tree encodes this connectivity. For regular data, we approximate region size with vertex count as in [Takahashi et al. 2004b]. For the integral of f over region R, we sum the sample values to get x R f (x): the correct integral is the limit of this sum as sample spacing approaches zero. When we prune a leaf to a saddle at height h, the integral over the region flattened is x R ( f (x) - h) = ( x R f (x)) - Ah where A is the area of region R. In three dimensions, vertex counting measures volume, and summing the samples gives hypervolume. This geometric measure is quite effective on the data sets we have tested in Section 6. 4.2 Combining Local Geometric Measures To compute local geometric measures, we must be able to combine upstart functions as we sweep a set of contours past a vertex. In Figure 4, we must combine the upstart functions for contours c 1 , c 2 and c 3 before sweeping past s. We must then update the combined upstart function as we sweep past the vertex. After sweeping past s, we know the combined upstart function d for contours d 1 , d 2 and d 3 . We remove the upstart functions for d 1 and d 2 from d to obtain the upstart function for d 3 . We assume that we have recursively computed the upstart functions for d 1 and d 2 by computing the downstart functions and then inverting them. Let us illustrate inversion, combination and removal for two local geometric measures in two dimensions. Contour Length: Contour length is independent of sweep direction , so these operations are simple: Inversion is the identity operation , combination sums the lengths of the individual contours, and contours are removed by subtracting their lengths. Area: Area depends on sweep direction, so inversion subtracts the function from the area of the entire field. Combining upstart functions at a saddle depends on whether the corresponding edges ascend or descend from the saddle. For ascending edges the upstart regions are disjoint, and the upstart functions are summed. For descending edges the upstart regions overlap, and the upstart functions are combined by inverting to downstart functions, summing, and re-inverting. Removing upstart functions reverses combination. Consider Figure 4 once more. The upstart region of d 1 contains s, as well as contours c 1 , c 2 and c 3 . Similarly, the upstart regions of d 2 and d 3 contain s and contours c 1 , c 2 and c 3 . However, the downstart regions of d 1 , d 2 and d 3 are disjoint, and can be summed, then inverted to obtain the combination of the upstart regions. In general, measures of contour size are independent of sweep direction and their computation follows the pattern of 2D contour length. Such measures include surface area in three dimensions, and hypersurface volume in four dimensions. Measures of region size depend on sweep direction and their computation follows the pattern of 2D cross-sectional area. Such measures include surface area and volume in two dimensions, and isosurface cross-sectional volume and hypervolume in three dimensions. Input : Fully Augmented Contour Tree C A local geometric measure f with operations Combine ( f 1 ,..., f m ) local geometric measures Update ( f ,v) that updates f for sweep past v Remove ( f , f 1 ,..., f m ) local geometric measures Invert () from down(e) to up(e) or vice versa Output : down (e) and up(e) for each edge e in C Make a copy C of C 1 for each vertex v do 2 If v is a leaf of C, enqueue v 3 while NumberOfArcs (C ) &gt; 0 do 4 Dequeue v and retrieve edge e = (u,v) from C 5 Without loss of generality, assume e ascends from v 6 Let d 1 , . . . , d k be downward arcs at v in C 7 Let upBelow = Combine(down(d 1 ),...,down(d k ) 8 Let upAbove = Update(upBelow,v) 9 Let e 1 , . . . , e m be upwards arcs at v in C, with e 1 = e 10 Let f i = Invert(down(e i )) for i = 2,...,m 11 Let up (e) = Remove(upAbove, f 2 ,..., f m ) 12 Let down (e) = Invert(up(e)) 13 Delete e from C 14 If u is now a leaf of C , enqueue u 15 Algorithm 1: Computing Local Geometric Measures 501 (a) Reduced by Height (Persistence) (b) Reduced by Volume (Vertex Count) (c) Reduced by Hypervolume (Riemann Sum) Figure 5: Comparison of Simplification Using Three Local Geometric Measures. In each case, the UNC Head data set has been simplified to 92 edges using the specified measure. Each trees were laid out using the dot tool, with no manual adjustment. 4.3 Computing Local Geometric Measures Algorithm 1 shows how to compute edge priorities up (e) and down (e) for a given local geometric measure. This algorithm relies on Combine (), Update(), Invert(), and Remove() having been suitably defined, and can be integrated into the merge phase of the contour tree algorithm in [Carr et al. 2003]. The algorithm builds a queue of leaf edges in Step 2, then works inwards, pruning edges as it goes. At each vertex, including regular points, the computation described in Section 4.2 is performed, and the edge is deleted from the tree. In this way, an edge is processed only when one of its vertices is reduced to a leaf: i.e. when all other edges at that vertex have already been processed. Unlike simplification, Algorithm 1 requires the fully augmented contour tree, which is obtained by adding every vertex in the input mesh to the contour tree. This makes the algorithm linear in the input size n rather than the tree size t: it cannot be used with the algorithms of [Pascucci and Cole-McLaughlin 2002] and [Chiang et al. 2002], which reduce running time by ignoring regular points. 4.4 Comparison of Local Geometric Measures In Figure 5, we show the results of simplifying the UNC Head data set with three different geometric measures: height (persistence), volume, and hypervolume. In each case, the contour tree has been reduced to 92 edges and laid out using dot with no manual intervention . In the left-hand image, height (persistence) is used as the geometric measure. All of the edges shown are tall as a result, but on inspection , many of these edges are caused by high-intensity voxels in the skull or in blood vessels. Most of the corresponding objects are quite small, while genuine objects of interest such as the eyes, ventricular cavities and nasal cavity have already been suppressed, because they are defined by limited ranges of voxel intensity. Also, on the corresponding simplification curve, we observe that there are a relatively large number of objects with large intensity ranges: again, on further inspection, these tended to be fragments of larger objects, particularly the skull. In comparison, the middle image shows the results of using volume (i.e. vertex count) as the geometric measure. Not only does this focus attention on a few objects of relatively large spatial extent, but the simplification curve shows a much more rapid drop-off, implying that there are fewer objects of large volume than there are of large height. Objects such as the eyeballs are represented, as they have relatively large regions despite having small height. However, we note that there are a large number of small-height edges at the bottom of the contour tree. These edges turn out to be caused by noise and artifacts outside the skull in the original CT scan, in which large regions are either slightly higher or lower in isovalue than the surrounding regions. Finally, the right-hand image shows the results of using hypervolume (the sum of sample values, as discussed above). In this case, we see a very rapid dropoff of importance in the simplification curve, with only 100 or so regions having significance. We note that this measure preserves small-height features such as the eyeballs, while eliminating most of the apparent noise edges at the bottom of the tree, although at the expense of representing more skull fragments than the volume measure. In general we have found that this measure is better for data exploration than either height or volume, since it balances representation of tall objects with representation of large objects. We do not claim that this measure is universally ideal: the choice of simplification measure should be driven by domain-dependent information. However, no matter what measure is chosen, the basic mechanism of simplification remains. Implementation We have combined simplification with the flexible isosurface interface of [Carr and Snoeyink 2003], which uses the contour tree as a visual index to contours. The interface window, shown in Figures 1, 6, and 7, is divided into data, contour tree, and simplification curve panels. The data panel displays the set of contours marked in the contour tree panel. Contours can be selected in either panel, then deleted, isolated, or have their isovalue adjusted. The simplification curve panel shows a log-log plot of contour tree size against "feature size": the highest cost of any edge pruned to reach the given level of simplification. Selecting a point on this curve determines the detail shown in the contour tree panel. For efficiency, we compute contour trees for the surfaces given by the Marching Cubes cases of [Montani et al. 1994] instead of a sim-502 plicial or trilinear mesh, because these surfaces generate roughly 60% fewer triangles than even a minimal simplicial subdivision of the voxels, with none of the directional biases identified by [Carr et al. 2001], and because they are significantly simpler to compute than the trilinear interpolant used by [Pascucci and Cole-McLaughlin 2002]. There is a loss of accuracy, but since our simplification discards small-scale details of the topology anyway, little would be gained from more complex interpolants. Finally, as in [Carr et al. 2003; Pascucci and Cole-McLaughlin 2002; Chiang et al. 2002], we use simulation of simplicity [Edelsbrunner and Mucke 1990] to guarantee uniqueness of isovalues, then collapse zero-height edges in the tree. Implementation details can be found in [Carr 2004]. Results and Discussion We used a variety of data sets to test these methods, including results from numerical simulations (Nucleon, Silicium, Fuel, Neghip, Hydrogen), analytical methods (ML, Shockwave), CT-scans (Lobster , Engine, Statue, Teapot, Bonsai), and X-rays (Aneurysm, Foot, Skull). Table 1 lists the size of each data set, the size of the unsimplified contour tree, the time for constructing the unsimplified contour tree, and the simplification time. Times were obtained using a 3 GHz Pentium 4 with 2 GB RAM, and the hypervolume measure. Data Set Data Tree Size Size CT (s) ST (s) Nucleon 41 41 41 49 0.28 0.01 ML 41 41 41 695 0.25 0.01 Silicium 98 34 34 225 0.41 0.01 Fuel 64 64 64 129 0.72 0.01 Neghip 64 64 64 248 0.90 0.01 Shockwave 64 64512 31 5.07 0.01 Hydrogen 128 128128 8 5.60 0.01 Lobster 301 324 56 77,349 19.22 0.10 Engine 256 256128 134,642 31.51 0.18 Statue 341 341 93 120,668 32.20 0.15 Teapot 256 256178 20,777 33.14 0.02 Aneurysm 256 256256 36,667 41.83 0.04 Bonsai 256 256256 82,876 49.71 0.11 Foot 256 256256 508,854 67.20 0.74 Skull 256 256256 931,348 109.73 1.47 CT Head 106 256256 92,434 21.30 0.12 UNC Head 109 256256 1,573,373 91.23 2.48 Tooth 161 256256 338,300 39.65 0.48 Rat 240 256256 2,943,748 233.33 4.97 Table 1: Data sets, unsimplified contour tree sizes, and contour tree construction time (CT) and simplification time (ST) in seconds. The size of the contour tree is proportional to the number of local extrema in the input data. For analytic and simulated data sets, such as the ones shown in the upper half of Table 1, this is much smaller than the input size. For noisy experimentally acquired data, such as the ones shown in the lower half of Table 1, the size of the contour tree is roughly proportional to the input size. The time required to simplify the contour tree using local geometric measures is typi-cally less than one percent of the time of constructing the original contour tree, plus the additional cost of pre-computing these measures during contour tree construction. 6.1 Examples of Data Exploration Figure 1 shows the result of exploring of the UNC Head data set using simplified contour trees. An appropriate level of simplification was chosen on the simplification curve and individual contours explored until the image shown was produced. Surfaces identifiable as part of the skull were not chosen because they occluded the view of internal organs, although two contours for the ventricular system were chosen despite being occluded by the brain surrounding them. The flexible isosurface interface is particularly useful in this context because it lets one manipulate a single contour at a time, as shown in the video submitted with this paper. embryo gut? lungs eyes brain windpipe? shoulder blades breastbone Figure 6: A Pregnant Rat MRI (240 256256). Despite low quality data, simplifying the contour tree from 2,943,748 to 125 edges allows identification of several anatomical features. spinal column spinal cord ventricles spinal cord spinal column ventricles Figure 7: CT of a Skull (256 256 106). Simplification of the contour tree from 92,434 to 20 edges isolates the ventricular cavity, spinal cord and spinal column. Similarly, Figure 6 shows the result of a similar exploration of a 240 256 256, low-quality MRI scan of a rat from the Whole Frog Project at http://www-itg.lbl.gov/ITG.hm.pg.docs/ Whole.Frog/Whole.Frog.html. Again, simplification reduces the contour tree to a useful size. Figure 7 shows a spinal column, spinal cord and ventricular cavity identified in a 256 256 106 CT data set from the University of Erlangen-Nuremberg. Other examples may be seen on the accompanying video. Each of these images took less than 10 minutes to produce after all pre-processing, using the dot tool from the graphviz package (http://www.research.att.com/sw/tools/graphviz/) to lay out the contour tree: we generally then made a few adjustments to the node positions for clarity. Although dot produces reasonable layouts for trees with 100 200 nodes, it is slow, sometimes taking several minutes, and the layout computed usually becomes unsatisfactory as edges are added or subtracted from the tree. 503 Note that in none of these cases was any special constant embedded in the code the result is purely a function of the topology of the isosurfaces of the input data. Conclusions and Future Work We have presented a novel algorithm for the simplification of contour trees based on local geometric measures. The algorithm is online , meaning that simplifications can be done and undone at any time. This addresses the scalability problems of the contour tree in exploratory visualization of 3D scalar fields. The simplification can also be reflected back onto the input data to produce an on-line simplified scalar field. The algorithm is driven by local geometric measures such as area and volume, which make the simplifications meaningful. Moreover, the simplifications can be tailored to a particular application or data set. We intend to explore several future directions. We could compute a multi-dimensional feature vector of local geometric measures, and allow user-directed simplification of the contour tree, with different measures being applied in different regions of the function. The simplified contour tree also provides a data structure for queries. With local feature vectors one could efficiently answer queries such as "Find all contours with volume of at least 10 units and an approximate surface-area-to-volume ratio of 5." If information about spatial extents (e.g., bounding boxes) is computed, then spatial constraints can also be included. Inverse problems could also be posed given examples of a feature (e.g., a tumor), what should the query constraints be to find such features? Some interface issues still need resolution, such as finding a fast contour tree layout that is clear over a wide range of levels of simplification but which also respects the convention that the y-position depends on the isovalue. We would also like to annotate contours using the flexible isosurface interface, rather than after the fact as we have done in Figure 1 and Figures 6 7, and to enable local simplification of the contour tree rather than the single-parameter simplification presented here. Isosurfaces are not the only way of visualizing volumetric data. Other methods include boundary propagation using level set methods or T-snakes. We believe that simplified contour trees can provide seeds for these methods, either automatically or through user interaction. We are adapting the flexible isosurface interface to generate transfer functions for volume rendering. These transfer functions would add spatial locality to volume rendering, based on the regions corresponding to edges of the simplified contour tree. Another possible direction is to develop more local geometric measure for multilinear interpolants. Lastly, the algorithms we describe work in arbitrary dimensions, but special consideration should be given to simplification of contour trees for time-varying data. Acknowledgements Acknowledgements are due to the National Science and Engineering Research Council of Canada (NSERC) for support in the form of post-graduate fellowships and research grants, and to the U.S. National Science Foundation (NSF) and the Institute for Robotics and Intelligent Systems (IRIS) for research grants. Acknowledgements are also due to those who made volumetric data available at volvis.org and other sites. References B AJAJ , C. L., P ASCUCCI , V., AND S CHIKORE , D. R. 1997. The Contour Spectrum. In Proceedings of IEEE Visualization 1997, 167173. B ENTLEY , J. L. 1979. Decomposable searching problems. Inform. Process. Lett. 8, 244251. B REMER , P.-T., E DELSBRUNNER , H., H AMANN , B., AND P ASCUCCI , V. 2003. A Multi-resolution Data Structure for Two-dimensional Morse-Smale Functions. In Proceedings of IEEE Visualization 2003, 139146. B RODLIE , K., AND W OOD , J. 2001. Recent advances in volume visualization. Computer Graphics Forum 20, 2 (June), 125148. C ARR , H., AND S NOEYINK , J. 2003. Path Seeds and Flexible Isosurfaces: Using Topology for Exploratory Visualization. In Proceedings of Eurographics Visualization Symposium 2003, 4958, 285. C ARR , H., M OLLER , T., AND S NOEYINK , J. 2001. Simplicial Subdivisions and Sampling Artifacts. In Proceedings of IEEE Visualization 2001, 99106. C ARR , H., S NOEYINK , J., AND A XEN , U. 2003. Computing Contour Trees in All Dimensions. Computational Geometry: Theory and Applications 24, 2, 7594. C ARR , H. 2004. Topological Manipulation of Isosurfaces. PhD thesis, University of British Columbia, Vancouver, BC, Canada. C HIANG , Y.-J., AND L U , X. 2003. Progressive Simplification of Tetrahedral Meshes Preserving All Isosurface Topologies. Computer Graphics Forum 22, 3, to appear. C HIANG , Y.-J., L ENZ , T., L U , X., AND R OTE , G. 2002. Simple and Output-Sensitive Construction of Contour Trees Using Monotone Paths. Tech. Rep. ECG-TR -244300-01, Institut f ur Informatik, Freie Universtat Berlin. E DELSBRUNNER , H., AND M UCKE , E. P. 1990. Simulation of Simplicity: A technique to cope with degenerate cases in geometric algorithms. ACM Transactions on Graphics 9, 1, 66104. E DELSBRUNNER , H., L ETSCHER , D., AND Z OMORODIAN , A. 2002. Topological persistence and simplification. Discrete Comput. Geom. 28, 511533. E DELSBRUNNER , H., H ARER , J., AND Z OMORODIAN , A. 2003. Hierarchical Morse-Smale complexes for piecewise linear 2-manifolds. Discrete Comput. Geom. 30, 87107. H ILAGA , M., S HINAGAWA , Y., K OHMURA , T., AND K UNII , T. L. 2001. Topology matching for fully automatic similarity estimation of 3d shapes. In SIGGRAPH 2001, 203212. K ETTNER , L., R OSSIGNAC , J., AND S NOEYINK , J. 2001. The Safari Interface for Visualizing Time-Dependent Volume Data Using Iso-surfaces and Contour Spectra. Computational Geometry: Theory and Applications 25, 1-2, 97116. L ORENSON , W. E., AND C LINE , H. E. 1987. Marching Cubes: A High Resolution 3D Surface Construction Algorithm. Computer Graphics 21, 4, 163169. M ATSUMOTO , Y. 2002. An Introduction to Morse Theory. AMS. M ILNOR , J. 1963. Morse Theory. Princeton University Press, Princeton, NJ. M ONTANI , C., S CATENI , R., AND S COPIGNO , R. 1994. A modified look-up table for implicit disambiguation of Marching Cubes. Visual Computer 10, 353355. P ASCUCCI , V., AND C OLE -M C L AUGHLIN , K. 2002. Efficient Computation of the Topology of Level Sets. In Proceedings of IEEE Visualization 2002, 187194. P ASCUCCI , V. 2001. On the Topology of the Level Sets of a Scalar Field. In Abstracts of the 13th Canadian Conference on Computational Geometry, 141144. R EEB , G. 1946. Sur les points singuliers d'une forme de Pfaff compl`etement integrable ou d'une fonction numerique. Comptes Rendus de l'Acad `emie des Sciences de Paris 222, 847849. T AKAHASHI , S., F UJISHIRO , I., AND T AKESHIMA , Y. 2004. Topological volume skeletonization and its application to transfer function design. Graphical Models 66, 1, 2449. T AKAHASHI , S., N IELSON , G. M., T AKESHIMA , Y., AND F UJISHIRO , I. 2004. Topological Volume Skeletonization Using Adaptive Tetrahedralization. In Geometric Modelling and Processing 2004. T ARASOV , S. P., AND V YALYI , M. N. 1998. Construction of Contour Trees in 3D in O (nlogn) steps. In Proceedings of the 14th ACM Symposium on Computational Geometry, 6875. VAN K REVELD , M., VAN O OSTRUM , R., B AJAJ , C. L., P ASCUCCI , V., AND S CHIKORE , D. R. 1997. Contour Trees and Small Seed Sets for Isosurface Traver-sal . In Proceedings of the 13th ACM Symposium on Computational Geometry, 212220. 504

Isosurfaces;topological simplification;contour trees

18

A Resilient Packt-Forwarding Scheme against Maliciously Packet-Dropping Nodes in Sensor Networks

This paper focuses on defending against compromised nodes' dropping of legitimate reports and investigates the misbehavior of a maliciously packet-dropping node in sensor networks . We present a resilient packet-forwarding scheme using Neighbor Watch System (NWS), specifically designed for hop-by-hop reliable delivery in face of malicious nodes that drop relaying packets, as well as faulty nodes that fail to relay packets. Unlike previous work with multipath data forwarding, our scheme basically employs single-path data forwarding, which consumes less power than multipath schemes. As the packet is forwarded along the single-path toward the base station, our scheme, however, converts into multipath data forwarding at the location where NWS detects relaying nodes' misbehavior. Simulation experiments show that, with the help of NWS, our forwarding scheme achieves a high success ratio in face of a large number of packet-dropping nodes, and effectively adjusts its forwarding style, depending on the number of packet-dropping nodes en-route to the base station.

INTRODUCTION Wireless sensor networks consist of hundreds or even thousands of small devices each with sensing, processing, and Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SASN'06, October 30, 2006, Alexandria, Virginia, USA. Copyright 2006 ACM 1-59593-554-1/06/0010 ... $ 5.00. communicating capabilities to monitor the real-world environment . They are envisioned to play an important role in a wide variety of areas ranging from critical military-surveillance applications to forest fire monitoring and the building security monitoring in the near future. In such a network, a large number of sensor nodes are distributed to monitor a vast field where the operational conditions are harsh or even hostile. To operate in such environments, security is an important aspect for sensor networks and security mechanisms should be provided against various attacks such as node capture, physical tampering, eavesdropping, denial of service, etc [23, 33, 38]. Previous research efforts against outsider attacks in key-management schemes [4, 13, 32] and secure node-to-node communication mechanisms [24, 32] in sensor networks are well-defined. Those security protections, however, break down when even a single legitimate node is compromised. It turns out to be relatively easy to compromise a legitimate node [14], which is to extract all the security information from the captured node and to make malicious code running for the attacker's purpose. Even a small number of compromised nodes can pose severe security threats on the entire part of the network, launching several attacks such as dropping legitimate reports , injecting bogus sensing reports, advertising inconsistent routing information, eavesdropping in-network communication using exposed keys, etc. Such disruption by the insider attacks can be devastating unless proper security countermeasures against each type of attacks are provided. In reality, detecting all of the compromised nodes in the network is not always possible, so we should pursue graceful degradation [35], with a small number of compromised nodes. The fundamental principle for defense against the insider attacks is to restrict the security impact of a node compromise as close to the vicinity of the compromised node as possible. When the attacker compromises a legitimate node, it may first try to replicate the captured node indefinitely with the same ID and spread them over the network. Against such attacks, a distributed detection mechanism (based on emergent properties [11]) has been proposed by Parno et al. [31]. In addition, Newsome et al. [30] have presented the techniques that prevent the adversary from arbitrarily creating new IDs for nodes. Using cryptographic information obtained from a captured node, attackers can establish pairwise keys with any legitimate nodes in order to eavesdrop communication any-59 where in the network. Localized key-establishment scheme by Zhu et al. [46] is a good solution against such an insider attack. Since the scheme does not allow a cloned node (by inside-attackers) to establish pairwise keys with any legitimate nodes except the neighbors of the compromised nodes, the cryptographic keys extracted from the compromised node are of no use for attackers. Compromised nodes can also inject false sensing reports to the network (i.e. report fabrication attacks [39]), which causes false alarms at the base station or the aggregation result to far deviate from the true measurement. Proposed en-route filtering mechanisms [8, 39, 41, 44, 47] that detect and drop such false reports effectively limit the impact of this type of attacks. Also, proposed secure aggregation protocols [34, 40] have addressed the problem of false data injection, and they ensure that the aggregated result is a good approximation to the true value in the presence of a small number of compromised nodes. Advertising inconsistent routing information by compromised nodes can disrupt the whole network topology. Hu et al. [19, 20] have proposed SEAD, a secure ad-hoc network routing protocol that uses efficient one-way hash functions to prevent any inside attackers from injecting inconsistent route updates. A few secure routing protocols [6, 27] in sensor networks have been proposed to detect and exclude the compromised nodes injecting inconsistent route updates. Compromised nodes also can silently drop legitimate reports (i.e. selective forwarding attacks [23]), instead of forwarding them to the next-hop toward the base station. Since data reports are delivered over multihop wireless paths to the base station, even a small number of strategically-placed packet-dropping nodes can deteriorate the network throughput significantly. In order to bypass such nodes, most work on secure routing and reliable delivery in sensor networks relies on multipath forwarding scheme [5, 6, 7, 10], or interleaved-mesh forwarding scheme [26, 29, 39, 42]. Among the insider attacks described above, this paper focuses on defense against compromised nodes' dropping of legitimate reports and we present a resilient packet-forwarding scheme using Neighbor Watch System (NWS) against maliciously packet-dropping nodes in sensor networks. We investigate the misbehavior of a maliciously packet-dropping node and show that an acknowledgement (ACK) that its packets were correctly received at the next-hop node does not guarantee reliable delivery from the security perspective. NWS is specifically designed for hop-by-hop reliable delivery in face of malicious nodes that drop relaying packets, as well as faulty nodes that fail to relay packets. Unlike previous work [10, 29, 42] with multipath data forwarding, our scheme basically employs single-path data forwarding, which consumes less power than multipath schemes. As the packet is forwarded along the single-path toward the base station, our scheme, however, converts into multipath data forwarding at the location where NWS detects relaying nodes' misbehavior . NWS exploits the dense deployment of large-scale static sensor networks and the broadcast nature of communication pattern to overhear neighbors' communication for free. The contribution of this paper is two-fold. First, we investigate the misbehavior of a maliciously packet-dropping node and propose a resilient packet-forwarding scheme, which basically employs single-path data forwarding, in face of such nodes, as well as faulty nodes. Second, our scheme can work with any existing routing protocols. Since it is designed not for securing specific protocols but for universal protocols, it can be applied to any existing routing protocols as a security complement. The rest of paper is organized as follows. Background is given in Section 2. We present our resilient packet-forwarding scheme in Section 3. An evaluation of the scheme is given and discussed in Section 4. We present conclusions and future work in Section 5. BACKGROUND Sensor networks typically comprise one or multiple base stations and hundreds or thousands of inexpensive, small, static, and resource-constrained nodes scattered over a wide area. An inexpensive sensor node cannot afford tamper-resistant packaging. We assume that a large number of sensor nodes are deployed in high density over a vast field, such that the expected degree of a node is high; each sensor has multiple neighbors within its communication range. Sensing data or aggregated data are sent along the multihop route to the base station. We assume that each sensor node has a constant transmission range, and communication links are bidirectional. Our sensor network model employs a key-establishment scheme that extends the one in LEAP [46] where the impact of a node compromise is localized in the immediate neighborhood of the compromised node, and our scheme is based on it. To evolve from LEAP, we will describe it briefly in Section 2.4. 2.2 Threat Model The attacks launched from outsiders hardly cause much damage to the network, since the rouge node, which does not possesses the legitimate credentials (e.g. the predistributed key ring from the key pool [13]), fails to participate in the network. On the other hand, there may be multiple attacks from insiders (e.g. dropping legitimate reports, injecting false sensing reports, advertising inconsistent route information , and eavesdropping in-network communication using exposed keys, etc), and the combination of such attacks can lead to disruption of the whole network. Thus, proper security countermeasures (specifically designed to protect against each type of the attacks) should be provided. Among them, in this paper, we focus on defending against compromised nodes' dropping of legitimate reports; Other attacks mentioned above are effectively dealt with by several proposed schemes as described in the previous section. We consider a packet-dropping node as not merely a faulty node, but also an arbitrarily malicious node. Some previous work [3, 29, 36] on reliable delivery uses an acknowledgement (ACK) that its packets were correctly received at the next-hop node, in order to find out unreliable links. However , in the presence of maliciously packet-dropping nodes, simply receiving ACK from a next-hop node does not guarantee that the packet will be really forwarded by the next-hop node. For example, node u forwards a packet to compromised node v, and node u waits for ACK from node v. Node v sends back ACK to node u, and then node v silently drops the packet. This simple example shows that receiving ACK is not enough for reliable delivery in face of maliciously packet-dropping nodes. 60 For more reliability, we should check whether the next-hop node really forwards the relaying packet to its proper next-hop node. Fortunately, due to the broadcast nature of communication pattern in sensor networks, we can overhear neighbors' communication for free (for now per-link encryption is ignored). After forwarding a packet to next-hop node v and buffering recently-sent packets, by listening in on node v's traffic, we can tell whether node v really transmits the packet. Watchdog [28] mechanism (extension to DSR [22]), implicit ACK in M 2 RC [29], and local monitoring in DICAS [25] detect misbehaving nodes in this way. However, this kind of simple overhearing schemes does not guarantee reliable delivery, either. With arbitrarily malicious nodes, we should be assured that the node, to which the next-hop node forwards the relaying packet, is really a neighbor of the next-hop node. For example, node u forwards a packet to compromised node v, and node u listens in on node v's traffic to compare each overheard packet with the packet in the buffer. Node v transmits the relaying packet whose intended next-hop id marked with any id in the network such as x that is not a neighbor of v. Then node u overhears this packet from node v, and considers it forwarded correctly despite the fact that none actually receives the packet. The packet is eventually dropped without being detected. We refer to this attack as blind letter attack. We consider packet-dropping attacks to be addressed in this paper as ones ranging from the naive case (e.g. a faulty node) to the most malicious one (e.g. a node launching blind letter attack). We focus on developing a solution to such attacks. 2.3 Notation We use the following notation throughout the paper: u, v are principals, such as communicating nodes. R u is a random number generated by u. f K is a family of pseudo-random function [12]. MAC(K, M 1 |M 2 ) denotes the message authentication code (MAC) of message - concatenation of M 1 and M 2 , with MAC key K. 2.4 Key-Establishment Scheme in LEAP LEAP supports the establishment of four types of keys for each sensor node - an individual key shared with the base station, a pairwise key shared with its neighbor, a cluster key shared with its surrounding neighbors, and a group key shared by all the nodes in the networks. It assumes that the time interval T est for a newly deployed sensor node to complete the neighbor discovery phase (e.g. tens of seconds) is smaller than the time interval T min that is necessary for the attacker to compromise a legitimate node (i.e. T min &gt; T est ). Some existing work [1, 39] has made similar assumptions, which are believed to be reasonable. The four steps for a newly added node u to establish a pairwise key with each of its neighbors are as follows: 1. Key Pre-distribution. Each node u is loaded with a common initial key K I , and derives its master key K u = f K I (u). 2. Neighbor Discovery. Once deployed, node u sets up a timer to fire after time T min , broadcasts its id, and waits for each neighbor v's ACK. The ACK from v is authenticated using the master key K v of node v. Since node u knows K I , it can derive K v = f K I (v). u - : u, R u . v - u : v, M AC(K v , R u |v). 3. Pairwise Key Establishment. Node u computes its pairwise key with v, K uv , as K uv = f K v (u). Node v also computes K uv in the same way. K uv serves as their pairwise key. 4. Key Erasure. When its timer expires, node u erases K I and all the master keys of its neighbors. Every node, however, keeps its own master key, in order to establish pairwise keys with later-deployed nodes. Once erasing K I , a node will not be able to establish a pairwise key with any other nodes that have also erased K I . Without K I , a cloned node (by an attacker compromising a legitimate node after T min ) fails to establish pairwise keys with any nodes except the neighbors of the compromised node. In such a way, LEAP localizes the security impact of a node compromise. A RESILIENT PACKET-FORWARDING SCHEME USING NEIGHBOR WATCH SYSTEM In this section, we present our resilient packet-forwarding scheme using Neighbor Watch System (NWS). NWS works with the information provided by Neighbor List Verification (NLV) to be described in Section 3.2. 3.1 Neighbor Watch System Our scheme seeks to achieve hop-by-hop reliable delivery in face of maliciously packet-dropping nodes, basically employing single-path forwarding. To the best of our knowledge , proposed works so far rely on multipath forwarding or diffusion-based forwarding, exploiting a large number of nodes in order to deliver a single packet. ACK-based technique is not a proper solution at all as explained in the previous section. With NWS, we can check whether the next-hop node really forwards the relaying packet to the actual neighbor of the next-hop node. The basic idea of our scheme is as follows : 1. Neighbor List Verification. After deployment, during neighbor discovery phase, every node u gets to know of not only its immediate neighbors, but also the neighbors' respective neighbor lists (i.e. u's neighbors' neighbor lists). The lists are verified using Neighbor List Verification to be described in Section 3.2. Every node stores its neighbors' neighbor lists in the neighbor table. 2. Packet Forwarding to Next-hop. If node u has a packet to be relayed, it buffers the packet and forwards the packet (encrypted with cluster key of node u so that neighbors of node u can overhear it) to its next-hop node v. As in LEAP, a cluster key is a key shared by a node and all its neighbors, for passive participation . 61 u v ? w y Figure 1: Neighbor Watch System. Sub-watch nodes w and y, as well as primary-watch node u listen in on v's traffic. 3. Designation of Watch Nodes. Overhearing the packet from node u to node v, among neighbors of node u, the nodes that are also neighbors of node v (in Figure 1, nodes w and y) are designated as sub-watch nodes and store the packet in the buffer. Other nodes (that are not neighbors of node v) discard the packet. Node u itself is a primary-watch node. A primary-watch node knows which nodes are sub-watch nodes, since every node has the knowledge of not only its neighbors but also their respective neighbor lists. 4. Neighbor Watch by Sub-Watch Node. Sub-watch nodes w and y listen in on node v's traffic to compare each overheard packet with the packet in the buffer. To defend against blind letter attack, each of them also checks whether the packet's intended next-hop is a verified neighbor of node v, by looking up the neighbor table. If all correct, the packet in the buffer is removed and the role of the sub-watch node is over. If the packet has remained in the buffer for longer than a certain timeout, sub-watch nodes w and y forward the packet (encrypted with their respective cluster keys) to their respective next-hop nodes other than node v. Then the role of a sub-watch node is over (each of them is now designated as a primary-watch node for the packet it has forwarded). 5. Neighbor Watch by Primary-Watch Node. Primary-watch node u does the same job as sub-watch nodes. The only difference, however, is that it listens in on not only node v's traffic, but also sub-watch nodes w's and y's. If the packet is correctly forwarded on by at least one of them (nodes v, w, or y), primary-watch node u removes the packet in the buffer and the role of the primary-watch node is over. Otherwise, after a certain timeout, primary-watch node u forwards the packet (encrypted with its cluster key) to its next-hop other than node v. As the packet is forwarded on, this procedure (except for Neighbor List Verification) of NWS is performed at each hop so that hop-by-hop reliable delivery can be achieved with mainly depending on single-path forwarding. On the other hand, in the previous approaches [29, 39, 42], when forwarding a packet, a node broadcasts the packet with no designated next-hop, and all neighbors with smaller costs 1 1 The cost at a node is the minimum energy overhead to Base Station u v ? ? Figure 2: An example of our packet-forwarding scheme. Only the nodes that relay the packet are presented. With the help of sub-watch nodes (grey ones), our scheme bypasses two packet-dropping nodes en-route to the base station. or within a specific geographic region continue forwarding the packet anyway. For example, in Figure 1, if nodes v, w, and y have smaller costs than node u in the previous approaches, they all forward 2 the packet from node u. In our scheme, however, sub-watch nodes w and y are just on watch in designated next-hop node v, instead of uncondi-tionally forwarding the packet. If no packet-dropping occurs en-route to the base station, the packet may be forwarded along single-path all the way through. However, a packet-dropping triggers the multipath forwarding for the dropped packet. If the designated next-hop node v in Figure 1 has not forwarded the relaying packet to its certified neighbor by a certain timeout, sub-watch nodes w and y forward the packet to their respective next-hop. At the point, the packet is sent over multiple paths. Since the location where the packet-dropping occurs is likely in an unreliable region, this prompt reaction of the conversion to multipath forwarding augments the robustness in our scheme. The degree of multipath depends on the number of the sub-watch nodes. Figure 2 shows an example of our packet-forwarding scheme, bypassing two packet-dropping nodes en-route to the base station. If a node utilizes a cache [16, 21] for recently-received packets, it can suppress the same copy of previously-received one within a certain timeout , as nodes u and v in Figure 2. Our scheme requires that a relaying packet should be encrypted with a cluster key of a forwarding node, in order that all its neighbors can decrypt and overhear it. In fact, per-link encryption provides better robustness to a node compromise, since a compromised node can decrypt only the packets addressed to it. Thus, there exists a tradeoff between resiliency against packet-dropping and robustness to a node compromise. However, encryption with a cluster key provides an intermediate level of robustness to a node compromise [24] (a compromised node can overhear only its immediate neighborhood), and also supports local broadcast (i.e. resiliency against packet-dropping), so that we can achieve graceful degradation in face of compromised nodes. forward a packet from this node to the base station. 2 It is the broadcast transmission with no designated next-hop , and, if needed, the packet should be encrypted with a cluster key in order for all neighbors to overhear it. 62 To make our scheme work (against blind letter attack), we must address the problem of how a node proves that it really has the claimed neighbors. It is the identical problem of how a node verifies the existence of its neighbors' neighbors. Apparently, a node has the knowledge of its direct neighbors by neighbor discovery and pairwise key establishment phases. However, in the case of two-hop away neighbors, as in Figure 1, malicious node v can inform its neighbor u that it also has neighbor node x (any possible id in the network ) which in fact is not a neighbor of node v. Node u has to believe it, since node x is not a direct neighbor of node u, and only the node v itself knows its actual surrounding neighbors. Then, how do we verify the neighbors' neighbors ? The answer to this critical question is described in the next subsection. 3.2 Neighbor List Verification To verify neighbors' neighbors, we present Neighbor List Verification (NLV) which extends the pairwise key establishment in LEAP. During neighbor discovery in LEAP, two messages are exchanged between neighbors to identify each other. On the other hand, NLV adopts three-way handshaking neighbor discovery, in order to identify not only communicating parties but also their respective neighbors. NLV has two cases of neighbor discovery. One is that neighbor discovery between two nodes that are both still within the initial T min3 (referred as pure nodes). The other is that neighbor discovery between a newly-deployed node within the initial T min and an existing node over the initial T min (referred as an adult node). Neighbor Discovery between Pure Nodes. Neighbor list verification process between pure nodes is quite simple. If a pure node broadcasts its neighbor list before the elapse of its initial T min , we can accept the list as verifiable. Thus, the key point here is to keep track of each other's T min , and to make sure that both broadcast their respective neighbor lists before their respective T min . The following shows the three-way handshaking neighbor discovery between pure node u and v: u - : u, R u . v - u : v, T v , R v M v , M AC(K v , R u |K u |M v ). u - v : u, T u M u , M AC(K uv , R v |M u ). where T v and T u are the amount of time remaining until T min of v and T min of u, respectively. Once deployed, node u sets up a timer to fire after time T min . Then, it broadcasts its id, and waits for each neighbor v's ACK. The ACK from every neighbor v is authenticated using the master key K v of node v. Since node u knows K I 4 , it can derive K v = f K I (v). The ACK from node v contains T v , the amount of time remaining until T min of node v. If T v is a non-zero value, node v claims to be a pure node. K u in MAC proves node v to be a pure node, since pure node v should know K I and derive K u = f K I (u). Node u records T v (T v added 3 T min is the time interval, necessary for the attacker to compromise a legitimate node as in LEAP [46]. 4 Each node u is loaded with a common initial key K I , and derives its master key K u = f K I (u). After time T min , node u erases K I and all the master keys of its neighbors. u v w x t z r q Figure 3: Neighbor Discovery between Pure node x and Adult node u. Grey and white nodes represent adult and pure nodes, respectively. to the current time of node u) in the entry for node v in the neighbor table. Node u computes its pairwise key with v, K uv = f K v (u). 5 Node u also generates M AC(K v , v|u) (which means that v certifies u as an immediate neighbor), and stores it as a certificate. The ACK from node u also contains T u , the amount of time remaining until T min of u. This ACK is authenticated using their pairwise key K uv , which proves node u a pure node and u's identity. Node v then records T u (T u added to the current time of v) in the entry for u in the neighbor table. It also generates M AC(K u , u|v) and stores it as a certificate. Then, the three-way handshaking is done. Every pure node u broadcasts its neighbor list just prior to T min of u. Each receiving neighbor v checks whether the receiving time at v is prior to T u in the neighbor table. If yes, the neighbor list of u is now certified by each neighbor v. Neighbor Discovery between A Pure Node and An Adult node. After most nodes have completed bootstrapping phase, new nodes can be added in the network. Consider Figure 3. The issue here is how adult node u can assure its existing neighbors (v and w) of the existence of its newly-added neighbor x. This is a different situation from the above neighbor list verification case between two pure nodes. Thus, the messages exchanged during the three-way handshaking are somewhat different in this case. The following shows the three-way handshaking neighbor discovery between pure node x and adult node u: x- : x, R x . u- x : u, T u , R u , v, certif icate M AC(K v , v|u), w, certif icate M AC(K w , w|u) M u , M AC(K u , R x |M u ). x- u : x, T x , certif icate M AC(K x , x|u), v, one-time cert. M AC(K v , x|u), w, one-time cert. M AC(K w , x|u) M x , M AC(K xu , R u |M x ). Newly-added node x sets up a timer to fire after time T min . Then, it broadcasts its id, and waits for each neighbor u's 5 Node v also computes K uv in the same way. K uv serves as their pairwise key. 63 ACK. The ACK from every neighbor u is authenticated using the master key K u of node u. Since node x knows K I , it can derive K u = f K I (u). The ACK from node u contains T u , the amount of time remaining until T min of u. If T u is zero, node u is an adult node that may already have multiple neighbors as in Figure 3. Node u reports its certified neighbor list (v and w) to x by including their respective certificates in the ACK. Node x verifies u's neighbor list by examining each certificate, since x can generate any certificate with K I . If all correct, x computes its pairwise key with u, K xu = f K u (x). Node x also generates M AC(K u , u|x) and stores it as a certificate. The ACK from x also contains T x , the amount of time remaining until T min of x. This ACK is authenticated using their pairwise key K xu , which proves node x a pure node and x's identity. Node u then records T x (T x added to the current time of u) in the entry for x in the neighbor table. Since adult node u cannot generate M AC(K x , x|u) by itself, pure node x provides the certificate for u in the ACK. Node x also provides one-time certificates 6 for each of u's certified neighbors (v and w). Then, the three-way handshaking is done. After that, adult node u broadcasts one-time certificates (from newly-discovered pure node x), in order to assure u's existing neighbors (v and w) of the discovery of new neighbor x. The packet containing one-time certificates is as follows: u- : u, x, v, one-time cert. M AC(K v , x|u), w, one-time cert. M AC(K w , x|u), K A u M u , M AC(K c u , M u ). where x is a new neighbor of u, K A u is a local broadcast authentication key in u's one-way key chain, K c u is the cluster key of u. Each receiving neighbor v of u verifies u's new neighbor x by examining the one-time certificate designated for v, M AC(K v , x|u) 6 . If ok, node x is now certified by each neighbor v of u. Then, one-time certificates can be erased, since they are of no use any more. Broadcast authentication only with symmetric keys such as cluster key K c u fails to prevent an impersonation attack, since every neighbor of u shares the cluster key of u. Thus, we employ the reverse disclosure of one-way key chain K A u as in LEAP. Just prior to T min of x, pure node x broadcasts its neighbor list. Each receiving neighbor u of x checks whether the receiving time at u is prior to T x in the neighbor table. If yes, the neighbor list of x is now certified by each neighbor u. In summary, through the proposed three-way handshaking neighbor discovery process, pure node u identifies each immediate neighbor v and v's certified neighbor list (if v is an adult node), and keeps track of T min of v. Just prior to T min of u, node u broadcasts its direct neighbor list so that every neighbor of u accepts the list as verifiable. Then, node u becomes an adult node. After that, if newly-added node x initiates neighbor discovery with adult node u, node u identifies pure node x, keeps track of T min of x, provides u's certified neighbor list to x, and, in return, takes one-time certificates from x. Node u then broadcasts these one-time 6 One-time certificate, for instance M AC(K v , x|u), assures v that x is an immediate neighbor of u. It is generated by pure node x with master key of v. Table 1: An example of the Neighbor Table of u. Neighbor ID Certificate Verified Neighbor List v M AC(K v , v|u) u, w, t w M AC(K w , w|u) u, v, z x M AC(K x , x|u) u, r, q certificates, in order to assure u's existing neighbors of the discovery of new neighbor x. Thus, every time adult node u discovers newly-added node x through three-way handshaking , node u informs (by broadcasting) its existing neighbors of the discovery of new neighbor x. Also, whenever receiving neighbor list information from pure neighbor x, node u checks whether the receiving time at u is prior to T x in the neighbor table. If yes, u now accepts the neighbor list of x as verifiable. Through the above neighbor list verification in the bootstrapping phase, every node gets the knowledge of its neighbors' certified neighbors. Our Neighbor Watch System makes use of this information to prevent blind letter attack. With this knowledge, watch nodes are able to check whether the relaying packet's intended next-hop is a verified neighbor of the forwarding node. 3.3 Neighbor Table Maintenance The information obtained through neighbor list verification (e.g. its direct neighbors, corresponding certificates, neighbors' neighbor lists, etc) is stored in the neighbor table of each node. Table 1 shows an example of the neighbor table of node u. In densely-deployed sensor networks, the expected degree of a node is high. However, in this example, for simplicity, node u has only three neighbors v, w, and x as in Figure 3. The entries in the neighbor table are accessed and maintained with immediate neighbor IDs. For example, if node u overhears the packet sent from w to v, node u begins to listen in on v's traffic as a sub-watch node (since the neighbor table of u has both v's and w's entries in it). Unless v forwards the packet to a node of the Verified Neighbor List in v's entry by a certain timeout, sub-watch node u will forward the packet to its next-hop other than v; many existing routing protocols [5, 18, 21, 27, 37, 43] enable each node to maintain multiple potential next-hop. Once forwarding the packet, sub-watch node u becomes a primary-watch node and begins to listen in on its next-hop's traffic as described above. If newly-added node y initiates the three-way handshaking with u, node u provides its neighbor list to y by sending certificates in the neighbor table. Node u, in return from node y, takes the certificate for y and one-time certificates for u's existing neighbors. Then, node u stores the certificate in the new entry for y. However, node u does not store the one-time certificates but broadcasts them to its neighbors. If new neighbor y broadcasts its neighbor list within T min , node u stores the list in the entry for y. If node u is compromised, not only cryptographic key information but also certificates in the neighbor table are exposed. However, the attacker cannot misuse these certificates for other purposes. Since a certificate only attests neighborship between two specific nodes, it cannot be applied to any other nodes. In fact, it can be made even public. However, colluding nodes can deceive a pure node anyway, 64 by fabricating a bogus certificate. We will describe this limitation in Section 4.4. EVALUATION In this section, we evaluate the communication and storage cost, and analyze the security of our resilient forwarding scheme (Neighbor Watch System) as well as Neighbor List Verification. We then present the simulation results of our forwarding scheme. 4.1 Communication Cost Unlike the previously proposed diffusion-based reliable-forwarding schemes [21, 29, 39, 42] that exploit a large number of nodes to deliver a single packet, our scheme requires only the designated next-hop node to relay the packet, under the supervision of watch nodes. We note that, like overhearing by watch nodes in our scheme, those diffusion-based schemes require each node to listen to all its neighbors, since they forward a packet by broadcasting with no designated next-hop. With a smaller number of relaying nodes, our scheme makes a report successfully reach the base station. Thus, the average communication cost of our forwarding scheme for delivery of a single packet is smaller than those of the previous schemes. Our neighbor list verification during the bootstrapping phase requires the three-way handshaking neighbor discovery . Unlike the neighbor discovery between two pure nodes, the size of the messages exchanged between a pure and an adult node varies with the degree of the adult node. A large number of certificates caused by the high degree can be overburdensome to a single TinyOS packet which provides 29 bytes for data. Considering 8-byte certificates and a 4-byte 7 message authentication code (MAC), the adult node is able to include at most two neighbors' information in a single TinyOS packet. Thus, when the entire neighbor list cannot be accommodated within a single packet, the node should allot the list to several packets and send them serially. In a network of size N with the expected degree d of each node, the average number of packets invoked by a newly-added node per each node is nearly (d - 1) 2 /2(N - 1). Therefore, as node density d grows, the total number of packets transmitted from adult nodes to a newly-added node increases. However, neighbor discovery between a pure and an adult node occurs much less than between two pure nodes, since most neighbor discoveries throughout the network are between two pure nodes in the early stage of the network. Neighbor discovery between a pure and an adult node occurs generally when a new node is added to the network . 4.2 Storage Overhead In LEAP, each node keeps four types of keys and a manageable length of hash chain, which is found to be scalable. In our scheme, each node needs to additionally store its direct neighbors' certificates and their respective neighbor lists as in Table 1. Thus, for a network of the expected degree d and the byte size l of node ID, the additional storage requirement for each node is d (8 + ld) bytes. Although our storage requirement for these neighbor lists is O(d 2 ), for a reasonable degree d, memory overhead does 7 4-byte MAC is found to be not detrimental in sensor networks as in TinySec [24] which employs 4-byte MAC. u v ? u v ? C 1 C 2 Figure 4: Examples of critical area C 1 and C 2 . not exceed 1 KB (a Berkeley MICA2 Mote with 128 KB flash memory and 4 KB SRAM). For example, when d = 20 and l = 2, a node needs 960 bytes of memory to store such information. If node density of a network is so high that the required space for those neighbor lists significantly increases and the storage utilization becomes an issue, we can employ a storage-reduction technique such as Bloom filter [2]. For example, when d = 30 and l = 2, a node requires 2,040 bytes of additional space mainly for the neighbor lists. Instead of storing neighbors' neighbor lists, applying each of the neighbor lists (480 bits) to a Bloom filter (of 5 hash functions mapping to a 256 bit vector), a node needs the reduced space of 1,200 bytes for such information (with the false positive probability = 0.02). 4.3 Resilience to Packet-Dropping Attacks In face of maliciously packet-dropping nodes, the higher degree of multipath we provide, the more resiliency our scheme achieves against such attacks. The average degree of multipath depends on the number of sub-watch nodes around a packet-dropping node. Sub-watch nodes should be located in the region within the communication range of both forwarding node u and designated next-hop v. We refer to such a region as critical area. As in Figure 4, if nodes u and v are located farther away, the size of critical area C 2 gets smaller than that of C 1 , and the probability (p c ) that at least one sub-watch node exists in the critical area goes down. The probability (p c ) is p c = 1 - (1 - c) d-1 , where c is the ratio of the critical area size to the node's communication range, and the expected degree d of the node. To determine the appropriate degree d, we set the smallest critical area C 2 in Figure 4 as a lower bound case (c = 0.4). Figure 5 shows that, even in the lower bound critical area, with d = 6 and d = 10, probability p c is above 0.9 and above 0.99, respectively. Since, in a network of degree d, the probability that there exist m sub-watch nodes in the critical area of the ratio c is p(m) = d - 1 m c m (1 - c) d-m-1 , the expected number of sub-watch nodes, m, in the critical area is given by E[m] = (d - 1)c. Thus, in the lower bound (c = 0.4) critical area, when d = 10, 15, 20, the number of sub-watch nodes (i.e. the degree of multipath) is 3.6, 5.6, 7.6 on average, respectively. This 65 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 5 10 15 20 Degree of a node P r ob ab i l i t y p c Figure 5: Probability (p c ) that at least one sub-watch node exists in the lower bound (c = 0.4) critical area. shows that the higher degree of each node has, our scheme has the higher degree of multipath and resiliency against packet-dropping nodes. 4.4 The Security of Neighbor List Verification Our Neighbor List Verification(NLV) keeps the nice properties of LEAP. Adult nodes fail to establish pairwise keys with any adult nodes in arbitrary locations, so that the impact of a node compromise is localized. NLV performs the three-way handshaking neighbor discovery, instead of two-message exchange in LEAP. The three-way handshaking enables each node to verify not only its direct neighbors but also their respective neighbor lists. Moreover, this this three-way handshaking can be a potential solution to deal with irregularity of radio range [15, 37, 45]. In reality, due to the noise and some environmen-tal factors, radio range of each node is not exactly circular . So, communication links among nodes are asymmetric; node u can hear node v which is unable to hear u. With two-message exchange, only the node initiating the neighbor discovery is assured of the link's bidirectionality. By the three-way handshaking, both of neighbors can be assured of their symmetric connectivity. With NLV, only the verified lists are stored and utilized for our packet-forwarding scheme. NLV verifies the neighbor list of an adult node with certificates. These certificates merely attest neighborship between two specific nodes. Even if a node is compromised, the attacker fails to abuse the certificates of the captured node for other purpose. However, collusion among compromised nodes can fabricate bogus certificates in order to deceive a newly-added node. For example, consider two colluding nodes u and v at the different locations. When compromised node u discovers newly-added node x, node u provides x with u's neighbor list (maliciously including v in it). Even though node v is not an actual neighbor of u, colluding node v can generate the bogus certificate for u, M AC(K v , v|u). Then, x falsely believes that v is a direct neighbor of u. This attack, however , affects only the one newly-added node x. Thus, when compromised node u tries to launch the blind letter attack 8 , 8 Compromised node u transmits the relaying packet with its other surrounding adult neighbors of u can still detect it anyway. The more serious case is that colluding nodes exploit a newly-added node to generate bogus one-time certificates. For example, consider two colluding nodes u and v that share all their secret information as well as all their certificates . When newly-added node x initiates the three-way handshaking with u, compromised node u pretends to be v and provides x with v' neighbor list. Then, x in return provides u with one-time certificates for each neighbor of v; these one-time certificates falsely attest that v has new neighbor x. Node u sends this information to v over the covert channel. Then, v broadcasts these one-time certificates , and neighbors of v falsely believe that x is a direct neighbor of v. Unfortunately, we do not provide a proper countermeasure to defend against this type of man-in-the-middle attacks . However, we point out that this type of attacks has to be launched in the passive manner. The adversary has to get the chance of discovery of a newly-added node. In other words, compromised nodes wait for the initiation of the three-way handshaking from a newly-added node. Since the attacker does not know where the new nodes will be added, it has to compromise a sufficient number of legitimate nodes in order to increase the probability of discovery of newly-added nodes. As an active defense against such man-in-the-middle attacks , we can apply a node replication detection mechanism such as Randomized or Line-Selected Multicast [31], which revokes the same ID node at the different location claims. To successfully launch such man-in-the-middle attacks, two colluding nodes should pretend to be each other so that each of them claims to be at two different locations with the same ID. Location-binding key-assignment scheme by Yang et al. [39] with a little modification also can be a good solution to such attacks. Since it binds secret keys with nodes' geographic locations, the key bound to the particular location cannot be used at any arbitrary locations. Adopting this, NLV can check whether the claimed neighbors are really located within geographically two hops away. 4.5 Simulations To further evaluate the performance of our resilient forwarding scheme, we run simulations of our scheme in the presence of packet-dropping nodes on a network simulator, ns-2 [9]. 4.5.1 Simulation Model In our simulations, we deploy N sensor nodes uniformly at random within 500 500m 2 target field, with N = 300 and 600. Each sensor node has a constant transmission range of 30m, so that the degree of each node is approximately 10 (N = 300) and 20 (N = 600) on average. We position a base station and a source node in opposite corners of the field, at a fixed point (50, 50) and (450, 450), respectively. They are located approximately 18 hops away from each other. We distribute compromised nodes over an inner square area with 200m each side (from 150m to 350m of each side of the 500 500m 2 target area). Thus, compromised nodes are strategically-placed in between the base station and the source node. In the simulations, those compromised nodes drop all the relaying packets. next-hop id as v, so that x considers it forwarded correctly. 66 ( 300 nodes ) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 5 10 15 20 25 30 35 40 45 50 Number of Packet-dropping Nodes S u cc ess R a t i o Single Path Forwarding with NWS (a) Success ratio (N = 300, x = 0 50) ( 600 nodes ) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 10 20 30 40 50 60 70 80 90 100 Number of Packet-dropping Nodes S u cc ess R a t i o Single Path Forwarding with NWS (b) Success ratio (N = 600, x = 0 100) ( 300 nodes ) 0 10 20 30 40 50 60 70 80 90 100 0 5 10 15 20 25 30 35 40 45 50 Number of Packet-Dropping Nodes N u m b e r of R e l a yi n g N o d e s . Single Path Forwarding with NWS (c) The number of relaying nodes with N = 300 ( 600 nodes ) 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 Number of Packet-dropping Nodes N u m b e r of R e l a yi n g N o d e s . Single Path Forwarding with NWS (d) The number of relaying nodes with N = 600 Figure 6: Simulation Results (averaged over 100 runs). We use the typical TinyOS beaconing [17] with a little modification as a base routing protocol in our simulations. We add a hop count value in a beacon message 9 . To have multiple potential next-hops, when receiving a beacon with the same or better hop count than the parent node's, each node marks the node sending the beacon as a potential next-hop . Each simulation experiment is conducted using 100 different network topologies, and each result is averaged over 100 runs of different network topologies. 4.5.2 Simulation Results In the presence of compromised node dropping all the relaying packets, we measure the success ratio (i.e. the percentage of the packets that successfully reach the base station from the source) and the number of relaying nodes by the primitive single-path forwarding and with NWS in a network of size N, with N = 300 and 600. 9 The base station initiates the beacon-broadcasting, which floods through the network, in order to set up a routing tree. Figure 6(a) shows the success ratio in face of x packet-dropping nodes (varying x=0 to 50) in a 300-sensor-node network with the approximate degree d = 10. Although the success ratio gently decreases with x, it keeps up above 0.8 even with x = 30, with the help of NWS. This tendency of decreasing success ratio can be attributed to the degree d = 10 (3.6 sub-watch nodes on average) as well as an increasing number of packet-dropping nodes. Due to the strategically-placement of compromised nodes in our simulations , as x increases on, it is likely that a forwarding node's all potential sub-watch nodes themselves are packet-dropping nodes. Figure 6(c) shows the number of nodes that relay the packet from the source to the base station in the same experiments. Since the source is located about 18 hops away from the base station, the number of relaying nodes only with the single-path forwarding remains at 18. With NWS, the number of relaying nodes increases with x, in order to bypass an increasing number of packet-dropping nodes. In face of such nodes, our scheme converts single-path forwarding into multipath data forwarding, with the 67 help of sub-watch nodes around such packet-dropping nodes. Utilizing a cache for recently-received packets can suppress the same copy within a certain timeout, which reduces the number of relaying nodes. Figure 6(b) shows the success ratio in a 600-sensor-node network with the approximate degree d = 20 with x packet-dropping nodes (varying x=0 to 100). Unlike that with N = 300, the success ratio stays constantly at around 0.99 even with x = 100, with the help of NWS. This tendency of high success ratio can be mainly attributed to the degree d = 20 (7.6 sub-watch nodes on average in the lower bound case), which is found to be high enough to bypass a large number of packet-dropping nodes. Figure 6(d) shows the number of relaying nodes from the source to the base station in the same experiments. With NWS, the increase in the number of relaying nodes with x is more conspicuous than that with N = 300, since more than twice as many as sub-watch nodes help forward the packets so that it can bypass a large number of packet-dropping nodes anyway. In the simulation results, we note that our forwarding scheme dynamically adjusts its forwarding style, depending on the number of packet-dropping nodes en-route to the base station. As in Figures 6(c) and 6(d), while there exist none or a small number of packet-dropping nodes on the way, our scheme works almost like the single-path forwarding with the help of a few additional relaying nodes. On the other hand, when confronting a large number of packet-dropping nodes, our scheme makes full use of the help from additional relaying nodes, in order to successfully deliver the packet to the base station at any cost to the best efforts. CONCLUSIONS AND FUTURE WORK In this paper we focus on defending against compromised nodes' dropping of legitimate reports. We have presented a resilient packet-forwarding scheme using Neighbor Watch System (NWS) against maliciously packet-dropping nodes in sensor networks. In face of such nodes, NWS is specifically designed for hop-by-hop reliable delivery, and the prompt reaction of the conversion from single-path to multipath forwarding augments the robustness in our scheme so that the packet successfully reach the base station. In future work, we plan on further improving NLV to defend against the man-in-the-middle attacks, collusion among compromised nodes. Such attacks can be prevented by using a master key derived with not only a node ID but also its geographic information. We will also seek to address O(d 2 ) storage requirement for the neighbors' neighbor lists. Finally , we would like to perform an intensive experimental evaluation to compare our scheme with other reliable delivery protocols [10, 29, 42]. ACKNOWLEDGMENTS This work was supported by grant No.R01-2006-000-10073-0 from the Basic Research Program of the Korea Science and Engineering Foundation. REFERENCES [1] R. Anderson, H. Chan, and A. Perrig, Key Infection: Smart Trust for Smart Dust, IEEE ICNP 2004 [2] Burton H. Bloom, Space/Time Trade-offs in Hash Coding with Allowable Errors, Communication of the ACM, vol. 13, 422-426, 1970 [3] B. Carbunar, I. Ioannidis, and C. Nita-Rotaru, JANUS: Towards Robust and Malicious Resilient Routing in Hybrid Wireless Networks, ACM workshop on Wireless security (WiSe'04), Oct. 2004 [4] H. Chan, A. Perrig, and D. Song, Random Key Predistribution Schemes for Sensor Networks, IEEE Symposium on Security and Privacy, pp. 197-213, May 2003. [5] B. Deb, S. Bhatnagar, and B. Nath, ReInForM: Reliable Information Forwarding Using Multiple Paths in Sensor Networks, IEEE Local Computer Networks (LCN 2003), pp. 406-415, Oct. 2003. [6] J. Deng, R. Han, and S. Mishra, A Performance Evaluation of Intrusion- Tolerant Routing in Wireless Sensor Networks, 2nd International Workshop on Information Processing in Sensor Networks (IPSN 03), pp. 349-364, Apr. 2003. [7] J. Deng, R. Han, and S. Mishra, Intrusion Tolerance and Anti-Traffic Analysis Strategies for Wireless Sensor Networks, IEEE International Conference on Dependable Systems and Networks (DSN), pp. 594-603, 2004. [8] J. Deng, R. Han, and S. Mishra, Defending against Path-based DoS Attacks in Wireless Sensor Networks, ACM Workshop on Security of Ad-Hoc and Sensor Networks (SASN'05) , Nov, 2005. [9] K. Fall and K. Varadhan (editors), NS notes and documentation, The VINT project, LBL, Feb 2000, http://www.isi.edu/nsnam/ns/ [10] D. Ganesan, R. Govindan, S. Shenker, and D. Estrin, Highly Resilient, Energy-Efficient Multipath Routing in Wireless Sensor Networks, Computing and Communications Review (MC2R) Vol 1., pp. 11-25, 2002. [11] V. D. Gligor, Security of Emergent Properties in Ad-Hoc Networks, International Workshop on Security Protocols, Apr. 2004. [12] O. Goldreich, S. Goldwasser, and S. Micali, How to Construct Random Functions, Journal of the ACM, Vol. 33, No. 4, 210-217, 1986 [13] L. Eschenauer and V. D. Gligor, A Key-Management Scheme for Distributed Sensor Networks, 9th ACM Conference on Computer and Communication Security (CCS), pp. 41-47, Nov. 2002. [14] C. Hartung, J. Balasalle, and R. Han, Node Compromise in Sensor Networks: The Need for Secure Systems, Technical Report CU-CS-990-05, Department of Computer Science University of Colorado at Boulder, Jan. 2005 [15] T. He, S. Krishnamurthy, J. A. Stankovic, T. F. Abdelzaher, L. Luo, R. Stoleru, T. Yan, L. Gu, J. Hui, and B. Krogh, An Energy-Efficient Surveillance System Using Wireless Sensor Networks, ACM MobiSys'04, June, 2004 [16] W.R. Heinzelman, J. Kulik, H. Balakrishnan, Adaptive Protocols for Information Dissemination in Wireless Sensor Networks, ACM MobiCom99, pp. 174.185, 1999. [17] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. Culler, and K. Pister, System Architecture Directions for Networked Sensors, ACU ASPLOS IX, November 2000. 68 [18] X. Hong, M. Gerla, W. Hanbiao, and L. Clare, Load Balanced, Energy-Aware Communications for Mars Sensor Networks, IEEE Aerospace Conference, vol.3, 1109-1115, 2002. [19] Y.-C. Hu, D. B. Johnson, and A. Perrig, SEAD: Secure Efficient Distance Vector Routing for Mobile Wireless Ad Hoc Networks, IEEE Workshop on Mobile Computing Systems and Applications, pp. 3-13, Jun. 2002. [20] Y.-C. Hu, A. Perrig, and D. B. Johnson, Efficient Security Mechanisms for Routing Protocols, NDSS 2003, pp. 57-73, Feb. 2003. [21] C. Intanagonwiwat, R. Govindan and D. Estrin, Directed Diffusion: A Scalable and Robust Communication Paradigm for Sensor Networks, MobiCom'00, Aug. 2000. [22] D. Johnson, D.A. Maltz, and J. Broch, The Dynamic Source Routing Protocol for Mobile Ad Hoc Networks (Internet-Draft), Mobile Ad-hoc Network (MANET) Working Group, IETF, Oct. 1999. [23] C. Karlof and D. Wagner, Secure Routing in Wireless Sensor Networks: Attacks and Countermeasures, The First IEEE International Workshop on Sensor Network Protocols and Applications, pp. 113-127, May 2003 [24] C. Karlof, N. Sastry, and D. Wagner, TinySec: A Link Layer Security Architecture for Wireless Sensor Networks, ACM SensSys'04, pp. 162-175, Nov. 2004. [25] I. Khalil, S. Bagchi, and C. Nina-Rotaru, DICAS: Detection, Diagnosis and Isolation of Control Attacks in Sensor Networks, IEEE SecureComm 2005, pp. 89 100 , Sep. 2005 [26] Y. Liu and W. K.G. Seah, A Priority-Based Multi-Path Routing Protocol for Sensor Networks, 15th IEEE International Symposium on Volume 1, 216 - 220, 2004 [27] S.-B. Lee and Y.-H. Choi, A Secure Alternate Path Routing in Sensor Networks, Computer Communications (2006), doi:10.1016/j.comcom.2006.08.006. [28] S. Marti, T.J. Giuli, K. Lai, and M. Baker, Mitigating Routing Misbehavior in Mobile Ad Hoc Networks, ACM/IEEE International Conference on Mobile Computing and Networking, pp. 255-265, 2000 [29] H. Morcos, I. Matta, and A. Bestavros, M 2 RC: Multiplicative-Increase/Additive-Decrease Multipath Routing Control for Wireless Sensor Networks, ACM SIGBED Review, Vol. 2, Jan 2005. [30] J. Newsome, E. Shi, D. Song, and A. Perrig, The Sybil Attack in Sensor Networks: Analysis and Defenses, IEEE IPSN'04, pp. 259-268, Apr. 2004. [31] B. Parno, A. Perrig, and V. D. Gligor, Distributed Detection of Node Replication Attacks in Sensor Networks, the 2005 IEEE Symposium on Security and Privacy, pp. 49-63, May 2005. [32] A. Perrig, R. Szewczyk, V. Wen, D. Culler, and J. Tygar, SPINS: Security Protocols for Sensor Networks, ACM MobiCom'01, pp. 189-199, 2001. [33] A. Perrig, J. Stankovic, and D. Wagner, Security in Wireless Sensor Networks, Communications of the ACM, 47(6), Special Issue on Wireless sensor networks, pp.53- 57, Jun. 2004 [34] B. Przydatek, D. Song, and A. Perrig, SIA: Secure Information Aggregation in Sensor Networks, 1st International Conference on Embedded Networked Sensor Systems, 255-256, 2003 [35] E. Shi and A. Perrig, Designing Secure Sensor Networks, Wireless Communications, IEEE Volume 11, Issue 6, pp. 38-43, Dec. 2004. [36] D. Tian and N.D. Georganas, Energy Efficient Routing with Guaranteed Delivery in Wireless Sensor Networks, IEEE Wireless Communications and Networking (WCNC 2003), IEEE Volume 3, 1923 1929 , March 2003 [37] A. Woo, T. Tong, and D. Culler, Taming the Underlying Challenges of Reliable Multhop Routing in Sensor Networks, ACM SenSys03, Nov, 2003 [38] A. Wood and J. Stankovic, Denial of Service in Sensor Networks, IEEE Computer, Vol.35, 54-62, Oct. 2002 [39] H.Yang, F. Ye, Y. Yuan, S. Lu and W. Arbough, Toward Resilient Security in Wireless Sensor Networks, ACM MobiHoc'05, 34-45, May 2005 [40] Y. Yang, X. Wang, S. Zhu, and G. Cao SDAP: A Secure Hop-by-Hop Data Aggregation Protocol for Sensor Networks, ACM MobiHoc'06 May 2006 [41] F. Ye, H. Luo, S. Lu and L. Zhang, Statictial En-route Filtering of Injected False Data in Sensor Networks, IEEE INFOCOM, 2004 [42] F. Ye, G. Zhong, S. Lu and L. Zhang, GRAdient Broadcast: A Robust Data Delivery Protocol for Large Scale Sensor Networks, ACM Wireless Networks (WINET), March 2005 [43] Y. Yu, R. Govindan, and D. Estrin, Geographical and Energy Aware Routing: a recursive data dissemination protocol for wireless sensor networks, UCLA Computer Science Department Technical Report UCLA/CSD-TR-01-0023, May 2001. [44] W. Zhang and G. Cao, Group Rekeying for Filtering False Data in Sensor Networks: A Predistribution and Local Collaboration-Based Approach, IEEE INFOCOM'05. Vol. 1, 503-514, March 2005 [45] G. Zhou, T. He, S. Krishnamurthy, and J. A. Stankovic, Impact of radio irregularity on wireless sensor networks, the 2nd International Conference on Mobile Systems, Applications, and Services (MobiSys04), June, 2004 [46] S. Zhu, S. Setia, and S. Jajodia, LEAP: Efficient Security Mechanisms for Large-Scale Distributed Sensor Networks, The 10th ACM Conference on Computer and Communications Security (CCS '03), 62-72, 2003 [47] S.Zhu, S. Setia, S. Jajodia, and P. Ning, An Interleaved Hop-by-Hop Authentication Scheme for Filtering False Data in Sensor Networks, IEEE Symposium on Security and Privacy, 2004 69

Neighbor Watch System;legitimate node;Reliable Delivery;Packet-dropping Attacks;aggregation protocols;malicious node;robustness;critical area;single-path forwarding;Sensor Network Security;cluster key;secure ad-hoc network routing protocol;Secure Routing;degree of multipath

180

SIMULATING OPTION PRICES AND SENSITIVITIES BY HIGHER RANK LATTICE RULES

In this paper we introduce the intermediate rank or higher rank lattice rule for the general case when the number of quadrature points is n t m, where m is a composite integer, t is the rank of the rule, n is an integer such that (n, m) = 1. Our emphasis is the applications of higher rank lattice rules to a class of option pricing problems. The higher rank lattice rules are good candidates for applications to finance based on the following reasons: the higher rank lattice rule has better asymptotic convergence rate than the conventional good lattice rule does and searching higher rank lattice points is much faster than that of good lattice points for the same number of quadrature points; furthermore, numerical tests for application to option pricing problems showed that the higher rank lattice rules are not worse than the conventional good lattice rule on average.

Introduction It is well known in scientific computation that Monte Carlo (MC) simulation method is the main method to deal with high dimensional ( 4) problems. The main drawback for this method is that it converges slowly with convergence rate O( 1 N ), where N is the number of points (or samples or simulations), even after using various variance reduction methods. To speed it up, researchers use quasi-random or low-discrepancy point sets, instead of using pseudo-random point sets. This is the so called quasi-Monte Carlo (QMC) method. There are two classes of low-discrepancy sequences (LDS). The first one is constructive LDS, such as Halton's sequence, Sobol's sequence, Faure's sequence, and Nieder-reiter's (t, m, s)-nets and (t, s)-sequence. This kind of LDS has convergence rate O( (log N ) s N ), where s is the dimension of the problem, N is, again, the number of points. The second class is the integration lattice points, for example , good lattice points (GLP). This type of LDS has convergence rate O( (log N ) s N ), where &gt; 1 is a parameter related to the smoothness of the integrand, s and N are the same as above. The monograph by Niederreiter [1] gives very detailed information on constructive LDS and good lattice points, while the monograph by Hua and Wang [2] and Sloan and Joe [3] describe good lattice rules in detail. Unlike the constructive sequences, the construction of good lattice points is not constructive in the sense that they could be found only by computer searches (except in the 2- dimensional case, where good lattice points can be constructed by using the Fibonacci numbers). Such searches are usually very time consuming, especially when the number of points is large or the dimension is high, or both. Therefore, to develop algorithms which can be used in finding good lattice points fast is of practical importance. This paper discusses the applications of the intermediate rank or higher rank lattice rules (HRLR) to option pricing problems. The motivations of using higher rank lattice points are as follows. For a class of finance problems , we found that using the randomized good lattice points (GLP) can reach much better convergence than the randomized constructive quasi-random sequences (such as, Sobol sequence), let alone the pseudo-random point sets, see [4] about this (in that paper, the lattice points were taken from [2]). The theory given in Section 2 shows that the error bound of a higher rank lattice rule is smaller than that of a good lattice rule, at least asymptotically. And searching higher rank lattice points is much faster than searching good lattice points. Our extensive numerical results confirmed this fact. Some results are listed in Section 2. Furthermore, the results in Section 3 showed that the standard errors of the randomized higher rank lattice points are smaller than those of the randomized good lattice points (most of the times), which are much smaller than the standard errors of the randomized Sobol sequence, when these quasi-random point sets are applied to some financial derivative pricing problems in simulating option values and sensitivities. Higher Rank Lattice Rules Detailed information about lattice rules can be found in the literature, such as [1], [2] and [3]. We start to introduce lattice rules briefly by considering an integral 530-113 258 If = C s f (x)dx, (1) where C s = [0, 1] s is the s-dimensional unit hyper-cube , f (x) is one-periodic in each component of x, i.e. f (x) = f (x + z), z Z s (the set of s- dimensional integer points), x R s (the s-dimensional real space). An s-dimensional integration lattice L is a discrete subset of R s that is closed under addition and subtraction and contains Z s as a subset. A lattice rule for (1) is a rule of the form Qf = 1 N N -1 j=0 f (x j ), (2) where {x 0 , , x N -1 } L U s with U s = [0, 1) s , N is called the order of the rule. Now we consider the intermediate rank or higher rank lattice rules, i.e. rules of the form Q t f = 1 n t m n-1 k t =0 n-1 k 1 =0 m-1 j=0 f ({ j m g+ k 1 n y 1 ++ k t n y t }) (3) for 1 t s, where (m, n) = 1 and g, y 1 , , y t Z s . Notice that t = 0 or t = 1 and n = 1 in (3) is just the conventional good lattice points rule (we refer it to the rank-1 rule in this paper). Under some conditions (see, for example, Theorem 7.1, [3]) on g, y 1 , , y t , the points in (3) are distinct, so that Q t is a lattice rule of order N = n t m, and it has rank t. Korobov (1959) gave the first existence of good lattice points in the case where N is a prime number. Niederreiter (1978) extended the existence to general number N . Disney and Sloan proved the existence and obtained the best asymptotic convergence rate for general N in good lattice points case. The existence of good rank t rules can be established , but much more complicated. We introduce Definition 1. For any integer N 2, let G = G(N ) = {g = (g 1 , , g s ) Z s , (g j , N ) = 1 and -N/2 &lt; g j N/2, 1 j s}. Let y 1 , , y t Z s be fixed. The mean of P (Q t ) over G is M (n) ,t (m) = 1 Card(G) gG P (Q t ), &gt; 1 (4) For the sake of simplicity, Sloan et al chose the special form of y j with all the components 0 except the jth which is 1 - the so-called copying rule. Thus (3) becomes Q t f = 1 n t m n-1 k t =0 n-1 k 1 =0 m-1 j=0 f ({ j m g+ (k 1, , k t , 0, , 0) n }). (5) With this choice, P (Q t ) is easily calculated as follows. For &gt; 1, 1 t s and n 2, define f (n) ,t (x) = ( t j=1 F (n) (x j )) s k=t+1 F (x k ), (6) where F (n) (x) = 1 + 1 n hZ | h | e (hx), and F (x) = 1 + hZ | h | e (hx). If Q (n) t f is the m-point lattice rule defined by Q (n) t f = 1 m m-1 j=0 f ({ jn m g 1 , , jn m g t , j m g t+1 , , j m g s }), (7) then P (Q t ) = Q (n) t f (n) ,t - 1. (8) For 1 t s and g = (g 1 , , g s )G, denote w = (ng 1 , , ng t , g t+1 , , g s ), and r t (h) =( t j=1 r(nh j )) s k=t+1 r(h k ), h =(h 1, , h s ). Then applying the rank-1 lattice rule with generating vector w, we have P (Q t ) = hw0 (mod m) r t (h) -1. (9) The existence of good rank-t rules and the error bounds for prime m was established by Joe and Sloan (Theorem 7.4, [3]). The corresponding results for general m were discovered and proved in [5], and is stated below. Theorem 1. For &gt; 1, 1 t s, n 1 integer, m &gt; 0 any integer with (n, m) = 1, then M (n) ,t (m) = 1 m t k=0 s-t l=0 ( t k )( s-t l ) (2()) n k k+l p|m F ,k+l (p ) - 1, (10) where the product is over all prime factors p of m, p is the highest power of p dividing m, p|m F ,0 (p ) = m, and for k 1, F ,k (p ) is given by F ,k (p ) = 1 + (-1) k (1 - 1/p -1 ) k (1 - 1/p (k-1) ) (p - 1) k-1 (1 - 1/p k-1 ) . (11) Remark: Using the Binomial Theorem, we can obtain the result of Theorem 7.4 in [3] (the case when m is prime) from Theorem 1, since the assumption that m is prime and n is not a multiple of m implies that (n, m) = 1. Moreover, the result of Theorem 1 also holds for n = 1 or t = 0. In either case, the right hand side of (10) is just 259 the case of rank-1 in [3], and if n 2 and t = s, then we obtain the result of maximal rank case in [3]. Now as in the case of rank-1, we give an upper bound for M (n) ,t (m) and hence P (Q t ). Corollary 1. Under the conditions of Theorem 1, we have M (n) ,t (m) 4() 2 (m) [( s-t 2 ) + 1 n 2 ( t 2 ) + (s - t)t n ] + 1 m {[a(1 + 2()) s-t + b(1 - 2()) s-t ] +[a(1 + 2() n ) t + b(1 - 2() n ) t ] + 1 n [a(1 + 2()) s + b(1 - 2()) s ]}, (12) where ( s-t 2 ) = 1 for s - t &lt; 2, ( t 2 ) = 1, for t &lt; 2; a = (3) (6) + 1 2 1.68, and b = a - 1. Hence M (n) ,t (m) = O( log log m m ), asm . (13) Theorem 2. Let (m) = (1 - s/ log m) -1 . If m &gt; e s/(-1) then there is a g G such that P (Q t ) M (n) (m),t (m) /(m) . (14) If s 3, then M (n) (m),t (N ) /(m) 1 n t ( 2e s ) s (log m) s m (15) as m , where f (x) h(x) as x means lim x f (x) h(x) = 1. It is hard to obtain a precise comparison result between the mean for the case of t = 0, i.e., M (n) ,t (m), and the corresponding mean for the case of t = 0 rule, i.e., M (n t m), even when m is prime, as pointed out in [3]. Notice that the number of points for rank t rule is n t m, and we should use the same number of points when comparing efficiency or convergence rate among different methods . We give an approximate result on this direction based on (15) and a result in [3] similar to (15). Corollary 2. For &gt; 1, 1 t s, n 1 integer , m &gt; 0 any integer with (n, m) = 1, let 1 (m) = (1 - s/ log m) -1 , 2 (n t m) = (1 - s/ log(n t m)) -1 . If m &gt; e s/(-1) then M (n) 1 (m),t (m) / 1 (m) M 2 (n t m) (n t m) / 2 (n t m) log m log m + t log n s &lt; 1. (16) From Corollary 2, we can roughly see that P (Q t ) &lt; P (Q 1 ) for t &gt; 1 with the same order (number of points) for both rules, at least asymptotically. Our numerical tests showed that it is true even for small number of points. Furthermore, higher rank good lattice points can also be found by computer search via minimizing P (Q t ) based on (8), but using m instead of using n t m (as in the case of good lattice points). Therefore, searching higher rank lattice points is much faster than searching rank-1 lattice points. Usually, the good lattice points were found by searching Korobov type g = (1, b, b 2 , ..., b s-1 ) mod m (compo-nentwise ) with (b, m) = 1. Sloan and Reztsov proposed a new searching algorithm - the component-by-component method. We searched extensively for both types of points. Based on the search results we found that these two types of lattice points are comparable in both errors and searching times for the same rank and the same number of points. We only report the Korobov type lattice points here limited to space. So far as we know, the theory of copying higher rank lattice rule is valid under the assumption that (n, m) = 1. We conjecture that this restriction can be relaxed. We are unable to prove this yet so far. But our numerical results strongly support our conjecture, see Table 1 (only partial results are listed). In this table, the comparison is based on the same number of points, where Kor t0 stands for the Korobov type lattice points with t = 0 (i.e., rank-1 case), sim-ilarly for Kor t4. Time is measured in seconds. Whenever time is zero, it just means that the time used in searching is less than 0.5 seconds. The CPU times used in searching may be machine dependent. Dev-C++ was used as our programming language (run on a laptop under Windows system). In order to measure CPU time as precise as possible , all the programs were run on the same machine and only one program, no any other programs, was run one at a time on the machine. Our searching results showed that within the same type of lattice points, the higher the rank, the smaller the P 2 , and the faster the search. The search time for rank=4 in the case of number of point =32768 is about 1 second, those for all the other cases are less than 0.5 seconds. Table 1 : Computer search results of t = 0 and t = 4, with (n, m) = 1, n = 2, m = a power of 2, dimension = 5. Kor t0 Kor t4 2 t m b P 2 Time b P 2 1024 189 0.735 0 5 0.373 2048 453 0.264 1 27 0.164 4096 1595 0.121 3 21 0.067 8192 2099 0.048 10 61 0.026 16384 2959 0.018 43 35 0.010 32768 1975 0.007 169 131 0.004 Applications to Option Pricing Under the Black-Scholes framework, many European options can be expressed in terms of multivariate normal distributions . Examples are options on maximum and minimum of n assets, discrete lookback options, discrete shout options, discrete partial barrier options, reset options, etc., see [6] and the references therein. In this section, we apply both the Monte Carlo and the quasi-Monte Carlo methods to applied finance area-260 option pricing, and compare the efficiencies among different methods. For the quasi-Monte Carlo methods, we use Sobol sequence and both rank-1 and higher rank lattice points. The Sobol sequence is usually the best among the constructive LDS based on our tests. To compare the efficiencies of different methods, we need a benchmark for fair comparisons. If the exact value of the quantity to be estimated can be found, then we use the absolute error or relative error for comparison. Otherwise , we use the standard error (stderr) for comparison. Here stderr = N , where 2 is the unbiased sample variance , N is the sample size. For LDS sequences, we define the standard error by introducing random shift as follows . Assume that we estimate = E[f (x)], where x is an s-dimensional random vector. Let {x i } m i=1 C s be a finite LDS sequence, {r j } n j=1 C s be a finite sequence of random vectors. For each fixed j, we have a sequence {y (j) i } m i=1 with y (j) i = x i + r j . It can be shown that such a sequence still has the same convergence rate as the original one. Denote j = 1 m m i=1 f (y (j) i ) and = 1 n n j=1 j .The unbiased sample variance is 2 = n j=1 ( j -) 2 n-1 = n n j=1 2 j n j=1 j 2 (n-1)n . Then the standard error is defined by stderr = n . The efficiency of a QMC method (after randomization) over the MC method is defined as the ratio of the standard error of the MC method to the standard error of a QMC method (both methods have the same number of points, otherwise the comparison is not fair). As an example, let us consider the computation of call options on maximum of s assets. Using martingale method, Dufresne et al derived in [7] that the value of a call option can be expressed in terms of multivariate normal distributions : V = V s max ({S i }, { i }, 0 , r, q, ) = s i=1 S i e -q i T N s (e i1 , ..., e i,i-1 , d (i) i (K, T ), e i,i+1 , e is ; i )Ke -rT 1 - N s (-d Q 1 (K, T ), ..., -d Q s (K, T ); 0 ) (17) where e ik = log(S i /S k ) + T 2 ik /2 ik T , ik = 2 i - 2 i k + 2 k , d (i) i (K, T ) = log(S i /K) + (r + 2 i /2)T i T , d Q i (K, T ) = log(S i /K) + (r 2 i /2)T i T , 0 = ( jk ) ss and for i = 1, ..., s, i = ( (i) jk ) ss with (i) jk = 2 i + jk j k ij i j ik i k ij ik , j, k = i; (i) ik = i ik k ik , i = k; ii = 1. Thus, in order to estimate the option values, we need to estimate the following s-variate normal distribution H(a, ) = 1 det()(2) s a 1 a s exp (1 2 x t -1 x)dx, where a = (a 1 , a 2 , ..., a s ), - a i + ( i = 1, 2, ..., s), x R s , dx =dx 1 ...dx s , = ( ij ) ss is a positive definite correlation matrix. Details about the computation of multivariate normal distributions can be found in [8]. Notice that after the transformation, the s-dimensional integral for H(a, ) is transformed into an s - 1 dimensional integral . For the numerical demonstration, we consider a call option on maximum of 6 stocks. In our simulations, each method was randomly shifted, including the MC method, so that each method has the same number of points. We took the number of random shifts to be 10, other parameters are s = 6, K {$90, $100, $110}, r = 10%, S i = $100, i = 0.2, ij = 0.5, i = j, i, j = 1, ..., 6. Besides the option values, the option sensitivities or Greek letters i = V S i , ij = 2 V S i S j , V i = V i , = V T and = V r are very important quantities in financial risk management and trading. They are usually harder to obtain than the option values themselves. The results where K = $100 are listed in the following Tables 2, 3 and 4, and the results where K = $90 and K = $110 are similar and are omitted here. In these tables, column 1 contains the numbers of points, numbers in the MC column are the standard errors, those in the columns of quasi-Monte Carlo methods are efficiencies of the corresponding methods over the Monte Carlo method. Here we do not include the CPU times for different methods since these programs were run on a main-frame using UNIX system, and there were many other programs were also running at the time I ran these programs. And I think that the CPU times measured in this way are not precise. Table 2 : Comparison of estimated call option values and efficiencies, the option value is $28.81 with standard error 1.5099e-06 obtained by higher rank lattice rule (rank=4) using 2 14 =16384 points with 10 random shifts. The standard error is zero in my simulation by the same rule using 2 15 =32768 points with 10 random shifts. N MC Sobol Kor t0 Kor t4 2 10 0.6832 10.8 219.9 200.5 2 11 0.4834 27.8 666.8 1194.9 2 12 0.3413 39.6 2104.7 371.8 2 13 0.2409 41.2 18129.2 22758.4 2 14 0.1703 110.2 30156.0 112804.8 2 15 0.1206 101.3 253540.7 * From Table 2, we observed that the randomized lattice rules achieve much better results than the randomized Sobol's sequence does, the latter is about 10 to 110 time more efficient than the MC method. The randomized Korobov type higher rank lattice points beat the randomized 261 rank-1 lattice points, except when N = 2 10 = 1024 and 2 12 = 4096. Table 3 : Comparison of estimated option sensitivity (Greek letter, 1 in this table) values and efficiencies, the value of 1 is 0.1898 with standard error 4.4427E-09 obtained by higher rank lattice rule (rank=4) using 2 15 =32768 points with 10 random shifts. N MC Sobol Kor t0 Kor t4 2 10 0.0151 7.3 135.5 98.1 2 11 0.0107 10.7 330.8 1226.6 2 12 0.0077 9.9 900.7 110.7 2 13 0.0054 19.7 8861.8 9833.5 2 14 0.0038 30.7 24742.9 70692.9 2 15 0.0027 42.8 137025.0 605473.8 Again, the randomized lattice rules are much more efficient than the randomized Sobol's sequence, the latter is about 8 to 43 time more efficient than the MC method. Kor t4 is more efficient than Kor t0 except when N = 2 10 = 1024 and 2 12 = 4096. Table 4 : Comparison of estimated gamma ( 11 = 2 V S 2 1 ) values and efficiencies, the value of 11 is 0.01631 with standard error 1.2418e-09 obtained by higher rank lattice rule (rank=4) using 2 15 =32768 points with 10 random shifts. N MC Sobol Kor t0 Kor t4 2 10 4.6E-4 8.0 107.6 12.8 2 11 3.3E-4 14.0 118.6 223.7 2 12 2.3E-4 19.6 484.3 30.3 2 13 1.7E-4 22.6 4217.8 4457.5 2 14 1.2E-4 24.1 3108.7 28713.5 2 15 8.2E-5 40.0 108047.3 66385.4 The conclusion is similar to that of Table 3. The randomized Kor t4 is more efficient than the randomized Kor t0 except when N = 2 10 = 1024 , 2 12 = 4096 and 2 15 = 32768. In our simulations, the pseudo random number generator we used is ran2() in [9]. The periodizing function used is (x) = 1 2 (2x - sin(2x)) . Conclusion In this paper, We introduced the higher rank lattice rules and gave a general expression for the average of P (Q t ) for higher rank lattice rule over a subset of Z s , an upper bound and an asymptotic rate for higher rank lattice rule. The results recovered the cases of good lattice rule and maximal rank rule. Computer search results showed that P 2 s by the higher rank lattice rule were smaller than those by good lattice rule, while searching higher rank lattice points was much faster than that of good lattice points for the same number of quadrature points. Numerical tests for applications to an option pricing problem showed that the higher rank lattice rules (t &gt; 0) usually beat the conventional good lattice rule (t = 0 case). Both of these rules showed significant superiority over the Sobol sequence. Our tests (not listed here) on other types of options showed similar efficiency gains of higher rank lattice rules over good lattice rules, though the gains may vary. Since searching higher rank lattice points is much faster than that of rank - 1 lattice points (say the rank is larger than 2); the search algorithm is simple; and the values of P 2 for higher rank lattice points are smaller than that for the rank - 1 points; furthermore, (standard) errors obtained by higher rank lattice rules to practical problems are not worse than those by the rank - 1 rules on average, the higher rank lattice rules are good candidates for applications . One unsolved problem in lattice rules (whether high rank or not) is the periodizing seems not work well in high dimensions. It needs futher exploration. Acknowledgements This research was partially supported by an Natural Sciences and Engineering Research Council of Canada (NSERC) grant. References [1] H. Niederreiter, Random Number Generation and Quasi-Monte Carlo Methods, SIAM, Philadelphia, 1992. [2] L. Hua and Y. Wang, Applications of Number Theory in Numerical Analysis, Springer-Verlag, 1980. [3] I. H. Sloan and S. Joe, Lattice Methods for Multiple Integration, Oxford University Press, New York, 1994. [4] P. Boyle, Y. Lai and K. S. Tan, Pricing Options Using lattice rules, North American Actuarial Journal, 9(3), 2005, 50-76. [5] Y. Lai, Monte Carlo and Quasi-Monte Carlo Methods and Their Applications, Ph. D Dissertation, Department of Mathematics, Claremont Graduate University, California, USA, 2000. [6] P. Zhang, Exotic Options, 2nd edition, World Scientific , 1998. [7] P.C. Dufresne, W. Keirstead and M. P. Ross, Pricing Derivatives the Martingale Way, working paper, 1996. [8] Y. Lai, Effcient Computations of Multivariate Normal Distributions with Applications to Finance, working paper, Departmetn of Mathematics, Wilfrid Laurier University, Waterloo, Ontario, Canada, 2005. [9] W. H. Press, S. A. Teukolsky, W. T. Vetterling and B. P. Flannery, Numerical recipes in C: The Art of Scientific Computing, Cambridge University Press, 1992. 262

Monte Carlo and Quasi-Monte Carlo methods;Simulation of multivariate integrations;Lattice rules;Option Pricing

181

SmartCrawl: A New Strategy for the Exploration of the Hidden Web

The way current search engines work leaves a large amount of information available in the World Wide Web outside their catalogues. This is due to the fact that crawlers work by following hyperlinks and a few other references and ignore HTML forms. In this paper, we propose a search engine prototype that can retrieve information behind HTML forms by automatically generating queries for them. We describe the architecture, some implementation details and an experiment that proves that the information is not in fact indexed by current search engines.

INTRODUCTION The gigantic growth in content present in the World Wide Web has turned search engines into fundamental tools when the objective is searching for information. A study in 2000 [11] discovered that they are the most used source for finding answers to questions, positioning themselves above books, for example. However, a great deal of relevant information is still hidden from general-purpose search engines like AlltheWeb.com or Google. This part of the Web, known as the Hidden Web [7], the Invisible Web [6, 9] or the Deep Web [1] is growing constantly, even more than the visible Web, to which we are accustomed [6]. This happens because the crawler (the program that is responsible for autonomous navigating the web, fetching pages) used by current search engines cannot reach this information . There are many reasons for this to occur. The Internet's own dynamics, for example, ends up making the index of search engines obsolete because even the quickest crawlers only manage to access only a small fraction each day of the total information available on the Web. The cost of interpretation of some types of files, as for example Macromedia Flash animations, compressed files, and programs (executable files) could be high, not compensating for the indexing of the little, or frequently absent, textual content. For this reason, that content is also not indexed for the majority of search engines. Dynamic pages also cause some problems for indexing. There are no technical problems, since this type of page generates ordinary HTML as responses for its requests. However , they can cause some challenges for the crawlers, called spider traps [8], which can cause, for example, the crawler to visit the same page an infinite number of times. Therefore, some search engines opt not to index this type of content. Finally, there are some sites that store their content in databases and utilize HTML forms as an access interface. This is certainly the major barrier in the exploration of the hidden Web and the problem that has fewer implemented solutions. Nowadays none of the commercial search engines that we use explore this content, which is called the Truly Invisible Web [9]. Two fundamental reasons make crawling the hidden Web a non-trivial task [7]. First is the issue of scale. Another study shows that the hidden content is actually much greater than what is currently publicly indexed [1]. As well as this, the interface for access to this information serves through the HTML forms are projected to be manipulated and filled by humans, creating a huge problem for the crawlers. In this paper, we propose a search engine prototype called SmartCrawl, which is capable of automatically attaining pages that are not actually recoverable by current search engines, and that are "secreted" behind HTML forms. The rest of this article is organised in the following manner . Section 2 shows related work and in the sequence, we explain the construction of HTML forms and how they can be represented. In sections 4 and 5, we describe the prototype . In Section 6 the experimental results are highlighted and finally, in Section 7, we conclude the paper. 9 RELATED WORK There are some proposals for the automatic exploration of this hidden content. Lin and Chen's solution [6] aims to build up a catalogue of small search engines located in sites and, given the user searching terms, choose which ones are more likely to answer them. Once the search engines are chosen by a module called Search Engine Selector, the user query is redirected by filling the text field of the form. The system submits the keywords and waits for the results that are combined subsequently and sent to the users' interface. The HiWE [7] is a different strategy which aims to test combinations of values for the HTML forms at the moment of the crawling (autonomous navigation), making the indexing of the hidden pages possible. Once a form in a HTML page is found, the crawler makes several filing attempts, analyses and indexes the results of the obtained pages. Moreover, the HiWE has a strategy to extract the labels of HTML forms by rendering the page. This is very useful to obtain information and classify forms and helps to fill in its fields. There are other approaches that focus on the data extraction . Lage et al. [4] claims to automatically generate agents to collect hidden Web pages by filling HTML forms. In addition to this, Liddle et al. [5] perform a more comprehensive study about form submissions and results processing . This study focus on how valuable information can be obtained behind Web forms, but do not include a crawler to fetches them. EXTRACTING DATA FROM BEHIND THE FORM HTML forms are frequently found on the web and are generally used for filtering a large amount of information. As shown in Figure 1, from a page with one form the user can provide several pieces of data which will be passed on to a process in the server, which generates the answer page. The current crawlers do not fill in form fields with values, making them the major barrier for exploration of the hidden Web. In order to achieve this, it is vital to extract several pieces of information from the form. An HTML form can be built on different manners, including various types of fields such as comboboxes, radio buttons, checkboxes, text fields, hidden fields and so on. However, the data sent to the server through the Common Gateway Interface (CGI) is represented by proper codified pairs (name, value). This way, we can characterise a form with which has n fields as a tuple: F = {U, (N 1 , V 1 ), (N 2 , V 2 ), ..., (N n , V n ) } (1) where U is the URL for the data that has been submitted, and (N n , V n ) are the pairs (name, value) [5]. However, this is a simplification, since there are much more information associated with HTML forms. An example is the method by which the form data will be sent to the server, that is, by HTTP GET or POST. Moreover, some fields possess domain limitations (e.g. text fields with a maximum size, comboboxes). To do an analysis of the form and extract relevant information is not an easy task, but the most difficult step surely is to extract the field's labels. This is because generally there is not a formal relationship between them in the HTML code. For example, the label for a text field can be placed above it, separated by a BR tag, it can be beside it, or it can be inserted inside table cells. All these pieces of data are absolutely necessary to be extracted for surpassing HTML forms and fetching the results page. THE SMARTCRAWL The aim of SmartCrawl is to bring a strategy that allows a more complete exploration of the hidden content. To achieve this, it managed to generate values for a largest number of forms. Furthermore, it has an architecture very similar to current commercial search engines, which means it permits an easier implantation of strategies vastly used to gain performance and scalability as presented by [10] or [2]. 4.1 Execution of the Prototype SmartCrawl is above all a search engine and, therefore, contains all its essential components. The difference is in the fact that each component has adaptations and some extra features which enable them to explore the hidden content of the Web. The main goal is to index only the pages that potentially are in the non-explorable part of the Web. To extract the content from behind these forms, SmartCrawl generates values for its fields and submits them. These values are chosen in two different moments: in the indexing and when an user performs a search. In the indexing phase, once it finds a form, the SmartCrawl extracts a set of pieces of information from it that allow queries (combinations of possible values for the form) to be created. New queries are also generated when the user performs a search. For this, the forms that are more likely to answer to the search receive the supplied keywords. However, contrary to the implementation of Lin and Chen [6] the obtained results are also scheduled for indexing and not only returned to the user interface. The process of execution of SmartCrawl is constituted in the following steps: (1) finding the forms, (2) generating queries for them, (3) going to the results and (4) searching created indexes. 4.1.1 Finding forms The first step in the execution process of the SmartCrawl is the creation of a number of crawlers that work in parallel searching for pages that include HTML forms. Every page found is then compressed and stored for further analysis. At this moment, it acts like a common crawler, following only links and references to frames. The pages stored by the crawler are decompressed afterwards by a indexing software component which extracts pieces of information from each of the forms found and catalogues them. Beyond this, every page is indexed and associated with the forms. If the same form is found in distinct pages, all of them are indexed. Nevertheless, there will be only one representation of the form. 4.1.2 Generating queries for the forms Another component of the indexing software is in charge of generating values for the encountered forms. The generation of queries is based on the collected information about the form and its fields. 10 Search Results 1. CD: Kinks Face To Face 2. CD: Kinks Muswell Hillbillies 3. CD: Kinks Misfits ... Processing (Server) CDs CDs DVDs DVDs : : : Kinks CDs CDs DVDs DVDs : CDs CDs DVDs DVDs Media: Title: Artist: Kinks Figure 1: A form processing The first generated query is always the default, that is, the one which uses all the defined values in the HTML code of the form. Next, a pre-defined number k of other possible combinations is generated. To generate values for the text fields (which possess an infinite domain) a table that stores a list of values for a data category is consulted based on the field label. For every generated query, a further visit is scheduled for the crawler. The parameters (set of field names and values) are stored and a new item is added to the queue of URLs that must be visited by crawlers. 4.1.3 Visiting the results The crawler is in charge of executing its second big goal which is to submit the scheduled queries. To accomplish this it needs an extra feature: the capacity to send parameters in HTTP requests using both the GET and POST methods . If it perceives that an item in the queue of URLs is a query, it submits the parameters and analyses the HTML code obtained as a result in the same way that it does to others. The page is then compressed, stored and associated with the information of the original query. The indexing software decompress pages that contain results of form submissions and indexes them. From this index , the classification and search software finds results that contain all the search terms formulated by the user. 4.1.4 Searching for stored pages As soon as the user performs a search, two steps are exe-cuted by the classification and search software. Firstly, the indexes created by the indexing software are consulted to find the keywords formed by the user and the results are returned in an organised form (the most relevant come first) in an HTML page. Subsequently, based on these indexed pages which are associated with the forms, SmartCrawl selects forms that are more likely to answer to the user's search and generate new queries which will be visited afterwards by the crawlers and indexed in the same way by the indexing software. ARCHITECTURE AND IMPLEMENTA-TION The architecture at the high level of this application is divided into: crawler, indexing software, ranking and search software and storage components. As we have seen, the crawler is responsible for obtaining Web pages, submitting queries and storing the results. The indexing software, on the other hand, indexes the obtained pages and generates form queries. The ranking and search software uses the indexes to answer searches made by the user or redirects other forms and storage components take care of the storage of all the information used by other components . Figure 2 shows how the main components are available and how they interact with the storage components, that are represented by the rounded boxes. They are the Form Parser, Form Inquirer and Form Result Indexer of the indexing software, the Document Seeker of the ranking and search software and the Crawler Downloader of the crawler. Two storage components support the crawling: URL Queue and URL List. The first is responsible for storing the line of URLs that the Crawlers Downloaders need to visit, in this way, the URLs which will serve as seeds to the autonomous are also added in it. As soon as a Crawler Downloader extracts links from an HTML page, it is in this component that the new URLs will also be inserted for a further visit. The URL list, on the other hand, stores the URLs that have already been visited, allowing the Crawler Downloader keep track of them. When the page includes a form, or a result to a query, it needs to be stored for subsequent indexing. The crawler stores the compressed content in a storage component named Warehouse, where it is given a number called storeId. Two components of the indexing software are responsible for decompressing these pages that have been stored in the Warehouse. The first, called the Form Parser, extracts information from all the forms contained within the page, and sends them to the storage component Form List, where a number, called formId, is associated with every form. The Form Parser is also responsible for indexing the page which contains forms and associating it to each formId of the forms contained within. To index these documents, SmartCrawl uses a technique called inverted index or inverted file. The Document List, Wordmatch and Lexicon are the three components that carry out storing an indexed page. The Document List stores the title, a brief description of each indexed page and its docId. The Lexicon and Wordmatch store the inverted index itself. The first contains a list of pairs (wordId, word) for each one of the words used in indexed documents. The second contains a list of occurrences of the words in the indexed documents and their position (offset) in the text. Wordmatch is formed therefore by the values of docId, wordId and offset The Form Inquirer is the indexing software component whose objective is to generate queries for the stored forms in the Form List. To generate values for the text fields, Form Inquirer consults the list of categories and values through the component Categories. Each query generated is sent to the Query List where a queryId is associated to it and a new URL is added to the URL Queue. 11 Warehouse URL Queue URL List Query List Form List Categories Wordmatch Document List Lexicon Crawler Downloader Form Parser Form Result Indexer Document Seeker New query Form Inquirer Figure 2: Architecture at a high level The second component that extracts compressed pages from the Warehouse is the Form Result Indexer. Its job is too much easier than the others, as its aim is only to index the pages that contain results of submitted queries and to associate the proper queryId. From the indexes created and stored in Lexicon and Wordmatch , the Document Seeker answers to user searches. Every document stored in the in the Document List possesses a formId associated with it, and optionally, a queryId when the document is the response to a query. With these two numbers, the Document Seeker consults the Form List and the Query List to obtain the necessary information about the way it locates an indexed page on the World Wide Web. For example, a query to a form that points to the URL http://search.cnn.com/cnn/search using the HTTP GET method can be represented by http://search.cnn.com/cnn/search?q=brazil, if it possesses only one parameter with name q and value brazil. The Document Seeker should return the result set in an ordered form, so that the most relevant documents will be taking first place. A simple solution for this is to only take into consideration the position of the words in the text, and the number of occurrences. Considering o i , as the offset of the i-th encountered word in the document which is amongst the terms of the user, the relevance is given by this: r = n X i=0 1000 (o i + 1) (2) In equation 2, we compute the rank of a page by summing the offsets of all user search terms that appears in the document and then divide the result by an arbitrary number (in this case 1000) so that the most relevant entry receives the smallest number. Another important role of the Document Seeker is to redirect the search terms supplied by the user to some of the several forms catalogued in the Form List. To accomplish this, it looks in the Document List for pages that contain search engines (forms with one, and only one, text field) and that possess in its text words related to the terms sought by the user. New queries are then added to the Query List and new URLs are added to the URL Queue to be visited by the Crawler Downloader. 5.1 Labels extraction algorithm A very important task is performed by a secondary component called Form Extractor. It is in charge of extracting diverse pieces of form information present in an HTML page. To facilitate the content analysis of a page, the HTML code is converted into a DOM 1 tree provided by Cyberneko Html Parser [3]. From the DOM tree, the Form Extractor looks for nodes which represent forms and separate them from the rest of the tree. Each of these sub-trees, which encompass all the tags which are positioned between the &lt;form&gt; and the &lt;/form&gt;, is submitted for processing. Amongst the data which should be obtained, undoubtedly the fields and their labels are the most challenging ones. In spite of having a tag in the HTML specification called label for the declaration of a label, it is almost not used and, therefore, we do not possess a formal declaration of labels in the HTML code. The solution encountered was to establish a standard that the labels must have. For the Form Extractor, labels are continuous segments of texts which use the same format and have the maximum of n words and k characters. These values can be defined in the configuration file. 1 Document Object Model 12 DVDs CDs DVDs Media: Title: Artist: Kinks (a) HTML form example 1 2 1 2 3 4 { "Artist:"} Label { "Title:"} Label { "Media:"} Label { "Submit", "Reset"} Button Button { "artist"} Textfield { "title"} Textfield { "cds", "CDs", "dvds", "DVDs"} Checkbox Label Checkbox Label (b) Table representing the position of the elements of the form Figure 3: Representing the positions of the components of an HTML form CheckboxField Label: "Media:" Name: "media" Options: TextField Label: "Artist:" Name: "artist" Value: "" Form Action: "Search.jsp" Request Method: POST Fields: TextField Label: "Title:" Name: "title" Value: "" SubmitButtonField Name: "Submit" Value: "submit" Option Label: "CDs" Value: "cd" Option Label: "DVDs" Value: "dvd" Figure 4: An example of a HTML form representation From the sub-tree which contains a certain form information , the Form Extractor generates a table which represents the positioning of the elements contained in it. Figure 3 shows an example of the table generated by a simple form. The table is generated considering the nodes in the DOM tree which represent a common HTML table. If there are more than one defined table in the HTML code, similar representations are created. Each cell has a collection of the form's elements. The third step of the process is to extract the labels of each one of the fields in the form (except for hidden fields and buttons) and generate an object-orientated representation for the forms. To extract the labels, the Form Extractor passes twice by the generated table to the form. The first time, for each field in the form which require a label, it is verified exactly what exists on the left side of the field (even in one adjacent cell). If a label is found in this position, this label is immediately associated to the field. In the second passage, the fields which still do not have association with labels are observed again, however this time the search for the label is done in the above cell. For the fields of the checkbox and combobox kind, the treatment is special, because apart from the conventional label, the items which represent their domain also have labels . As in the case of "DVDs" and "CDs" labels in Figure 3(a). The domain labels are extracted from the right, and the label of the set of items is obtained from the left. In the example, the checkboxes are grouped by their names, which in this case is "media." The labels of each item are extracted from the right ("CDs" and "DVDs") and the label of the set of checkboxes is extracted from the left of the first field ("Media:"). Figure 4 shows how the object-oriented structure for the example above would be. 5.2 The list of categories and values The Categories component is in charged of controlling a list of categories and values which helps the Form Inquirer to generate values for the text fields. In order to guarantee better results, the name of the category is normalized before comparing to the field label. The normalization aims to: (1) remove punctuation (leaving just the words), (2) convert graphic signing and other special characters into simple ones and (3) remove stop words. Stop words is a concept given to the words which can be taken from sentences without changing its meaning, being largely used in normalization and some search engines even extinguish these words from their indexes. Great part of the Stop Words are prepositions, articles and auxiliary verbs, such as "the", "of" or "is." The list of categories is automatically built by the Form Parser. Once it finds a field with finite domain (e.g. comboboxes ), the values are extracted and added to the categories list associated to the field's label. When the same value is added more than once to a category , it gets more priority in relation to the others. This is obtained by using a number which means the relevance of this value in the category. It is based on this number that the set of values is put in order before it is repassed to the Query Inquirer. Therefore, the most relevant values are tested first in text fields. 13 5.3 Redirecting queries As mentioned before, associated to each form, there is a set of indexed pages where it has been found. These pages allow the Document Seeker to choose the forms to which the user's search terms will be redirected to. In order to choose amongst several forms, two steps are taken: (1) finding which words or sequences of words are related to the terms of the user's search and (2) looking for these words on the pages which have small search engines. To solve the first problem, we could use the stored index itself. However, the volume of indexed information is not so large as to provide a good set of words. It was used, therefore , the catalogue of a general purpose search engine called Gigablast 2 , since it implements data mining techniques and provides, for the searching terms, a list of words or sentences which frequently appear in the returning documents. From this set of words, the Document Seeker performs search on pages which have search engines looking for these terms, also putting them in order according to equation 2. For the first n selected search engines, the Document Seeker creates new queries and add them to the queue so that they can be submitted afterwards in the same way done by the Form Inquirer. The queries are created by filling in form's only text field with the searching terms of the user and the other fields with default values. Search form Results New queries Figure 5: Interface for the search of indexed documents 5.4 The searching interface A searching interface was build for the purpose of carrying out the tests. As shown in Figure 5, it is divided into three parts. The searching form allows the user to provide the terms of the search. Furthermore, it is possible to choose what the target of the search is: all the pages, only pages which have forms or pages containing the results of forms. The results are shown on a list of documents that contain all the searching terms offered by the user. On pages that contain a 2 http://www.gigablast.com/ query associated to them, it is possible to visualise also the parameters. In the area called new queries the generated queries which have been scheduled for the crawler's visit are displayed. EXPERIMENTAL RESULTS The tests which have been carried out aim to test some strategies used in the implementation of the prototype, hence some important aspects of the system were tested separately. Besides that, an analysis of the indexed content regarding its absence or not in the current searching engines was carried out. In order to support the tests, we started up the crawlers and kept them up until we have 15 thousands indexed pages (including only pages with HTML forms and form results). It worth repeat that our strategy does not index pages that do not offer any challenge to regular crawlers. 6.1 Label extraction algorithm evaluation This phase aims to evaluate the algorithm used for the extraction of the labels from the form fields that was de-scribed in section 5.1. To do so, 100 forms were manually observed and compared to the information extracted by the Form Extractor. For each one of these fields it was verified whether the choice made by the algorithm was correct or not. From the 100 forms evaluated, 5 of them (5%) were not extracted. The reason for this is that the HTML was malformed and the API used for the extraction of the DOM tree (NekoHtml [3]) did not manage to recover the error. This way, 189 fields from the 95 remaining forms were verified. For 167 of them (88%), the algorithm extracted the label correctly, making mistakes only in 22 labels (see Figure 6). Labels extracted (88%) correctly Labels extracted (12%) incorrectly Figure 6: Fraction of labels extracted correctly Some labels were not extracted correctly because they did not fit within the restrictions defined in section 5.1. Another problem faced was when the labels were not defined inside the tag FORM. In this case they were not present in the sub-tree analysed by the Form Extractor, making its extraction impossible. Although our solution did not reach the HiWE [7] accuracy to extract labels, we prove that it is possible to get very close results without rendering the page (that consumes much computing resources). Moreover, many of the problems faced in this experiment can be fixed without much effort. 14 6.2 Relevances of the queries generated by the Document Seeker In order to analyse the results obtained by the new generated queries from the searches of users, 80 search queries were submitted to the prototype by using arbitrary terms that are commonly used in general purpose search engines, such as "World Cup" or "Music Lyrics". For each list of queries generated by the Document Seeker, the first five ones were submitted and analysed manually, totalizing 155 pages with results. Each page was verified whether the query was successful or not. A successful query is one which has one or more results in it, in contrast to pages with no result or pages that was not considered a search engine results page (e.g. mailing list registration form). Errors (10%) Queries with no results (24%) Successful queries (66%) Figure 7: Utilizations of the new queries generated by Document Seeker The result obtained was that 66% of the submitted queries brought some results back, 24% were not successful and 10% of the pages, for some reason, could not be recovered. Figure 7 illustrates better the obtained utilization. Once the most relevant queries are returned taking first place, it is probable that, with a larger number of indexed pages, we will get better results. 6.3 Visibility of the indexed content The implementation of SmartCrawl has as aim the pages generated from the filling of the HTML forms and which are potentially part of the hidden web. Despite the fact that the current commercial search engines do not have an automatic mechanism which fills in the forms fields and obtain these data, as the SmartCrawl does, through common links, part of this information can be explored. For instance, a page with the results of a form that utilizes the HTTP method GET can be accessed through an usual link because all its parameters can be passed in the URL string itself. In addition to this, once the content is stored in databases, it is not so difficult to find pages that offer the same information through different interfaces. This phase aimed to verify how much of the indexed content can also be accessed by Google. In order to do this, 300 pages with results of GET and POST forms were observed if, for one of the reasons stated before, they were not indexed by this general purpose search engine. We found out that 62% of these pages are not indexed by Google. When only queries which use the method HTTP POST (which are 59% from the total) are observed, this number becomes even greater, leaving just 14% reachable by Google. CONCLUSIONS AND FUTURE WORK This work proposed a search engine prototype which is capable of handling with HTML forms as well as filling them in automatically in order to obtain information which is un-reachable by the current search engines. When compared to other solutions, as the HiWE [7] and the solution that redirects queries by Lin and Chen [6], the SmartCrawl brings a big differential which is the ability of surpassing great number of forms. The mentioned solutions have severe restrictions which directly affect the number of forms that receives queries. There is a great deal of work still to be attached to the solution for a better exploration of the recovered content. An example of this is that SmartCrawl does not make any analysis of the pages obtained as results of the queries, therefore indexing pages which contain errors and no results. The implementation of an algorithm which recognizes these pages would increase the quality of the indexed data. Besides, a high performance structure was not used for the storage of the indexes. This resulted in slow searching and indexing. A future work will be the implementation of a new indexing module. REFERENCES [1] M. K. Bergman. The Deep Web: Surfacing Hidden Value. 2001. [2] S. Brin and L. Page. The anatomy of a large-scale hypertextual Web search engine. Computer Networks and ISDN Systems, 30(17):107117, 1998. [3] A. Clark. Cyberneko html parser, 2004. http://www.apache.org/ andyc/. [4] J. Lage, A. Silva, P. Golgher, and A. Laender. Collecting hidden web pages for data extraction. In Proceedings of the 4th ACM International Workshop on Web Information and Data Management, 2002. [5] S. Liddle, D. Embley, D. Scott, and S. H. Yau. Extracting data behind web forms. In Proceedings of the Workshop on Conceptual Modeling Approaches for e-Business, pages 3849, 2002. [6] K.-I. Lin and H. Chen. Automatic information discovery from the invisible web. In Proceedings of the The International Conference on Information Technology: Coding and Computing (ITCC'02), pages 332337, 2002. [7] S. Raghavan and H. Garcia-Molina. Crawling the hidden web. In Proceedings of the 27th International Conference on Very Large Databases, pages 129138, 2001. [8] A. Rappoport. Checklist for search robot crawling and indexing, 2004. http://www.searchtools.com/robots/robot-checklist .html. [9] C. Sherman and G. Price. The Invisible Web: Uncovering Information Sources Search Engines Can't See. CyberAge Books, 2001. [10] V. Shkapenyuk and T. Suel. Design and implementation of a high-performance distributed web crawler. In Proceedings of the 18th International Conference on Data Engineering, pages 357368. [11] D. Sullivan. Internet Top Information Resource, Study Finds, 2001. 15

implementation;architecture;Label Extraction;experimentation;html form;SmartCrawl;web crawler;hidden web content;information retrieval;search engine;Search Engine;extraction algorithm;Hidden Web

182

Sparsha: A Comprehensive Indian Language Toolset for the Blind

Braille and audio feedback based systems have vastly improved the lives of the visually impaired across a wide majority of the globe. However, more than 13 million visually impaired people in the Indian sub-continent could not benefit much from such systems. This was primarily due to the difference in the technology required for Indian languages compared to those corresponding to other popular languages of the world. In this paper, we describe the Sparsha toolset. The contribution made by this research has enabled the visually impaired to read and write in Indian vernaculars with the help of a computer.

INTRODUCTION The advent of computer systems has opened up many avenues for the visually impaired. They have benefited immensely from computer based systems like automatic text-to-Braille translation systems and audio feedback based virtual environments. Automatic text-to-Braille translation systems are widely available for languages like English, French, Spanish, Portuguese, and Swedish [7, 26, 18, 16]. Similarly audio feedback based interfaces like screen readers are available for English and other languages [ref c, 8, 20]. These technologies have enabled the visually impaired to communicate effectively with other sighted people and also harness the power of the Internet. However, most of these technologies remained unusable to the large visually impaired population in the Indian sub-continent [17]. This crisis can be attributed to primarily two reasons. First, the languages in the mentioned region differ widely from other popular languages in the world, like English. These languages or vernaculars also use relatively complex scripts for writing. Hence, the technologies used for English and other such languages cannot be easily extended to these languages. Secondly, the development of these technologies for Indian languages, right from scratch, is not trivial as the various Indian languages also differ significantly amongst themselves. The Sparsha toolset uses a number of innovative techniques to overcome the above mentioned challenges and provides a unified framework for a large number of popular Indian languages. Each of the tools of Sparsha will be discussed in detail in the following sections. Apart from English the languages supported by Sparsha include Hindi, Bengali, Assamese, Marathi, Gujarati, Oriya, Telugu and Kannada. The motivation for this work is to enable the visually impaired to read and write in all Indian languages. The toolset set has been named Sparsha since the word "Sparsha" means "touch" in Hindi, something which is closely associated with how Braille is read. BHARATI BRAILLE TRANSLITERATION Bharati Braille is a standard for writing text in Indian languages using the six dot format of Braille. It uses a single script to represent all Indian languages. This is done by assigning the same Braille cell to characters in different languages that are phonetically equivalent. In other words, the same combination of dots in a cell may represent different characters in each of the different Indian languages. However, a single character in an Indian language may be represented by more than one Braille cell. The above mentioned characteristics of Bharati Braille code is illustrated in Figure 1. There are many other issues and rules related to Bharati Braille. These will be discussed in the following sections along with the methods used for implementing them. Figure 1. Examples of characters in Indian languages and their corresponding Bharati Braille representation 2.1 Transliteration to Bharati Braille As shown in Figure 1, characters from different Indian languages can be mapped to the same Braille representation. Thus, in order to implement this, the system uses separate code tables for each of the languages and depending on the users choice of input language the corresponding code table is used. The said method of implementation also makes the system highly scalable and allows the inclusion of more languages in future if required. For instance this technique is being used successfully to extend the system to include Urdu and Sinhala. This work is expected to be completed in the near future. Figure 2. Formation of Conjugates Another important aspect of Indian languages is the formation of consonant clusters or conjugates. In traditional hand written text this may be expressed conceptually as the first consonant followed by a special character called halanth which in turn is followed by the second character. The consonant cluster may again be followed by a vowel. However, the visual representation of such a consonant cluster or conjugate may be quite different from the visual representation of each of the individual consonants included in it, as shown in Figure 2. However, while translating the same text into Bharati Braille the special character halanth must precede both the consonants to be combined into a single conjugate. The above constraints necessarily mean that the Braille translation for a particular character also depends on the sequence of characters preceding and following it. Hence, in order to perform the tasks efficiently the system uses a finite state machine based approach similar to that of lexical analyzers [3, 6]. The mentioned approach also proves to be suitable for handling other issues associated with standard Braille translation like detection of opening and closing quotation marks, string of uppercase characters. Apart from Indian languages the Sparsha system supports the translation of English language texts into grade 1 and grade 2 Braille. The system maintains a database of all standard Braille contractions which is used for generating grade 2 Braille. Furthermore, the system allows the user to add new contractions to the existing database. The Sparsha system also supports the proper translation of a document containing text both in English as well as an Indian Language. According to standard Braille notations [11] the change in language is indicated through the proper use of the letter sign. However, a single document containing text in more than one Indian language cannot be translated into Braille, such that the reader is able to distinguish each of the languages correctly. This is due to the following reason. As mentioned previously the same Braille representation can refer to different characters in different Indian languages, this leads to the inherent ambiguity. Figure 3. A screenshot of the interface for translating and editing Braille in the Sparsha system 2.2 Reverse transliteration The Sparsha system allows reverse transliteration of Braille to text both for Indian languages as well as English. This allows the visually impaired to communicate seamlessly with other sighted people. The Braille code to be translated may be entered into the computer using a standard six key Braille keyboard. After translating the Braille code into text, the visually readable text may then be checked for correctness using a file reading system which will be described in later section. In order to achieve reverse translation from Braille to text, the system uses a finite state machine based approach similar to that used for translating text to Braille as described previously. The task of reverse translation also uses the code tables corresponding to the language to which the text is being translated. Thus the system can easily be extended to other languages just by adding the corresponding code tables to achieve both forward and reverse translation. 2.3 Methods of Input - Output Sparsha can accept English text, for translation, in the form of plain text files, HTML (hyper text markup language) files and Microsoft Word documents. Apart from English the Sparsha Braille translation system, as described, can take input text in Indian languages. This input can be given to the system in a number of forms as follows: ISCII (Indian Script Code for Information Interchange) [24] documents generated by applications like iLeap [10] LP2 documents generated by iLeap [10] Unicode text generated by any standard editor supporting Unicode [25]. This technique will be discussed in detail in a later section The output of the Braille translation can be obtained on a large variety of commercial Braille embossers [23]. The Sparsha system has been tested on the following Braille embossers: Index Basic-S Index Basic-D Index 4X4 PRO + + + = 115 Braillo 400 Modified Perkins Brailler [15] Alternatively the output may be obtained on tactile Braille displays [1]. BRAILLE MATHEMATICS At the time of this development there existed a few translators for converting mathematics to Braille [b]. However, these were found to be unsuitable for the visually impaired in the Indian sub-continent due to a number of reasons. Firstly the Braille code used for mathematics in India is slightly different from those used in other parts of the world [4], however, it bears close resemblance to the Nemeth code [5]. Secondly the interleaving of Braille mathematics with text in Indian languages was also not possible with the available systems. Thirdly many of these systems require a working knowledge of LaTex [13]. This cannot be expected from every user. Finally, most of these systems are unaffordable to the visually impaired in the Indian sub-continent. The above mentioned reasons warranted the development of a mathematics-to-Braille translation system for the Indian subcontinent . The system thus developed can translate almost all mathematic and scientific notations. It also allows the user to interleave mathematic and scientific expressions with text in both Indian languages and English. In order to allow the user to write complex mathematic and scientific expressions, the system provides a special editor for the purpose. The above mentioned editor is named "Nemeth editor" after Abraham Nemeth [t]. Thus the user is exempted from the task of learning LaTex. The editor provides a GUI (Graphic User Interface) as shown in Figure 4 for writing a mathematic or scientific expression in a form similar to that used by LaTex. This string can then be readily converted into Braille by the translation engine. However, the mathematical expression formed by the editor must be enclosed within a pair of special character sequences. This needs to be done so that when the mathematic or scientific expression is embedded within another English or Indian language text, it is properly translated to Braille using the standard for mathematic and scientific notation. Figure 4. Screenshot of the Nemeth Editor The selection of mathematical symbols and notations is done by the user in a menu driven fashion using the GUI. The set of all mathematic and scientific notations is partitioned in to separate collections, each consisting of similar notations. Alternatively the text may be entered by the user in a LaTex like format using any standard text editor. SPARSHA CHITRA Elementary tactile graphics is one of the best methods for introducing certain subjects, like geometry, to visually impaired students. However, such tactile graphics have remained outside the reach of the common man. This is due to the fact that sophisticated Braille embossers and expensive image conversion software are necessary for the purpose. Sparsha Chitra aims to provide relatively simple tactile graphics which can be obtained even by using low cost Braille embossers like the modified Perkins Brailler [15]. In other words no assumptions have been about any special feature of the Braille embosser being used. This allows tactile graphics to be embossed using just the Braille embossing capability of the embosser. The tactile graphics obtained from any image may be viewed and edited before finally being embossed. The image may also be scaled up or down to a size suitable for embossing. The system also allows the image color to be inverted in order to improve the contrast. Sparsha Chitra takes its input in HTML format such that additional text can be included along with the tactile representation of the image. Sparsha is the feeling of touch and "Chitra" in Hindi means "picture" and thus this tool is named Sparsha Chitra. The primary limitation of this tool is that complex images cannot be represented very clearly. However, the effect of this drawback is mitigated by the fact that the amount of detail that can be observed through touch is also limited. Furthermore the size of the tactile image is restricted by the bounds imposed by the sheet on which it is embossed. The functions for scaling the tactile image may prove to be useful in such a case. Figure 5. Screenshot of Spasha Chitra FILE READER In order to enter text into the computer, in English, a visually impaired user can take the help of any standard screen reader [12, 8, 20 ]. Screen readers have proved to be vital to visually impaired computer users [19].Such screen readers are commercially available. However, such screen readers are not available for Indian languages. 116 This was primarily due to the reasons mentioned at the beginning of this paper. The file reader which will be described in this section will redeem the situation and allow the user to type in text in Indian languages using Microsoft Word. For performing other tasks related to the operating system the user may use any of the standard screen readers. The construction of such a file reader requires a number of vital components [2]. These include text-to-speech engines for Indian languages, fonts for Indian languages, keyboard layouts for them, proper rendering engines and a text editor which can support Indian languages. Each of these components will be described briefly in the following sections. This will be followed by a description of the overall architecture of the system and its functioning. 5.2 Speech synthesis system A speech synthesis system is vital for the functioning of any screen reader. It is responsible for producing human voice rendition of the text provided to it by the screen reader. In case of screen readers the speech synthesis system should be able to deliver the voice in real-time . This is necessary for visually impaired users to get instantaneous audio feedback. A multilingual screen reader necessarily needs a speech synthesis system for each of the languages that it supports. The mentioned file reader uses a speech synthesis engine for Indian languages called Shruti [22]. Shruti support two popular Indian languages namely Hindi and Bengali. It uses a method of di-phone concatenation for speech synthesis. This allows the speech synthesis system to produce reasonable real-time performance, at the same time maintaining a low memory space requirement. 5.3 Fonts and Rendering There are number issues involved with Indian language fonts and their rendering. This is due to the fact that Indian language scripts are generally complex in nature. The Microsoft Windows system can be configure for correctly rendering these complex Indian language scripts. Correct rendering of fonts is achieved through the use of Uniscribe (Unicode Script Processor) and OTLS (OpenType Layout Services) libraries [9, 14]. Furthermore glyph substitution and glyph repositioning, as shown in Figure 2, are closely associated with the rendering of text in Indian languages. For this reason OpenType fonts have been found to be suitable for Indian languages as they carry, within the font file, explicit information about glyph substitution and glyph positioning. This maintained in the form of two tables namely GSUB (Glyph Substitution) and GPOS (Glyph Positioning). 5.4 Editor for Indian Languages A number of text editor are available for Indian languages. Many of these editors are difficult to use and are non-intuitive. On the other hand it has been observed that Microsoft Word XP (Word 2002) performs reasonably well for Indian languages when proper fonts and rendering engines are used. Thus, Microsoft Word has been used instead of creating a new editor for the file reading application as shown in Figure 7. Microsoft Word also provides certain additional features which have been used extensively for the development of the file reader. These features have been discussed in detail in the following paragraphs. The use of Microsoft Word also motivates visually impaired users to switch to main stream applications and also eliminates the effort of learning another system. Microsoft Word supports Unicode [25], hence it can accept text in any Indian language. However, in order to enter text in an Indian language in the Windows system a keyboard layout or IME (Input Method Editor) [21] for that language is required. Keyboard layouts are available for some popular Indian languages like Hindi. For other Indian languages it may have to be created. In our case a keyboard layout had to be created for Bengali. Figure 6. File reader System Architecture 117 The capabilities of the Microsoft Word can be extended using COM (Component Object Model) Add-Ins. Such Add-Ins are basically programs that run within the framework provided by Microsoft Word. The file reader has been developed in the form of such an Add-In. It interacts closely with the editor to provide necessary audio feedback for text in Indian languages. Such interaction takes place through the object model exposed by Microsoft Word. The file reader may be configured to start up every time Microsoft Word is used. 5.5 Overall System Structure and Operation The overall architectural structure of the file reader system is shown in Figure 6. Most of the components of the system shown in the figure have been discussed in the last few sections. The interaction between the different components and how they operate as a system will be discussed in this section. Keyboard hooks are placed within the operating system by the file reader Add-In. The keyboard hooks are responsible for trapping the keystrokes entered by the user through the keyboard. A copy of the entered keystrokes is passed to the file reader Add-In. The keystrokes are then passed to the keyboard layout or IME which is integrated with the operating system. The keyboard layout translates the keystrokes into Unicode characters and passes them to the editor. Figure 7. Screenshot of the file reader in operation using Microsoft Word In the mean while the file-reader Add-In, on receiving the keystrokes, provides appropriate audio feedback by invoking the speech synthesis engine. Again, certain combinations of keystrokes are recognized by the file reader Add-In as special commands. These request the file reader to read out a certain portions of the text. This selected text is then passed to the speech synthesis system for producing human voice rendition. Thus providing an audio feedback based virtual environment for Indian languages. The file reader can be further extended to provide full screen reading functionality by using Microsoft Active Accessibility. SYSTEM EVALUATION A subset of the Sparsha system known as the Bharati Braille Transliteration System * has been deployed by Webel Mediatronics Limited in a number of organizations for the visually impaired all over India as a part of a project sponsored by the Ministry of Communication and Information Technology, Government of India. As a result of these field tests the system underwent an iterative process of refinement to reach its current form. A plethora of request and suggestions from visually impaired users led to the development and inclusion of a number of additional features and tools that were added to the toolset. These include the Sparsha Chitra and the file readers for Indian languages. The process of continuous feedback helped the Sparsha toolset mature over the years. It also helped in weeding out many bugs and shortcomings of the initial versions of the system. 6.2 Obtained Results The Sparsha system is under a continuing process of use and evaluation. This feedback is being used to make the system more usable to the visually impaired and to enhance the features provided by the system. Training and deployment of the system has also been carried out at a number of premier organizations for the visually impaired. These include The National Association for the Blind, Delhi Blind Peoples' Association, Ahmedabad National Institute for the Visually Handicapped, Dehradun The Braille translation system has been tested on a large number of computers in these organizations. The typical performance characteristics of the Sparsha Braille translation system is as shown in Figure 8. The performance characteristics have been measured for two different personal computer systems. English Grade - II Braille 0 1000 2000 3000 4000 5000 6000 7000 200 600 1000 1400 1800 2200 2600 3000 3400 3800 5000 Word Count Ti m e ( m i l l i s e c on ds ) P4 P3 (a) Grade - I Braille 0 100 200 300 400 500 200 600 1000 1400 1800 2200 2600 3000 3400 3800 5000 Word Count T i m e ( m illis e c onds ) P4 - English P4 - Indian Languages P3 - English P3 - Indian Languages (b) Figure 8. Graphs showing the computation time taken during Braille translation for (a) Grade II English (b) Grade I English and Indian languages 118 The computers have the following specifications (Processor Type, Primary Memory, Hard disk): Intel Pentiun 4 3GHz, 512MB, 80GB referred to as P4 Intel Pentium III 550MHz, 256MB, 40GB referred to as P3 The Sparsha Chitra tool was tested by visually impaired users and the obtained results are given in Figure 9. This was done by handing them sheets of paper Braille paper with tactile diagrams created by Sparsha Chitra and asking them to guess the image on the sheets. Correct guess 40% Cannot guess 20% Close guess 35% Wrong guess 5% Figure 9. User response to tactile images generated by Sparsha Chitra Most of these images were geometric figures. The majority of the guesses were correct while a large percentage of the guesses were very close like identifying a rectangle as square or a triangle as a mountain. Such misinterpretations often occur due to lack of color information or misjudging dimensions which is indeed quite difficult estimate from tactile representations. The file reader tool needs extensive training before a nave user can use it efficiently. Visually impaired users who are already familiar with Jaws or other screen readers can adapt to this system very quickly. This tool was primarily tested by visually impaired users having reasonable experience with Jaws. 0 5 10 15 20 25 30 Jaws File reader for Indian Languages W o rd s p e r M i n u t e Figure 10. Comparison of the typing speed of a visually impaired user using Jaws and the Indian language file reader The Indian language file reader could not be experimented with a large number of users since a good level of expertise with screen readers is required for using the file reader efficiently. The experimental results shown in Figure 10 and 11 pertain to a particular visually impaired user having some experience with Jaws. The experiments were carried out by dictating a paragraph of about hundred words to the user while he typed it into the computer using the Indian language file reader. However, this is only a preliminary experiment. It was also found that both the typing speed and the error rates improved significantly with practice. No Errors 10% One Error 30% Two Errors 40% More than 2 Errors 20% Figure 11. Number of words with errors for every ten words LIMITATIONS AND FUTURE WORK The Sparsha toolset is in the process of being extended to a number of other languages in the Indian subcontinent. This includes Urdu and Sinhala. The file reader in the Sparsha toolset is also limited by the availability of text to speech synthesis engines for all Indian languages. Hopefully these will be available in the near future and allow the system to be extended to more languages. It is also envisioned that the Sparsha system will be ported to mobile handheld systems. This will enable the visually impaired to communicate on the move. As of now, the required text to speech synthesis engines have been ported onto the Microsoft Pocket PC platform as well as on an ARM-Linux platform. It is only a matter of time before the file reader becomes functional on such mobile platforms. CONCLUSION The Sparsha system named after the feeling of touch has been the first attempt to help visually impaired users, in the Indian subcontinent, read and write in their native tongues. In this paper the various aspects of Indian languages and how they differ from other languages in the world has been explained. It also been discussed how these issues have been tackled in the Sparsha system. The paper describes in depth the various tools included in the Sparsha toolset. These tools form a comprehensive toolset for Indian languages. It can be hoped that the Sparsha system would help increase the literacy rates among the 13 million visually impaired in the Indian subcontinent. ACKNOWLEDGMENTS The authors would like to thank Media Lab Asia for sponsoring a part of the work related to the file reader. The authors would also like to thank the National Association for the Blind, Delhi and many other organizations for the blind for their sustained help and cooperation during the entire development process. The authors owe special thanks to Mr. Samit Patra, Director, Electrosoft Consultants for his enormous help with many technical aspects of the work. 119 REFERENCES [1] Basu Anupam, Roy S., Dutta P. and Banerjee S., "A PC Based Multi-user Braille.Reading System for the Blind Libraries", IEEE Transactions on Rehabilitation Engineering, Vol. 6, No. 1, March 1998, pp.60--68 [2] Blenkhorn, P. "Requirements for Screen Access Software using Synthetic Speech". Journal of Microcomputer Applications, 16, 243-248, 1993. [3] Blenkhorn Paul, "A System for Converting Braille to Print", IEEE Transactions on Rehabilitation Engineering, Vol. 3, No. 2, June 1995, pp. 215-221 [4] Braille Mathematics Code for India Manual, Prepared under the project "Adoption and Introduction of an Appropriate Braille Mathematics Code for India", sponsored by UNICEF, Published by National Institute for Visually Handicapped, Dehra Dun and National Association for the Blind, Bombay, India [5] Cranmer T. V. and Abraham Nemeth, A Uniform Braille Code, memo to the members of the BANA Board (January 15, 1991); Available at: http://www.nfb.org or http://world.std.com/~iceb/ [6] Das, P.K.; Das, R.; Chaudhuri, A., "A computerised Braille transcriptor for the visually handicapped". Engineering in Medicine and Biology Society, 1995 and 14th Conference of the Biomedical Engineering Society of India. An International Meeting, Proceedings of the First Regional Conference, IEEE. 15-18 Feb. 1995 Page(s):3/7 - 3/8 [7] Duxbury Braille Translator, 2000, http://www.duxburysystems.com/products.asp [8] HAL. Dolphin Computer Access, http://www.dolphinuk.co.uk/products/hal.htm [9] Hudson, John for Microsoft Typography, "Windows Glyph Processing : an Open Type Primer", November 2000, http://www.microsoft.com/typography/glyph%20processing/i ntro.mspx [10] iLeap. Centre for Development of Advanced Computing. http://www.cdacindia.com/html/gist/products/ileap.asp [11] International Council on English Braille (ICEB), Unified English Braille Code (UEBC) Research Project, http://www.iceb.org/ubc.html [12] JAWS for Window. Freedom Scientific. http://www.freedomscientific.com/fs_products/software_jaws. asp [13] Lamport L. LaTeX - A Document Preparation System, Addison-Wesley, 1985, ISBN 0-201-15790-X. [14] Microsoft Typography, "Specifications : overview", http://www.microsoft.com/typography/SpecificationsOvervie w.mspx [15] Modified Perkins Brailler, Webel Mediatronics Limited. http://www.braille-aids.com/emboss.htm [16] MONTY, VisuAide. http://www.visuaide.com/monty.html [17] National Association for the Blind, India, 2002. Available at http://www.nabindia.org/sited/infor06.htm [18] NFBTRANS. National Federation of the Blind, 2004, http://www.nfb.org/nfbtrans.htm [19] Pennington C.A. and McCoy K.F., Providing Intelligent Language Feedback or Augmentative Communication Users, Springer-Verlag, 1998. [20] Raman T.V. (1996). "Emacspeak a speech interface". Proceedings of CHI96, April 1996 [21] Rolfe, Russ "What is an IME (Input Method Editor) and how do I use it?" http://www.microsoft.com/globaldev/handson [22] Shruti, Media Lab Asia Research Laboratory, Indian Institute of Technology, Kharagpur. http://www.mla.iitkgp.ernet.in/projects/shruti.html [23] Taylor Anne, "Choosing your Braille Embosser", Braille Monitor, October 200. Available at http://www.nfb.org/bm/bm01/bm0110/bm011007.htm [24] Technology Development for Indian Languages, Department of Information Technology, Ministry of Communication & Information Technology, Government of India. Available at http://tdil.mit.gov.in/standards.htm [25] Unicode. http://www.unicode.org [26] WinBraille. Index Braille. http://www.braille.se/downloads/winbraille.htm 120

audio feedback;Indian languages;Braille;Visual impairment

183

StyleCam: Interactive Stylized 3D Navigation using Integrated Spatial & Temporal Controls

This paper describes StyleCam, an approach for authoring 3D viewing experiences that incorporate stylistic elements that are not available in typical 3D viewers. A key aspect of StyleCam is that it allows the author to significantly tailor what the user sees and when they see it. The resulting viewing experience can approach the visual richness and pacing of highly authored visual content such as television commercials or feature films. At the same time, StyleCam allows for a satisfying level of interactivity while avoiding the problems inherent in using unconstrained camera models. The main components of StyleCam are camera surfaces which spatially constrain the viewing camera; animation clips that allow for visually appealing transitions between different camera surfaces; and a simple, unified, interaction technique that permits the user to seamlessly and continuously move between spatial-control of the camera and temporal-control of the animated transitions. Further, the user's focus of attention is always kept on the content, and not on extraneous interface widgets. In addition to describing the conceptual model of StyleCam, its current implementation, and an example authored experience, we also present the results of an evaluation involving real users.

INTRODUCTION Computer graphics has reached the stage where 3D models can be created and rendered, often in real time on commodity hardware, at a fidelity that is almost indistinguishable from the real thing. As such, it should be feasible at the consumer level to use 3D models rather than 2D images to represent or showcase various physical artifacts. Indeed, as an example, many product manufacturers' websites are beginning to supply not only professionally produced 2D images of their products, but also ways to view their products in 3D. Unfortunately, the visual and interactive experience provided by these 3D viewers currently fall short of the slick, professionally produced 2D images of the same items. For example, the quality of 2D imagery in an automobile's sales brochure typically provides a richer and more compelling presentation of that automobile to the user than the interactive 3D experiences provided on the manufacturer's website. If these 3D viewers are to replace, or at the very least be at par with, the 2D imagery, eliminating this r, viewpoint in the scene difference in quality is critical. The reasons for the poor quality of these 3D viewers fall roughly into two categories. First, 2D imagery is usually produced by professional artists and photographers who are skilled at using this well-established artform to convey information, feelings, or experiences, whereas creators of 3D models do not necessarily have the same established skills and are working in an evolving medium. Howeve this problem will work itself out as the medium matures. The second issue is more troublesome. In creating 2D images a photographer can carefully control most of the elements that make up the shot including lighting and viewpoint, in an attempt to ensure that a viewer receives the intended message. In contrast, 3D viewers typically allow the user to interactively move their to view any part of the 3D model. Figure 1. StyleCam authored elements This results in a host of problems: a user may "get lost" in the scene, view the model from awkward angles that present it in poor light, miss seeing important features, experience frustration at controlling their navigation, etc. As such, given that the author of the 3D model does not have control over all aspects of what the user eventually sees, they cannot ensure that 3D viewing conveys the intended messages. In the worse case, the problems in 3D viewing produce an experience completely opposite to the authors intentions! The goal of our present research is to develop a system, which we call StyleCam (Figure 1), where users viewing 3D models can be guaranteed a certain level of quality in terms of their visual and interactive experience. Further, we intend that the system should not only avoid the problems suggested earlier, but also have the capability to make the interactive experience adhere to particular visual styles. For example, with StyleCam one should be able to produce an interactive viewing experience for a 3D model of an automobile "in the style of" the television commercial for that same automobile. Ultimately, a high-level goal of our research is to produce interactive 3D viewing experiences where, to use an old saying from the film industry, "every frame is a Rembrandt". 1.1. Author vs. User Control Central to our research is differentiating between the concept of authoring an interactive 3D experience versus authoring a 3D model which the user subsequently views using general controls. If we look at the case of a typical 3D viewer on the web, in terms of interaction, the original author of the 3D scene is limited to providing somewhat standard camera controls such as pan, tumble and zoom. Essentially, control of the viewpoint is left up to the user and the author has limited influence on the overall experience. From an author's perspective this is a significant imbalance. If we view an interactive experience by cinematic standards, an author (or director) of a movie has control over several major elements: content/art direction, shading/lighting, viewpoint, and pacing. It is these elements that determine the overall visual style of a movie. However, in the interactive experience provided by current 3D viewers, by placing control of the viewpoint completely in the hands of the user, the author has surrendered control of two major elements of visual style: viewpoint and pacing. Thus we desire a method for creating 3D interactive experiences where an author can not only determine the content and shading but also the viewpoints and pacing. However, intrinsic in any interactive system is some degree of user control and therefore, more accurately, our desire is to allow the author to have methods to significantly influence the viewpoints and pacing in order to create particular visual styles. Thus, we hope to strike a better balance between author and user control. In order to achieve this end, StyleCam incorporates an innovative interaction technique that seamlessly integrates spatial camera control with the temporal control of animation playback. CONCEPTUAL MODEL In order to provide author control or influence over viewpoints and pacing, we need a way for an author to express the viewpoints and the types of pacing they are interested in. Thus we have developed three main elements upon which our StyleCam approach is based. 1. Camera surfaces an author-created surface used to constrain the users' movement of the viewpoint 2. Animation clips an author-created set of visual sequences and effects whose playback may be controlled by the user. These can include: sophisticated camera movements. Slates 2D media such as images, movies, documents, or web pages. visual effects such as fades, wipes, and edits. animation of elements in the scene. 3. Unified UI technique The user utilizes a single method of interaction (dragging) to control the viewpoint, animation clips, and the transitions between camera surfaces. 2.1. Camera Surfaces In the motion picture industry a money-shot is a shot with a particular viewpoint that a director has deemed "important" in portraying a story or in setting the visual style of a movie. Similarly, in advertising, money-shots are those which are the most effective in conveying the intended message. We borrow these concepts of a money-shot for our StyleCam system. Our money-shots are viewpoints that an author can use to broadly determine what a user will see. Further, we use the concept of a camera surface as introduced by Hanson and Wernert [19, 36] . When on a camera surface, the virtual camera's spatial movement is constrained to that surface. Further, each camera surface is defined such that they incorporate a single money-shot. Figure 2 illustrates this notion. Camera surfaces can be used for various purposes. A small camera surface can be thought of as an enhanced money-shot where the user is allowed to move their viewpoint a bit in order to get a sense of the 3-dimensionality of what they are looking at. Alternatively, the shape of the surface could be used to provide some dramatic camera movements, for example, sweeping across the front grill of a car. The key idea is that camera surfaces allow authors to conceptualize, visualize, and express particular ranges of viewpoints they deem important. Intrinsic in our authored interactions is the notion that multiple camera surfaces can be used to capture multiple money-shots. Thus authors have the ability to influence a user's viewpoint broadly, by adding different camera surfaces, or locally by adjusting the shape of a camera 102 Volume 4, Issue 2 surface to allow a user to navigate through a range of viewpoints which are similar to a single particular money-shot . For example, as shown in Figure 2, camera surfaces at the front and rear of the car provide two authored viewpoints of these parts of the car in which a user can "move around a bit" to get a better sense of the shape of the front grille and rear tail design. Figure 2. Camera surfaces. The active camera is at the money-shot viewpoint on the first camera surface. The rate at which a user moves around on a camera surface (Control-Display gain) can dramatically affect the style of the experience. In order to allow an author some control over visual pacing, we provide the author with the ability to control the rate at which dragging the mouse changes the camera position as it moves across a camera surface. The intention is that increasing/decreasing this gain ratio results in slower/faster camera movement and this will influence how fast a user moves in the scene, which contributes to a sense of pacing and visual style. For example, if small mouse movements cause large changes in viewpoint this may produce a feeling of fast action while large mouse movement and slow changes in movement produce a slow, flowing quality. Figure 3 illustrates an example of variable control-display gain, where the gain increases as the camera gets closer to the right edge of the camera surface. Figure 3. Variable control-display gain on a camera surface 2.2. Animation Clips To support transitions between two camera surfaces, we use animation clips as illustrated in Figure 4. An animation clip can be thought of as a "path" between the edges of camera surfaces. When a user navigates to the edge of a camera surface, this triggers an animation. When the animation ends, they resume navigating at the destination camera surface. One obvious type of animation between the camera surfaces would simply be an automatic interpolation of the camera moving from its start location on the first camera surface to its end location on the second camera surface (Figure 4a). This is similar to what systems such as VRML do. While our system supports these automatic interpolated animations, we also allow for authored, stylized, animations. These authored animations can be any visual sequence and pacing, and are therefore opportunities for introducing visual style. For example, in transitioning from one side of the car to the other, the author may create a stylized camera animation which pans across the front of the car, while closing in on a styling detail like a front grille emblem (Figure 4b). The generality of using animation clips allows the author the stylistic freedom of completely abandoning the camera-movement metaphor for transitions between surfaces and expressing other types of visual sequences. Thus animation clips are effective mechanisms for introducing slates -- 2D visuals which are not part of the 3D scene but are momentarily placed in front of the viewing camera as it moves from one camera surface to another (Figure 4c). For example, moving from a view of the front of the car to the back of the car may be accomplished using a 2D image showing the name of the car. This mechanism allows the use of visual elements commonly found in advertising such as real action video clips and rich 2D imagery. In the computer realm, slates may also contain elements such as documents or webpages. Figure 4. Three example animated transitions between camera surfaces. (a) automatic transition, (b) authored stylized transition, (c) slate transition. The use of animation clips also allows for typical visual transitions effects such as cross fades, wipes etc. In addition to using animation clips for transitions between camera surfaces, StyleCam also supports the animation of elements in the 3D scene. These scene element animations can occur separately or concurrently with transition animations. For example, while the animation clip for the visual transition may have the camera sweeping down the side of the car, an auxiliary animation may open the trunk to reveal cargo space. Volume 4, Issue 2 103 The animation of scene elements can also be used to affect extremely broad changes. For example, entire scene transitions (similar to level changes in video games) may occur when a user hits the edge of particular camera surface. At the author's discretion, temporal control of animation clips can either be under user control or uninterruptable. Overall, in terms of visual expression, these varying types of animation clips allow an author to provide rich visual experiences and therefore significantly influence the pacing and style of a user's interaction. 2.3. Unified User Interaction Technique While animation clips are effective for providing a means to move between camera surfaces and introduce visual styling elements, they also highlight the fundamental issue of arbitrating between user control and system control. At the heart of our system are two distinct types of behavior: 1) user control of the viewpoint, and 2) playback of animation clips. In other systems these two types of behavior are treated as distinct interactions. Specifically, the user must stop dragging the camera viewpoint, then click on something in the interface to trigger the animation, dividing their attention and interrupting the visual flow. In our system we wanted to use animations as a seamless way of facilitating movement between camera surfaces. Thus we needed a mechanism for engaging these animations that did not require an explicit mouse click to trigger animation. Ideally we wanted to leave the user with the impression that they "dragged" from one camera surface to another even though the transition between the surfaces was implemented as an authored animation. These two behaviors are fundamentally different in that viewpoint control is spatial navigation and animation control is temporal navigation. From a user interaction standpoint, spatial behavior can be thought of as "dragging the camera" while temporal control is "dragging a time slider" or "scrubbing". Given this we required an interaction model which allowed these two types of drags to be combined together in a way that was well defined, controllable, and corresponded to user's expectations. Figure 5, which uses the finite-state-machine model to describe interaction as introduced by [5, 26] , shows the interaction model we developed. The key feature of this model is the ability to transition back and forth from spatial to temporal control during a contiguous drag. As a user drags the camera across a camera surface (State 1, Spatial Navigation) and hits the edge of the surface, a transition is made to dragging an invisible time slider (State 2, Temporal Navigation). As the user continues to drag, the drag controls the location in the animation clip, assuming that the author has specified the clip to be under user control. Upon reaching the end of the animation, a transition is made back to dragging the camera, however, on a different, destination camera surface (State 1). Button Up Clip Finished Button Up State 0 State 1 State 2 State 3 Button Down Enter Surface Exit Surface Button Down Tracking Dragging in Space Dragging in Time Tracking during Automatic Playback Stop Playback Spatial Navigation Temporal Navigation Figure 5. StyleCam interaction model. The interaction model also handles a variety of reasonable variations on this type of dragging behavior. A user may stop moving when dragging an animation clip, thus pausing the animation. If, however, when in State 2 the user releases the mouse button during a drag, automatic playback is invoked to carry the user to the next camera surface (State 3). Should the user press the mouse button during this automatic playback, playback is stopped and temporal control by the user is resumed (return to State 2). We found in practice that this interaction design enhanced the user's feeling of being in control throughout the entire experience. DESIGN RATIONALE At first glance, it may appear that the incorporation of animation clips into StyleCam unnecessarily complicates its authoring and use. After all, without animated transitions, we would not have had to develop an interaction technique that blended between spatial and temporal control. Indeed, when we first began our research, our hope was to create a system that simply involved spatial control of a constrained camera. Our first variation used a single camera surface that surrounded the 3D object of interest. The camera was constrained to remain normal to this single camera surface at all times. While this gave the author more control than using a simple unconstrained camera, we found that it was difficult to author a single camera surface that encompassed all the desirable viewpoints and interesting transitions between those viewpoints. In order to guarantee desirable viewpoints, we introduced the concept of money-shots that were placed on the single camera surface. The parameters of the camera were then determined based on its location on the camera surface and a weighted average of the surrounding money-shots. At this point, it was still difficult to author what the user would see when not directly on a money-shot. In other words, while money-shots worked well, the transitions between them worked poorly. To address this problem of unsatisfactory transitions, we first replaced the concept of a single global camera surface with separate local camera surfaces for each money-shot. 104 Volume 4, Issue 2 Then, to define transitions between these local camera surfaces, we introduced the idea of animating the camera. This led to the use of the three types of animation clips as described earlier. Simply playing back the animation clips between camera surfaces gave users the sense that they lost control during this period. To maintain the feeling of continuous control throughout, we developed our integrated spatial-temporal interaction technique. AN EXAMPLE EXPERIENCE We illustrate how StyleCam operates by an example. Figure 6 illustrates the system components and how they react to user input, as well as screen shots of what the user actually sees. The user starts by dragging on a camera surface (position A). The path A-B shows the camera being dragged on the surface (spatial navigation). At B, the user reaches the edge of the camera surface and this launches an animation that will transition the user from B to E. The zigzag path from B to D indicates that the user is scrubbing time on the animation (temporal navigation). Position C simply illustrates an intermediate point in the animation that gets seen three times during the interaction. At position D, the user releases the mouse button, whereupon the system automatically completes playing back the remainder of the animation at the authored pacing. At position E, the user enters another camera surface and resumes spatial navigation of the camera as shown by path E-F. When the user exits this camera surface at position F, another animation is launched that will transition the user to position J. Since the user releases the mouse button at position F, the animation from F to J is played back at the authored pacing. Since this animation is a slate animation, the intermediate shots at positions G, H, and I along the path F to J are of slates containing information on the car fading in and out as the camera pans over the top of the car. The net result of this StyleCam experience is a view of the car that is far more visually rich and influenced by an author who intends to convey a certain message, rather than using simple camera controls as is typical in current 3D viewers. RELATED WORK Much prior research has explored camera techniques for 3D virtual environments. Many of the techniques use a 2D mouse or stylus as an input device and introduce metaphors to assist the user. Perhaps the most ubiquitous metaphor, the cinematic camera, enables users to tumble, track and dolly a viewpoint. Various other metaphors have been explored by researchers, including orbiting and flying [32] , through-the-lens control [18] , points and areas of interests Figure 6. Example StyleCam experience. Top: system components and their reaction to user input. Bottom: what the user sees. Volume 4, Issue 2 105 [22] , using constraints [24, 29], drawing a path [21] , two-handed techniques [1, 38] , and combinations of techniques [30, 37]. Bowman et. al. present taxonomies and evaluations of various schemes [3, 4] . Other techniques involve automatic framing of the areas of interest as typically found in game console based adventure games which use a "chase airplane" metaphor for a third person perspective. Systems that utilize higher degree-of-freedom input devices offer additional control and alternative metaphors have been investigated, including flying [7, 34] , eyeball-in-hand [35] , and worlds in miniature [31] . The major difference between this body of prior research and our work is that we attempt to give the author substantially more influence over the types of views and transitions between them as the user navigates in the virtual space. Beyond techniques for navigating the scene, extra information can also be provided to aid navigation. These include global maps in addition to local views [12, 14] , and various landmarks [9, 33] . Others have investigated integrating global and local views, using various distorted spaces including "fisheye" views [6, 15] . At present, in an attempt to keep the visual space uncluttered, our work does not have mechanisms for providing global information to the user, however, this is something we may incorporate as our system progresses. Approaches which give the author more influence include guided tours where camera paths are prespecified for the end user to travel along. Galyean [17] proposes a "river analogy" where a user, on a metaphorical boat, can deviate from the guided path, the river, by steering a conceptual "rudder". Fundamental work by Hanson and Wernert [19, 36] proposes "virtual sidewalks" which are authored by constructing virtual surfaces and specifying gaze direction, vistas, and procedural events (e.g., fog and spotlights) along the sidewalk. Our system builds upon the guided tour and virtual sidewalk ideas but differs by providing authoring elements that enable a much more stylized experience. Specifically, we offer a means of presenting 3D, 2D, and temporal media experiences through a simple, unified, singular user interaction technique that supports both spatial and temporal navigation. Robotic planning algorithms have been used to assist or automatically create a guided tour of a 3D scene, in some cases resulting in specific behaviors trying to satisfy goals and constraints [10, 11] . Individual camera framing of a scene has been used to assist in viewing or manipulation tasks [27] . Rules can be defined for cameras to automatically frame a scene that follow cinematic principles such as keeping the virtual actors visible in the scene; or following the lead actor [20] . Yet another system [2] allows authors to define storyboard frames and the system defines a set of virtual cameras in the 3D scene to support the visual composition. This previous work assists in the authoring aspects by ceding some control to the system. Our work too involves some automatic system control, but we emphasize author control. Image based virtual reality environments such as QuicktimeVR [8] utilize camera panning and zooming and allow users to move to defined vista points. The driving metaphor has also been used for navigating interactive video, as seen in the Movie-Maps system [23] . More recently, the Steerable Media project [25] for interactive television aims to retain the visual aesthetic of existing television but increase the level of user interactivity. The user is given the ability to control the content progression by seamlessly integrating video with augmented 2D and 3D graphics. While our goals are similar in that we hope to enhance the aesthetics of the visual experience, we differ in that our dominant media type is 3D graphics with augmented temporal media (animations and visual effects) and traditional 2D media (video, still images). Lastly, we note that widely available 3D viewers or viewing technologies such as VRML, Cult3D, Shockwave, Viewpoint, Virtools, and Pulse3D, are becoming very popular but offer the standard camera controls of vista points, track, tumble, and zoom. We hope our explorations will ultimately assist in offering new experience and interaction approaches for future incarnations of these 3D viewers. IMPLEMENTATION StyleCam is implemented using Alias|wavefront's MAYA 3D modeling and animation package. We use MAYA to author the 3D content to be visualized, the required camera surfaces, animation clips, and required associations between them. A custom written MAYA plugin allows the user to control their view of the 3D content based on their mouse input and the authored camera surfaces, animation clips, and associations. The following description of our implementation assumes some knowledge of MAYA, although we have endeavoured to be as general as possible without sacrificing accuracy. 6.1. Authoring First, money-shots are created by defining a MAYA camera with specific position, orientation, and other camera parameters. Then, a camera surface which intersects the position of the money-shot camera is defined by creating an appropriate non-trimmed NURBS surface within MAYA. To include an optional camera look-at point, the author simply defines a point in 3D space (using a MAYA locator). Finally, to make these components easily locatable by the plugin, they are grouped under a named MAYA node within its dependency graph. Then, StyleCam animation clips are created as one would normally create animations in MAYA, using its TRAX non-linear animation editor. Animation clips at this stage are given meaningful, consistent, names in order to facilitate their identification later when associating them with events. 106 Volume 4, Issue 2 StyleCam allows the author to create scripts and associate them with events. Supported events are session startup, camera surface entry, camera surface exit, and camera surface timeout (Figure 7). We implement variable control-display gain on a camera surface (Figure 3) by varying the separation between the isoparms on the NURBS surface. As shown in Figure 4, StyleCam supports three types of transitions: automatic, authored, and slate. Automatic transitions are those that smoothly move the camera from one camera surface to another without requiring any authored animation clips. This is done by having the system perform quaternion [28] interpolation of camera orientation, combined quaternion and linear interpolation of camera position, and linear interpolation of other camera properties such as focal length. Using quaternion interpolation ensures smooth changes in orientation while defining a smooth arcing path for the position. At each time step in the transition, two quaternions representing the required fractional rotations of the position and orientation vectors of the camera are calculated and applied to the source vectors. In addition, the magnitude of the position vector is adjusted by linear interpolation between the source and destination position vector magnitudes. The result is a series of intermediate camera positions and orientations as Figure 8 illustrates. Figure 7. StyleCam events The session startup event is triggered only once when the user initially begins using StyleCam to view a scene. Exit events are triggered when the user leaves a camera surface from one of four directions. Associated scripts can specify destination camera surfaces and types of transitions to be performed. Time-out events are triggered when the mouse is idle for a given duration while on a particular camera surface, and can be used to launch an automatic presentation. StyleCam's event and script mechanism provides for the use of logic to dynamically alter the presentation. For example, scripts can ensure that some surfaces are only visited once, while others are shown only after certain surfaces have already been visited. 6.2. Interaction When the StyleCam plugin is activated, the first money-shot of the first camera surface is used as the initial view. If a look-at point is defined for this camera surface, the orientation of the user camera is set such that the camera points directly at the look-at point. Otherwise, the orientation is set to the normal of the camera surface at the money-shot viewpoint's position. Figure 8. Combined quaternion and linear interpolation Authored transitions involve the playback of preauthored animation clips. This gives the author complete control over the user experience during the transition including the pacing, framing and visual effects. User's mouse movements and button presses are monitored by the StyleCam plugin. Mouse drags result in the camera moving along the current camera surface. Specifically, for a given mouse displacement (dx, dy), the new position of the camera on the camera surface (in uv-coordinates local to the camera surface) is given by Slate transitions are a special case of authored transitions. Used to present 2D media, slate transitions are authored by placing an image plane in front of the camera as it transitions between camera surfaces. Various visual effects can be achieved by using multiple image planes simultaneously and by animating transparency and other parameters of these image planes. While the slate transition is in progress, the camera is simultaneously being smoothly interpolated towards the destination camera surface. This essentially allows for a "soft" fade from a camera view, to a slate, and back, as Figure 9 illustrates. (u1,v1) = (u0,v0) + c*(dx, dy) where (u0, v0) is the last position of the camera, and c is the gain constant. If either the u or v coordinate of the resulting position is not within the range [0,1], the camera has left the current camera surface. At this point, the author-scripted logic is executed to determine the next step. First, the destination money-shot is resolved. Next, an appropriate transition is performed to move to the next camera surface. Volume 4, Issue 2 107 Figure 9. Slate transitions StyleCam supports temporal control or "scrubbing" of animations. During navigation mode, the user's mouse drags control the camera's position on the camera surface. However, when the user moves off a camera surface into an animated transition, mouse drags control the (invisible) timeslider of the animation. Time is advanced when the mouse is dragged in the same direction that the camera exited the camera surface and reversed if the directions are also reversed. When the mouse button is released, the system takes over time management and smoothly ramps the time steps towards the animation's original playback rate. Our present implementation supports scrubbing only for automatic transitions. Authored and slate transitions are currently uninterruptible. There is however no technical reason why all transitions cannot support scrubbing. In future versions we intend to give the author the choice of determining whether or not any given transition is scrubable. This is important since in some cases it may be desirable to force the animation to playback uninterrupted at a certain rate. EVALUATION We conducted an informal user study to get a sense of users' initial reactions to using StyleCam. Seven participants, three of whom had experience with 3D graphics applications and camera control techniques, and four who had never used a 3D application or camera controls, were asked to explore a 3D car model using StyleCam. In order to ensure the study resembled our intended casual usage scenario, we gave participants only minimal instructions. We explained the click-and-drag action required to manipulate the camera, a brief rationale for the study, and to imagine they were experiencing an interactive advertisement for that car. We did not identify the various components (camera surfaces, animated transitions, etc) nor give any details on them. This was deliberately done so that the participants could experience these components in action for themselves and give us feedback without knowing in advance of their existence. One very promising result was that none of the participants realized that they were switching between controlling the camera and controlling the time slider on the animations. They felt that they had the same type of control throughout, indicating that our blending between spatial and temporal control worked remarkably well. Also the simplicity of the interaction technique essentially a single click and drag action was immediately understood and usable by all our users. Another reaction from all the participants was that, to varying degrees, they sometimes felt that they were not in control of the interaction when the uninterruptable animations occurred. This was particularly acute when the information in the animations seemed unrelated to their current view. In these cases, participants indicated that they had no idea what triggered these animations and were often annoyed at the sudden interruptions. However when the information was relevant the interruptions were not as annoying and often actually appreciated. In some cases participants indicated that they would have liked to be able to replay the animation or to have it last longer. This highlights the importance of carefully authoring the intermingling of uninterruptable animations with the rest of the interaction experience. Participants also indicated that they would have liked the ability to click on individual parts of the car model in order to inspect them more closely. This request is not surprising since we made no effort in our current implement to support pointing. However, we believe that in future research StyleCam could be extended to include pointing. As we expected, all the participants with prior 3D graphics camera experience stated that they at times would have liked full control of the camera, in addition to the constrained control we provided. Participants without this prior experience, however, did not ask for this directly although they indicated that there were some areas of the car model that they would have liked to see but could not get to. However, this does not necessarily imply full control of the camera is required. We believe that this issue can be largely alleviated at the authoring phase by ascertaining what users want to see for a particular model and ensuring that those features are accessible via the authored camera surfaces. Interestingly, the participant with the most 3D graphics experience commented that the automatic transitions and smooth camera paths during those transitions were very good and that "for those who don't know 3D and stuff, this would be very good"! DISCUSSION & CONCLUSIONS Central to our StyleCam system is the integration of spatial and temporal controls into a single user interaction model. The implications of this interaction model go far beyond a simple interaction technique. The blending of spatial and temporal control presents a completely new issue that an author needs to understand and consider when creating these interactive visual experiences. As evident from the comments of our users, temporal control can feel very much like spatial control even when scrubbing backwards 108 Volume 4, Issue 2 in an animation when the animation consists of moving the viewing camera around the central object of interest. However, if the animation is not around the central object of interest, for example in some of our slate animations, temporal control can produce very different sensations. These include the feeling of moving backwards in time, interruption of a well paced animation, jarring or ugly visuals, and sometimes even nonsensical content. As a result, the author needs to be extremely cognizant of these artefacts and make design decisions as to when and where to relinquish control - and how much control - to the user. At one extreme, the author can specify that certain animations are completely uninterruptible by the user. In the experience we authored for our user study, we included several of these types of transitions. As discussed earlier, whether users favored this depended heavily on the content. In other words, in some cases, as authors, we did not make the right decision. Further improvements could include partially interruptable animations. For example, we may not allow movement backwards in time but allow the user to control the forward pacing. This will largely solve the nonsensical content problem but may still result in occasionally jarring visuals. If we intend to support these various types of control, we must also be able to set the users' expectations of what type of control they have at any given time. It is clear that the current StyleCam switching between spatial and temporal control without any explicit indication to the user that a switch is happening works in most cases. In the cases where it fails, either the visual content itself should indicate what control is possible, or some explicit mechanism is required to inform the user of the current or upcoming control possibilities. In addition to the obvious solution of using on-screen visual indicators (e.g., changing cursors) to indicate state, future research could include exploring "hint-ahead" mechanisms that indicate upcoming content if the user chooses to stay on their current course of travel. For example, as the user reaches the edge of a camera surface, a "voice-over" could say something like "now we're heading towards the engine of the car". Alternatively, a visual "signpost" could fade-in near the cursor location to convey this information. These ideas coincide with research that states that navigation routes must be discoverable by the user [16] . It is very clear from our experiences with StyleCam that the user's viewing experience is highly dependent on the talent and skill of the author. It is likely that skills from movie making, game authoring, advertising, and theme park design would all assist in authoring compelling experiences. However, we also realize that authoring skills from these other genres do not necessarily directly translate due to the unique interaction aspects of StyleCam. While StyleCam has the appropriate components for creating compelling visual experiences, it is still currently a research prototype that requires substantial skills with MAYA. We envision a more author-friendly tool that is based on the conceptual model of StyleCam. Some future avenues that we intend to explore include supporting soundtracks, extensions to enable pointing to elements in the 3D scene, and mechanisms for authoring animation paths using alternate techniques such as Chameleon [13] . Finally, it is important to note that StyleCam is not limited to product or automobile visualization. Other domains such as visualization of building interiors and medical applications could also utilize the ideas presented in this paper. Figures 10, 11, and 12 illustrate some examples. ACKNOWLEDGEMENTS We thank Scott Guy and Miles Menegon for assistance in figure and video creation. REFERENCES 1. Balakrishnan, R., & Kurtenbach, G. (1999). Exploring bimanual camera control and object manipulation in 3D graphics interfaces. ACM CHI 1999 Conference on Human Factors in Computing Systems. p. 56-63. 2. Bares, W., McDermott, S., Boudreaux, C., & Thainimit, S. (2000). Virtual 3D camera composition from frame constraints. ACM Multimedia. p. 177-186. 3. Bowman, D.A., Johnson, D.B., & Hodges, L.F. (1997). Travel in immersive virtual environments. IEEE VRAIS'97 Virtual Reality Annual International Symposium. p. 45-52. 4. Bowman, D.A., Johnson, D.B., & Hodges, L.F. (1999). Testbed environment of virtual environment interaction. ACM VRST'99 Symposium on Virtual Reality Software and Technologies. p. 26-33. 5. Buxton, W., ed. Three-state model of graphical input. Human-computer interaction - INTERACT'90, ed. D. Diaper. 1990, Elsevier Science Publishers B. V. (North-Holland ): Amsterdam. 449-456. 6. Carpendale, M.S.T., & Montagnese, C.A. (2001). A framework for unifying presentation space. ACM UIST'2001 Symposium on User Interface Software and Technology. p. 61-70. 7. Chapman, D., & Ware, C. (1992). Manipulating the future: predictor based feedback for velocity control in virtual environment navigation. ACM I3D'92 Symposium on Interactive 3D Graphics. p. 63-66. 8. Chen, S.E. (1995). QuickTime VR: An image-based approach to virtual environment navigation. ACM SIGGRAPH'95 Conference on Computer Graphics and Interactive Techniques. p. 29-38. 9. Darken, R., & Sibert, J. (1996). Wayfinding strategies and behaviours in large virtual worlds. ACM CHI'96 Conference on Human Factors in Computing Systems. p. 142-149. Volume 4, Issue 2 109 10. Drucker, S.M., Galyean, T.A., & Zeltzer, D. (1992). CINEMA: A system for procedural camera movements. ACM Symposium on Interactive 3D Graphics. p. 67-70. 11. Drucker, S.M., & Zeltzer, D. (1994). Intelligent camera control in a virtual environment. Graphics Interface. p. 190-199. 12. Elvins, T., Nadeau, D., Schul, R., & Kirsh, D. (1998). Worldlets: 3D thumbnails for 3D browsing. ACM CHI'98 Conf. on Human Factors in Computing Systems. p. 163-170. 13. Fitzmaurice, G.W. (1993). Situated information spaces and spatially aware palmtop computers. Communications of the ACM, 36(7). p. 38-49. 14. Fukatsu, S., Kitamura, Y., Masaki, T., & Kishino, F. (1998). Intuitive control of bird's eye overview images for navigation in an enormous virtual environment. ACM VRST'98 Sympoisum on Virtual Reality Software and Technology. p. 67-76. 15. Furnas, G. (1986). Generalized fisheye views. ACM CHI 1986 Conference on Human Factors in Computing Systems. p. 16-23. 16. Furnas, G. (1997). Effective view navigation. ACM CHI'97 Conference on Human Factors in Computing Systems. p. 367-374. 17. Galyean, T.A. (1995). Guided navigation of virtual environments. ACM I3D'95 Symposium on Interactive 3D Graphics. p. 103-104. 18. Gliecher, M., & Witkin, A. (1992). Through-the-lens camera control. ACM SIGGRAPH' Conf. on Computer Graphics and Interactive Techniques. p. 331-340. 19. Hanson, A.J., & Wernet, E. (1997). Constrained 3D navigation with 2D controllers. p. 175-182. 20. He, L., Cohen, M.F., & Salesin, D. (1996). The virtual cinematographer: a paradigm for automatic real-time camera control and directing. ACM SIGGRAPH'96 Conference on Computer Graphics and Interactive Techniques. p. 217-224. 21. Igarashi, T., Kadobayashi, R., Mase, K., & Tanaka, H. (1998). Path drawing for 3D walkthrough. ACM UIST 1998 Symposium on User Interface Software and Technology. p. 173-174. 22. Jul, S., & Furnas, G. (1998). Critical zones in desert fog: aids to multiscale navigation. ACM Symposium on User Interface Software and Technology. p. 97-106. 23. Lippman, A. (1980). Movie-maps: an application of the optical videodisc to computer graphics. ACM SIGGRAPH'80 Conference on Computer Graphics and Interactive Techniques. p. 32-42. 24. Mackinlay, J., Card, S., & Robertson, G. (1990). Rapid controlled movement through a virtual 3D workspace. ACM SIGGRAPH 1990 Conference on Computer Graphics and Interactive Techniques. p. 171-176. 25. Marrin, C., Myers, R., Kent, J., & Broadwell, P. (2001). Steerable media: interactive television via video synthesis. ACM Conference on 3D Technologies for the World Wide Web. p. 7-14. 26. Newman, W. (1968). A system for interactive graphical programming. AFIPS Spring Joint Computer Conference. p. 47-54. 27. Phillips, C.B., Badler, N.I., & Granieri, J. (1992). Automatic viewing control for 3D direct manipulation. ACM Symposium on Interactive 3D Graphics. p. 71-74. 28. Shoemake, K. (1985). Animating rotation with quartenion curves. ACM SIGGRAPH Conf Computer Graphics & Interactive Techniques. p. 245-254. 29. Smith, G., Salzman, T., & Stuerzlinger, W. (2001). 3D Scene manipulation with 2D devices and constraints. Graphics Interface. p. 135-142. 30. Steed, A. (1997). Efficient navigation around complex virtual environments. ACM VRST'97 Conference on Virtual Reality Software and Technology. p. 173-180. 31. Stoakley, R., Conway, M., & Pausch, R. (1995). Virtual reality on a WIM: Interactive worlds in miniature. ACM CHI 1995 Conference on Human Factors in Computing Systems. p. 265-272. 32. Tan, D., Robertson, G., & Czerwinski, M. (2001). Exploring 3D navigation: combining speed-coupled flying with orbiting. ACM CHI'2001 Conference on Human Factors in Computing Systems. p. 418-425. 33. Vinson, N. (1999). Design guidelines for landmarks to support navigation in virtual environments. ACM CHI'99 Conference on Human Factors in Computing Systems. p. 278-285. 34. Ware, C., & Fleet, D. (1997). Context sensitve flying interface. ACM I3D'97 Symposium on Interactive 3D Graphics. p. 127-130. 35. Ware, C., & Osborne, S. (1990). Exploration and virtual camera control in virtual three dimensional environments. ACM I3D'90 Symposium on Interactive 3D Graphics. p. 175-183. 36. Wernert, E.A., & Hanson, A.J. (1999). A framework for assisted exploration with collaboration. IEEE Visualization. p. 241-248. 37. Zeleznik, R., & Forsberg, A. (1999). UniCam - 2D Gestural Camera Controls for 3D Environments. ACM Symposium on Interactive 3D Graphics. p. 169-173. 38. Zeleznik, R., Forsberg, A., & Strauss, P. (1997). Two pointer input for 3D interaction. ACM I3D Symposium on Interactive 3D Graphics. p. 115-120. 110 Volume 4, Issue 2

3D viewers;camera controls;3D navigation;3D visualization;interaction techniques

185

Tactons: Structured Tactile Messages for Non-Visual Information Display

Tactile displays are now becoming available in a form that can be easily used in a user interface. This paper describes a new form of tactile output. Tactons, or tactile icons, are structured, abstract messages that can be used to communicate messages non-visually. A range of different parameters can be used for Tacton construction including : frequency, amplitude and duration of a tactile pulse, plus other parameters such as rhythm and location. Tactons have the potential to improve interaction in a range of different areas, particularly where the visual display is overloaded, limited in size or not available, such as interfaces for blind people or in mobile and wearable devices . . This paper describes Tactons, the parameters used to construct them and some possible ways to design them. Examples of where Tactons might prove useful in user interfaces are given.

Introduction The area of haptic (touch-based) human computer interaction (HCI) has grown rapidly over the last few years. A range of new applications has become possible now that touch can be used as an interaction technique (Wall et al., 2002). However, most current haptic devices have scant provision for tactile stimulation, being primarily pro-grammable , constrained motion force-feedback devices for kinaesthetic display. The cutaneous (skin-based) component is ignored even though it is a key part of our experience of touch (van Erp, 2002). It is, for example, important for recognising texture, and detecting slip, compliance and direction of edges. As Tan (1997) says "In the general area of human-computer interfaces ... the tactual sense is still underutilised compared with vision and audition". One reason for this is that, until recently, the technology for tactile displays was limited. Tactile displays are not new but they have not received much attention from HCI researchers as they are often engineering prototypes or designed for very specific applications (Kaczmarek et al., 1991). They have been used in areas such as tele-operation or displays for blind people to provide sensory substitution where one sense is used to receive information normally received by another (Kaczmarek et al.). Most of the development of these devices has taken place in robotics or engineering labs and has focused on the challenges inherent in building low cost, high-resolution devices with realistic size, power and safety performance. Little research has gone into how they might actually be used at the user interface. Devices are now available that allow the use of tactile displays so the time is right to think about how they might be used to improve interaction. In this paper the concept of Tactons, or tactile icons, is introduced as a new communication method to complement graphical and auditory feedback at the user interface . Tactons are structured, abstract messages that can be used to communicate messages non-visually. Conveying structured messages through touch will be very useful in areas such as wearable computing where screens are limited . The paper gives some background to the perception and use of tactile stimuli and then describes the design of Tactons. It finishes with examples of potential uses for Tactons. Background and previous work The skin is the largest organ in the body, about 2 m 2 in the average male (Montagu, 1971). Little direct use is made of it for displaying information in human-computer interfaces (Tan and Pentland, 1997, van Erp, 2002), yet a touch on the hand or other parts of the body is a very rich experience. The skin can therefore potentially be used as a medium to communicate information. As a receiving instrument the skin combines important aspects of the eye and the ear, with high acuity in both space and time (Gunther, 2001) giving it good potential as a communication medium. The human sense of touch can be roughly split in to two parts: kinaesthetic and cutaneous. "Kinaesthetic" is often used as catch-all term to describe the information arising from forces and positions sensed by the muscles and joints. Force-feedback haptic devices (such as the PHANToM from SensAble) are used to present information to the kinaesthetic sense. Cutaneous perception refers to the mechanoreceptors contained within the skin, and includes the sensations of vibration, temperature, pain and indentation. Tactile devices are used to present feedback to the cutaneous sense. 15 Current haptic devices use force-feedback to present kinaesthetic stimuli. This works well for some aspects of touch (e.g. identifying the geometric properties of objects ) but is poor for features such as texture (normally perceived cutaneously). Oakley et al. (2000) found that trying to use texture in a user interface with a force-feedback device actually reduced user performance. One reason for this is that the textures had to be made large so that they could be perceived kinaesthetically, but they then perturbed users' movements. The use of a tactile haptic device to present texture would not have this problem as small indentations in the fingertip would not affect hand movements. At present, however, there are no haptic devices that do a good job of presenting both tactile and force-feedback cues to users. Current force-feedback devices use a point interaction model; the user is represented by a single point of contact corresponding to the tip of a stylus. This is analogous to exploring the world by remote contact through a stick thus depriving the user of the rich, spatially varying cutaneous cues that arise on the finger pad when contacting a real object (Wall and Harwin, 2001). Users must integrate temporally varying cues as they traverse the structure of virtual objects with the single point of contact, which places considerable demands on short-term memory (Jansson and Larsson, 2002). Even when exploring simple geometric primitives, performance is greatly reduced compared to natural touch. Lederman and Klatzky (1999) have shown that such removal of cutaneous input to the fingertip impedes perception of edge direction, which is an essential component of understanding haptic objects. It can therefore be seen that tactile feedback and cutaneous perception are key parts of touch that must be incorporated into haptic displays if they are to be effective and usable. 2.1 Vibrotactile actuators There are two basic types of vibrotactile display device. These evoke tactile sensations using mechanical vibration of the skin (usually in the range 10-500Hz) (Kaczmarek et al., 1991). This is commonly done by vibrating a small plate pressed against the skin or via a pin or array of pins on the fingertip. These are very easy to control from standard PC hardware. Other types of actuator technology are available, including pneumatic and electrotactile (Stone, 2000), but these tend to be bulkier and harder to control so are less useful in many situations. Figure 1: The pins arrays on the VirTouch tactile mouse (www.virtouch.com). The first type of vibrotactile display uses a pin or array of small pins (e.g. the VirTouch mouse in Figure 1 or those produced by Summers et al. (2001)) to stimulate the fingertip . Such devices can present very fine cues for surface texture, edges, lines, etc. The second type uses larger point-contact stimulators (e.g. Figure 2 or alternatively small loudspeaker cones playing tones, or other simple vibrating actuators placed against the skin as used by Tan (1997) and in devices such as the CyberTouch glove www.immersion.com). The cues here are much lower resolution but can exert more force; they can also be distributed over the body to allow multiple simultaneous cues (often mounted in a vest on the user's back or in a belt around the waist). These devices are both easy to control and use. For a full review see Kaczmarek et al. (1991). Figure 2: Audiological Engineering Corp. VBW32 transducers (www.tactaid.com). 2.2 Previous work on tactile display One common form of tactile output is Braille, and dynamic Braille cells are available. A display is made up of a line of `soft' cells (often 40 or 80), each with 6 or 8 pins that move up and down to represent the dots of a Braille cell. The user can read a line of Braille cells by touching the pins of each cell as they pop up (for more information see www.tiresias.org). The focus of the work reported here is not on Braille as it tends to be used mainly for representing text (although other notations are used, e.g. music) and the cells are very low resolution (8 pins maximum). These displays are also very expensive with an 80 cell display costing around 4000. There have been many other tactile devices for blind people, such as the Optacon (TeleSensory Inc.), which used an array of 144 pins to display the input from a camera to the fingertip, but again these are mainly used for reading text. Pin arrays produce Braille but can do much more, especially the higher resolution displays such as shown in Figure 1. Our research also builds on the work that has been done on tactile graphics for blind people (this mainly takes the form of raised lines and dots on special `swell' paper). Kurze (1997, 1998) and Challis (2001) have developed guidelines which allow images and objects to be presented that are understandable through touch by blind users. Two other examples show that the cutaneous sense is very effective for communication. Firstly, Tadoma is a tactile language used by deaf/blind people. The transmitter speaks normally and the receiver puts a hand on the face of the speaker, covering the mouth and neck (Tan and Pentland, 2001). Tadoma users can listen at very high 16 speeds (normal speaking speed for experts) and pick up subtleties of the speech such as accent. In the second example , Geldard (1957) taught participants a simple tactile language of 45 symbols, using three intensities, three durations and five locations on the chest. Participants were able to learn the alphabet quickly and could recognise up to 38 words per minute in some cases. Other sensory substitution systems convert sound into vibration for hearing-impaired people (e.g. the TactAid system from Audiological Engineering). Again this shows that cutaneous perception is very powerful and if we can make use of it at the user interfaces we will have a rich new way to present information to users. Research and existing applications have shown that the cutaneous sense is a very powerful method of receiving information. Other work has shown that it can be used in user interfaces and wearable computers (Gemperle et al., 1998). Tan has begun to investigate the use of tactile displays on wearable computers (Tan and Pentland, 1997). She used a 3x3 grid of stimulators on a user's back to provide navigation information. Informal results suggested it was useful but no formal evaluation has taken place. Other relevant work has taken place in aircraft cockpits to provide pilots with navigation information (van Veen and van Erp, 2001, Rupert, 2000). In these examples only simple tactile cues for direction have been provided. For example, an actuator maybe vibrated on one side of the body to indicate the direction to turn. More sophisticated cues could be used to provide much more information to users without them needing to use their eyes. Gunther et al. have used tactile cues to present `musical' compositions to users (Gunther, 2001, Gunther et al., 2002). They say: "The approach taken ... views haptic technologies in particular the vibrotactile stimulator as independent output devices to be used in conjunction with the composition and perception of music. Vibrotactile stimuli are viewed not as signals carrying information per se, but as aesthetic artifacts themselves". He used an array of 13 transducers across the body of a `listener' so that he/she could experience the combined sonic/tactile presentation. Gunther created a series of compositions played to listeners who appeared to enjoy them. This work was artistic in nature so no formal usability assessments were made but the listeners all liked the experience . In order to create a tactile composition (the same is true for the Tactons described below) a good understanding of the experience of touch is needed. However, as Gunther et al. suggest: "It is indeed premature to hammer out the details of a language for tactile composition. It seems more productive at this point in time to identify the underpinnings of such a language, specifically those dimensions of tactile stimuli that can be manipulated to form the basic vocabulary elements of a compositional lan-guage" . Research is needed to gain a more systematic understanding of cutaneous perception for use in the presentation of such messages. Enriquez and MacLean (2003) recently proposed `haptic icons', which they define as "brief programmed forces applied to a user through a haptic interface, with the role of communicating a simple idea in a manner similar to visual or auditory icons". The problem they are trying to address is different to that of Tactons, as they say "With the introduction of "active" haptic interfaces, a single handle e.g. a knob or a joystick can control several different and perhaps unrelated functions. These multi-function controllers can no longer be differentiated from one another by position, shape or texture... Active haptic icons, or "hapticons", may be able to solve this problem by rendering haptically distinct and meaningful sensations for the different functions". These use one degree-of -freedom force-feedback devices, rather than tactile displays, so encode information very differently to Tactons . They report the construction of a tool to allow a user to create and edit haptic icons. This is early work and they do not report results from the use of hapticons in any interfaces. Their results, however, will be directly relevant to Tactons. Tactons Given that the cutaneous sense is rich and a powerful communication medium currently little utilised in HCI, how can we make effective use of it? One approach is to use it to render objects from the real world more realisti-cally in virtual environments, for example in improving the presentation of texture in haptic devices. It could also be used to improve targeting in desktop interactions along the lines suggested by Oakley et al. (2000). In this paper it is suggested that it can additionally be used to present structured informational messages to users. Tactons are structured, abstract messages that can be used to communicate complex concepts to users non-visually. Shneiderman (1998) defines an icon as "an image, picture or symbol representing a concept". Tactons can represent complex interface concepts, objects and actions very con-cisely . Visual icons and their auditory equivalent earcons (Blattner et al., 1989, Brewster et al., 1994) are very powerful ways of displaying information but there is currently no tactile equivalent. In the visual domain there is text and its counterpart the icon, the same is true in sound with synthetic speech and the earcon. In the tactile domain there is Braille but it has no `iconic' counterpart. Tactons fill this gap. Icons/Earcons/Tactons form a simple , efficient language to represent concepts at the user interface. Tactons are similar to Braille in the same way that visual icons are similar to text, or earcons are similar to synthetic speech. For example, visual icons can convey complex information in a very small amount of screen space, much smaller than for a textual description. Earcons convey information in a small amount of time as compared to synthetic speech. Tactons can convey information in a smaller amount of space and time than Braille. Research will also show which form of iconic display is most suitable for which type of information. Visual icons are good for spatial information, earcons for temporal. One property of Tactons is that they operate both spatially and temporally so they can complement both icons and earcons . Further research is needed to understand how these different types of feedback work together. 17 Using speech as an example from the auditory domain: presenting information in speech is slow because of its serial nature; to assimilate information the user must hear a spoken message from beginning to end and many words may have to be comprehended before the message can be understood. With earcons the messages are shorter and therefore more rapidly heard, speeding up interactions. The same is true of Tactons when compared to Braille. Speech suffers from many of the same problems as graphical text in text-based computer systems, as this is also a serial medium. Barker & Manji (1989) claim that an important limitation of text is its lack of expressive capability: It may take many words to describe a fairly simple concept. Graphical iconic displays were introduced that speeded up interactions as users could see a picture of the thing they wanted instead of having to read its name from a list (Barker and Manji, 1989). In the same way, an encoded tactile message may be able to communicate its information in fewer symbols. The user feels the Tacton then recalls its meaning rather than having the meaning described in Braille (or speech or text). The icon is also (in principle) universal: it means the same thing in different languages and the Tacton would have similar universality. Designing with Tactons Tactons are created by encoding information using the parameters of cutaneous perception. The encoding is similar to that of earcons in sound (Blattner et al., 1989, Brewster et al., 1994) where each of the musical parameters (e.g. timbre, frequency, amplitude) is varied to encode information. Similar parameters can be used for Tactons (although their relative importance is different). As suggested by Blattner, short motifs could be used to represent simple objects or actions and these can then be combined in different ways to represent more complex messages and concepts. As Tactons are abstract the mapping between the Tacton and what it represents must be learned, but work on earcons has shown that learning can take place quickly (Brewster, 1998b). The properties that can be manipulated for Tactons are similar to those used in the creation of earcons. The parameters for manipulation also vary depending on the type of transducer used; not all transducers allow all types of parameters. The general basic parameters are: Frequency: A range of frequencies can be used to differentiate Tactons. The range of 20 1000 Hz is perceivable but maximum sensitivity occurs around 250 Hz (Gunther et al., 2002). The number of discrete values that can be differentiated is not well understood, but Gill (2003) suggests that a maximum of nine different levels can be used. As in audition, a change in amplitude leads to a change in the perception of frequency so this has an impact on the use of frequency as a cue. The number of levels of frequency that can be discriminated also depends on whether the cues are presented in a relative or absolute way. Making relative comparisons between stimuli is much easier than absolute identification, which will lead to much fewer discriminable values, as shown in the work on earcon design (Brewster et al., 1994). Amplitude: Intensity of stimulation can be used to encode values to present information to the user. Gunther (2002) reports that the intensity range extends to 55 dB above the threshold of detection; above this pain occurs. Craig and Sherrick (1982) indicate that perception deteriorates above 28 dB so this would seem to be a useful maximum. Gunther (2001) reports that various values, ranging from 0.4dB to 3.2dB, have been reported for the just noticeable difference (JND) value for intensity. Gill states that that no more than four different intensities should be used (Gill, 2003). Again the number of useful discriminable values will depend on absolute or relative presentation of stimuli. Due to the interactions between this and frequency several researchers have suggested that they be combined into a single parameter to simplify design Waveform: The perception of wave shape is much more limited than with the perception of timbre in sound. Users can differentiate sine waves and square waves but more subtle differences are more difficult (Gunther, 2001). This limits the number of different values that can be encoded and makes this a much less important variable than it is in earcon design (where it is one of the key variables ). Duration: Pulses of different durations can encode information . Gunther (2001) investigated a range of subjective responses to pulses of different durations. He found that stimuli lasting less than 0.1 seconds were perceived as taps or jabs whereas stimuli of longer duration, when combined with gradual attacks and decays, may be perceived as smoothly flowing tactile phrases. He suggests combining duration with alterations in the envelope of a vibration, e.g. an abrupt attack feels like a tap against the skin, a gradual attack feels like something rising up out of the skin. Rhythm: Building on from duration, groups of pulses of different durations can be composed into rhythmic units. This is a very powerful cue in both sound and touch. Gunther (2001) suggests that differences in duration can be used to group events when multiple events occur on the same area of skin. Specific transducer types allow other parameters to be used: Body location: Spatially distributed transducers can encode information in the position of stimulation across the body. The choice of body location for vibrotactile display is important, as different locations have different levels of sensitivity and spatial acuity. A display may make use of several body locations, so that the location can be used as another parameter, or can be used to group tactile stimuli. The fingers are often used for vibrotactile displays because of their high sensitivity to small amplitudes and their high spatial acuity (Craig and Sherrick, 1982). However , the fingers are often required for other tasks, so other body locations may be more suitable. Craig and Sherrick suggest the back, thigh and abdomen as other suitable body locations. They report that, once subjects have been trained in vibrotactile pattern recognition on the back, they can almost immediately recognise the same patterns when they are presented to the thigh or abdomen. This transfer also occurs to some extent when patterns are 18 presented to different fingers after training on one finger, but is not so immediate. Certain body locations are particularly suitable, or particularly unsuitable, for certain types of vibrotactile displays . For example, transducers should not be placed on or near the head, as this can cause leakage of vibrations into the ears, resulting in unwanted sounds (Gunther et al., 2002). An example of a suitable body location is in Gunther's Skinscape display, where he positions low frequency transducers on the torso as this is where low frequencies are felt when loud music is heard. The method of attaching the transducers to a user's body is also important. The pressure of the transducer against the body has a significant effect on the user's perception of the vibrations. Transducers should rest lightly on the skin, allowing the user to feel the vibration against the skin, and to isolate the location of the vibration with ease. Exerting too much pressure with the transducer against the user's body will cause the vibrations to be felt in the bone structure, making them less isolated due to skeletal conduction. In addition, tightening the straps holding the transducer to achieve this level of pressure may impede circulation (Gunther, 2001). Rupert (2000) suggests using the full torso for displaying 3D information, with 128 transducers distributed over the body. His system displays information to pilots about the location of objects around them in 3D space, by stimulating the transducers at the part of their body corresponding to the location of the object in 3D space around them. This could be used to indicate horizons, borders, targets, or other aircraft. Spatiotemporal patterns: Related to position and rhythm, spatial patterns can also be "drawn" on the user's body. For example, if a user has a 3x3 array of stimulators lo-cated on his/her back, lines and geometric shapes can be "drawn" on the back, by stimulating, in turn, the stimulators that make up that shape. In Figure 3, an `L' shaped gesture can be drawn by activating the stimulators: 1-4-78 -9 in turn. Patterns can move about the body, varying in time and location to encode information. Cholewiak (1996) and Sherrick (1985) have also looked at low-level perception of distributed tactile cues. . Figure 3: "Drawing" an L-shaped gesture. Now that the basic parameters for Tactons have been described , we will give some examples of how they might be designed to convey information. The fundamental design of Tactons is similar to that of earcons. 4.1 Compound Tactons A simple set of Tactons could be created as in Figure 4. A high-frequency pulse that increases in intensity could represent `Create', a lower frequency pulse that decreases in intensity could represent `Delete'. A two note falling Tacton could represent a file and a two rising notes a folder. The mapping is abstract; there is no intuitive link between what the user feels and what it represents. Create Delete File Folder Create File Delete Folder Figure 4: Compound Tactons (after Blattner et al., 1989). These Tactons can then be combined to create compound messages. For example, `create file' or `delete folder'. The set of basic elements could be extended and a simple language of tactile elements created to provide feedback in a user interface. 4.2 Hierarchical Tactons Tactons could also be combined in a hierarchical way, as shown in Figure 5. Each Tacton is a node in a tree and inherits properties from the levels above it. Figure 5 shows a hierarchy of Tactons representing a hypothetical family of errors. The top of the tree is a family Tacton which has a basic rhythm played using a sinewave (a different family of errors would use a different rhythm so that they are not confused). The rhythmic structure of Level 2 inherits the Tacton from Level 1 and adds to it. In this case a second, higher frequency Tacton played with a squarewave. At Level 3 the tempo of the two Tactons is changed. In this way a hierarchical structure can be presented . The other parameters discussed above could be used to add further levels. 4.3 Transformational Tactons A third type of Tacton is the Transformational Tacton. These have several properties, each represented by a different tactile parameter. For example, if Transformational Tactons were used to represent files in a computer interface , the file type could be represented by rhythm, size by frequency, and creation date by body location. Each file type would be mapped to a unique rhythm. Therefore, two files of the same type, and same size, but different creation date would share the same rhythm and frequency , but would be presented to a different body location . If two files were of different types but the same size they would be represented by different rhythms with the same frequency. 19 Uses for Tactons We are interested in three areas of use for Tactons, although there are many others where they have potential to improve usability. 5.1 Enhancements of desktop interfaces The first, and simplest, area of interest is in the addition of Tactons to desktop graphical interfaces. The addition of earcons to desktops has shown many advantages in terms of reduced errors, reduced times to complete tasks and lowered workload (Brewster, 1998a). One problem with audio is that users believe that it may be annoying to use (although no research has actually shown this to be the case) and it has the potential to annoy others nearby (for a discussion see (Brewster, 2002)). The addition of Tactons to widgets has the same potential to indicate usability problems but without the potential to annoy. One reason for enhancing standard desktop interfaces is that users can become overloaded with visual information on large, high-resolution displays. In highly complex graphical displays users must concentrate on one part of the display to perceive the visual feedback, so that feedback from another part may be missed. This becomes very important in situations where users must notice and deal with large amounts of dynamic data or output from multiple applications or tasks. If information about secondary tasks was presented through touch then users could concentrate their visual attention on the primary one but feel information about the others. As a simple example, the display of a progress bar widget could be presented tactually. Two sets of tactile pulses could be used to indicate the current and end points of a download. The time between the two pulses would indicate the amount of time remaining, the closer the two pulses the nearer the download is to finishing. The two pulses could use different waveforms to ensure they were not confused. Different rhythms for each pulse could be used to indicate different types of downloads. If a more sophisticated set of transducers on a belt around the waist was available then the position of a pulse moving around the body in a clockwise direction (starting from the front) would give information about progress: when the pulse was at the right side of the body the download would be 25% of the way through, when it was on the left hand side 75%, and when it got back around to the front it would be finished. There would be no need for any visual presentation of the progress bar, allowing users to focus their visual attention on the main task they are involved with. Tactons could also be used to enhance interactions with buttons, scrollbars, menus, etc. to indicate when users are on targets and when certain types of errors occur. Others have shown that basic tactile feedback can improve pointing and steering type interactions (Akamatsu et al., 1995, Campbell et al., 1999). There are some commercial systems that give simple tactile feedback in desktop user interfaces, e.g. the software that comes with the Logitech iFeel mouse (www.logitech.com). This provides basic targeting: a brief pulse is played, for example, when a user moves over a target. We believe there is much more that can be presented with tactile feedback. 5.2 Visually impaired users Tactons will be able to work alongside Braille in tactile displays for blind and visually impaired users, in the same way as earcons work alongside synthetic speech. They will allow information to be delivered more efficiently. In addition, hierarchical Tactons could help users navigate Sine Sine Square Error Operating system error Execution error Sine Square Overflow Sine Square Underflow Sine Square Fast tempo Slow tempo Figure 5: Hierarchical Tacton composition. Level 1 Level 2 Level 3 20 around Braille media by providing navigation information (Brewster, 1998b). One of our main interests is in using Tactons to improve access to graphical information non-visually. Text can be rendered in a relatively straightforward manner by speech or Braille, but graphics are more problematic. One area that we and others have focused on is visualisation for blind people. Understanding and manipulating information using visualisations such as graphs, tables, bar charts and 3D plots is very common for sighted people. The skills needed are learned early in school and then used throughout life, for example, in analysing information or managing home finances. The basic skills needed for creating and manipulating graphs are necessary for all parts of education and employment. Blind people have very restricted access to information presented in these visual ways (Edwards, 1995). As Wise et al. (2001) say "Inac-cessibility of instructional materials, media, and technologies used in science, engineering, and mathematics education severely restricts the ability of students with little or no sight to excel in these disciplines". To allow blind people to gain the skills needed for the workplace new technologies are necessary to make visualisations usable. Tactons provide another route through which information can be presented. Research has shown that using haptic devices is an effective way of presenting graphical information non-visually (Yu and Brewster, 2003, Wies et al., 2001, Van Scoy et al., 2000). The most common approach has been to use haptic devices to present graphs, tables or 3D plots that users can feel kinaesthetically by tracing a line or shape with a finger using a device like the PHANToM (www.sensable.com). Lederman and Klatzky (1999) have shown that removal of cutaneous input to the fingertip impedes perception of edge direction, which is an essential component of tracing a haptic line graph. This lack of cutaneous stimulation leads to problems with navigation (exploring using a single point of contact means it is difficult to locate items as there is no context, which can be given in a tactile display), exploring small scale features (these would be perceived cutaneously on the finger pad in real life), and information overload (all haptic information is perceived kinaesthetically rather than being shared with cutaneous perception). Incorporating a tactile display into a force-feedback device will alleviate many of these problems and potentially increase user efficiency and comprehension of visualisations. Tactons could be presented as the user moves the force-feedback device over the visualisation. Dimensions of the data can be encoded into a Tacton to give information about the current point, using the parameters described in Section 4. This would allow more data to be presented more efficiently. For example, with multidimensional data one dimension might be mapped to the frequency of a pulse in a Tacton, another might map to rhythm and another to body locatoin. As the user moves about the data he/she would feel the different parameters. In addition to the finger pad, we can also include tactile displays to other parts of the body (e.g. to the back) using spatially distributed transducers to provide even more display area. As long as this is done in a comprehensible manner users will be able to gain access to their data in a much more effective way than with current force-feedback only visualisation tools. 5.3 Mobile and wearable devices Our other main application area is mobile and wearable device displays (for both sighted and blind people). Mobile telephones and handheld computers are currently one of the fastest growth areas of computing and this growth will extend into more sophisticated, fully wearable computers in the future. One problem with these devices is their limited output capabilities. Their small displays easily become cluttered with information and widgets and this makes interface design difficult. In addition, users are not always looking at the display of a device as they must walk or navigate through their environment which requires visual attention. One way to solve this problem is to use other display modalities and so reduce demands on visual display, or replace it if not available. Work has gone into using speech and non-speech sounds to overcome the display bottleneck. Tactile displays have great potential here too but are much less well investigated. Sound has many advantages but it can be problematic; in loud environments it can be impossible to hear auditory output from a device, in quiet places the audio may be disturbing to others nearby. Blind people often do not like to wear headphones when outdoors as they mask important environmental sounds. Tactile displays do not suffer from these problems (although there may be other problems for example, perceiving tactile stimuli whilst running due to the difficulties of keeping the transducers in contact with the skin). Mobile telephones commonly have a very simple point-contact tactile stimulator built-in that can alert the user to a call. These are often only able to produce pulses of different durations. A pin array would be possible on such a device as the user will be holding it in a hand when in use. Such a sophisticated tactile display could do much more, e.g. it could give information on the caller, replace or enhance items on the display (like icons, progress indicators, games) or aid in the navigation of the devices' menus so that the user does not need to look at the screen. In a wearable device users could have body mounted transducers so that information can be displayed over their body. In the simplest case this could be used to give directional information by vibrating one side of the body or other to indicate which way to turn (Tan and Pentland, 1997). A belt of transducers around the waist could give a compass-like display of direction; a pulse could be played continuously at north so the user can maintain orientation after turning (useful when navigating in the dark) or at the position around the waist corresponding to the direction in which to head. A more sophisticated display might give information about the user's context. For example, presenting Tactons describing information such as the type of building (shop, bank, office-block, house), the type of shop (clothes, phones, food, furniture) the price-bracket of a shop (budget, mid-range, expensive), or information more related to the concerns of visually impaired people, such as the number of stairs leading up to the entrance (for firefighters, whose vision is impaired 21 due to smoke and flames, a tactile display could also provide information on the location of rooms and exits in a burning building). A tactile display could also present information on stock market data (building on from the work on tactile visualisation in the section above) so that users could keep track of trades whilst away from the office. Such tactile displays could also work alongside auditory or visual ones. Future work and conclusions This paper has laid out some of the foundations of information display through Tactons. There is still much work to be done to fully understand how they should be designed and used. There are many lower level perceptual questions to be addressed before higher level design issues can be investigated. Many of the parameters of touch described in Section 4 are not fully understood and the full usable ranges of the parameters are not known. Studies need to be undertaken to explore the parameter space so that the relative importance of the different parameters can be discovered. Once the range of parameters is understood then the construction of Tactons can be examined. Basic studies are needed to understand how the parameters can be combined to construct Tactons. Parameters which work well alone may not work well when combined with others into a Tacton. For example, one parameter may mask another. When the basic design of Tactons is understood the composition of simple Tactons into more complex messages, encoding hierarchical information into Tactons, and their learnability and memorability can be investigated. The concurrent presentation of multiple Tactons must also be studied. These studies will answer some of the main questions regarding the usability of Tactons and a good understanding of their design and usability will have been a-chieved . Another important task is to investigate the strong relationships between hearing and touch by examining cross-modal uses of audio and tactile multimodal displays (Spence and Driver, 1997), e.g. combined audio and tactile cues, redundant tactile and audio cues, and moving from an audio to a tactile presentation of the same information (and vice versa). This is important in a mo-bile/wearable context because at different times different display techniques might be appropriate. For example, audio might be inappropriate in a very noisy environment , or tactile cues might be masked when the user is running. One important issue is to identify the types of information best presented in sound and those best presented tactually. For example, the range of the vibrotactile frequency response is roughly 20 times less than that of the auditory system. Such discrepancies must be accounted for when performing cross-modal mappings from hearing to touch. In conclusion, this paper has proposed a new form of tactile output called Tactons. These are structured tactile messages that can be used to communicate information. Tactile output is underused in current interfaces and Tactons provide a way of addressing this problem. The basic parameters have been described and design issues discussed . A technique is now available to allow tactile display to form a significant part of the set of interaction and display techniques that can be used to communicate with users at the interface. Acknowledgements This research was conducted when Brewster was on sabbatical in the Department of Computer Science at the University of Canterbury, Christchurch, New Zealand. Thanks to Andy Cockburn for his thoughts and comments on this work. The sabbatical was funded by an Erskine Fellowship from the University of Canterbury. The work was part funded by EPSRC grant GR/S53244. Brown is funded by an EPSRC studentship. References Akamatsu, M., MacKenzie, I. S. and Hasbrouq, T. (1995): A comparison of tactile, auditory, and visual feedback in a pointing task using a mouse-type device . Ergonomics, 38, 816-827. Barker, P. G. and Manji, K. A. (1989): Pictorial dialogue methods. International Journal of Man-Machine Studies , 31, 323-347. Blattner, M., Sumikawa, D. and Greenberg, R. (1989): Earcons and icons: Their structure and common design principles. Human Computer Interaction, 4, 11-44 . Brewster, S. A. (1998a): The design of sonically-enhanced widgets. Interacting with Computers, 11, 211-235. Brewster, S. A. (1998b): Using Non-Speech Sounds to Provide Navigation Cues. ACM Transactions on Computer-Human Interaction, 5, 224-259. Brewster, S. A. (2002): Chapter 12: Non-speech auditory output. In The Human Computer Interaction Handbook (Eds, Jacko, J. and Sears, A.) Lawrence Erlbaum Associates, pp. 220-239. Brewster, S. A., Wright, P. C. and Edwards, A. D. N. (1994): A detailed investigation into the effectiveness of earcons. In Auditory Display (Ed, Kramer, G.) Addison -Wesley, Reading, MA, pp. 471-498. Campbell, C., Zhai, S., May, K. and Maglio, P. (1999): What You Feel Must Be What You See: Adding Tactile Feedback to the Trackpoint. Proceedings of IFIP INTERACT'99, Edinburgh, UK, 383-390, IOS Press Challis, B. and Edwards, A. D. N. (2001): Design principles for tactile interaction. In Haptic Human-Computer Interaction, Vol. 2058 (Eds, Brewster, S. A. and Murray-Smith, R.) Springer LNCS, Berlin, Germany, pp. 17-24. Cholewiak, R. W. and Collins, A. (1996): Vibrotactile pattern discrimination and communality at several body sites. Perception and Psychophysics, 57, 724-737 . Craig, J. C. and Sherrick, C. E. (1982): Dynamic Tactile Displays. In Tactual Perception: A Sourcebook (Ed, Foulke, E.) Cambridge University Press, pp. 209-233. 22 Edwards, A. D. N. (Ed.) (1995) Extra-Ordinary Human-Computer Interaction, Cambridge University Press, Cambridge, UK. Enriquez, M. J. and Maclean, K. (2003): The Hapticon editor: A tool in support of haptic communication research . Haptics Symposium 2003, Los Angeles, CA, 356-362, IEEE Press Geldard, F. A. (1957): Adventures in tactile literacy. The American Psychologist, 12, 115-124. Gemperle, F., Kasabach, C., Stivoric, J., Bauer, M. and Martin, R. (1998): Design for wearability. Proceedings of Second International Symposium on Wearable Computers, Los Alamitos, CA, 116-122, IEEE Computer Society Gill, J. (2003), Vol. 2003 Royal National Institute of the Blind, UK. Gunther, E. (2001): Skinscape: A Tool for Composition in the Tactile Modality. Massachusetts Institute of Technology . Masters of Engineering. Gunther, E., Davenport, G. and O'Modhrain, S. (2002): Cutaneous Grooves: Composing for the Sense of Touch. Proceedings of Conference on New Instruments for Musical Expression, Dublin, IR, 1-6, Jansson, G. and Larsson, K. (2002): Identification of Haptic Virtual Objects with Differing Degrees of Complexity. Proceedings of Eurohaptics 2002, Edinburgh , UK, 57-60, Edinburgh University Kaczmarek, K., Webster, J., Bach-y-Rita, P. and Tomp-kins , W. (1991): Electrotacile and vibrotactile displays for sensory substitution systems. IEEE Transaction on Biomedical Engineering, 38, 1-16. Kurze, M. (1997): Rendering drawings for interactive haptic perception. Proceedings of ACM CHI'97, Atlanta , GA, 423-430, ACM Press, Addison-Wesley Kurze, M. (1998): TGuide: a guidance system for tactile image exploration. Proceedings of ACM ASSETS '98, Marina del Rey, CA, ACM Press Lederman, S. J. and Klatzky, R. L. (1999): Sensing and Displaying Spatially Distributed Fingertip Forces in Haptic Interfaces for Teleoperator and Virtual Environment Systems. Presence: Teleoperators and Virtual Environments, 8, 86-103. Montagu, A. (1971): Touching: The Human Significance of the Skin, Columbia University Press, New York. Oakley, I., McGee, M., Brewster, S. A. and Gray, P. D. (2000): Putting the feel in look and feel. Proceedings of ACM CHI 2000, The Hague, Netherlands, 415-422, ACM Press, Addison-Wesley Rupert, A. (2000): Tactile situation awareness system: proprioceptive prostheses for sensory deficiencies. Aviation, Space and Environmental Medicine, 71, 92-99 . Sherrick, C. (1985): A scale for rate of tactual vibration. Journal of the Acoustical Society of America, 78. Shneiderman, B. (1998): Designing the user interface, 3 rd Ed. Addison-Wesley, Reading (MA). Spence, C. and Driver, J. (1997): Cross-modal links in attention between audition, vision and touch: implications for interface design. International Journal of Cognitive Ergonomics, 1, 351-373. Stone, R. (2000): Haptic feedback: A potted history, from telepresence to virtual reality. The First International Workshop on Haptic Human-Computer Interaction, Glasgow, UK, 1-7, Springer-Verlag Lecture Notes in Computer Science Summers, I. R., Chanter, C. M., Southall, A. L. and Brady, A. C. (2001): Results from a Tactile Array on the Fingertip. Proceedings of Eurohaptics 2001, Birmingham , UK, 26-28, University of Birmingham Tan, H. Z. and Pentland, A. (1997): Tactual Displays for Wearable Computing. Proceedings of the First International Symposium on Wearable Computers, IEEE Tan, H. Z. and Pentland, A. (2001): Chapter 18: Tactual displays for sensory substitution and wearable computers . In Fundamentals of wearable computers and augmented reality (Eds, Barfield, W. and Caudell, T.) Lawrence Erlbaum Associates, Mahwah, New Jersey, pp. 579-598. van Erp, J. B. F. (2002): Guidelines for the use of active vibro-tactile displays in human-computer interaction. Proceedings of Eurohaptics 2002, Edinburgh, UK, 18-22, University of Edinburgh Van Scoy, F., Kawai, T., Darrah, M. and Rash, C. (2000): Haptic Display of Mathematical Functions for Teaching Mathematics to Students with Vision Disabilities: Design and Proof of Concept. Proceedings of the First Workshop on Haptic Human-Computer Interaction, Glasgow, UK, University of Glasgow van Veen, H. and van Erp, J. B. F. (2001): Tactile information presentation in the cockpit. In Haptic Human-Computer Interaction (LNCS2058), Vol. 2058 (Eds, Brewster, S. A. and Murray-Smith, R.) Springer, Berlin , Germany, pp. 174-181. Wall, S. A. and Harwin, W. S. (2001): A High Bandwidth Interface for Haptic Human Computer Interaction. Mechatronics. The Science of Intelligent Machines. An International Journal, 11, 371-387. Wall, S. A., Riedel, B., Crossan, A. and McGee, M. R. (Eds.) (2002) Eurohaptics 2002 Conference Proceedings , University of Edinburgh, Edinburgh, Scotland. Wies, E., Gardner, J., O'Modhrain, S., Hasser, C. and Bulatov, V. (2001): Web-based touch display for accessible science education. In Haptic Human-Computer Interaction, Vol. 2058 (Eds, Brewster, S. A. and Murray-Smith, R.) Springer LNCS, Berlin, pp. 52-60.

tactile displays;multimodal interaction;Tactons;non-visual cues

186

TCP/IP Performance over 3G Wireless Links with Rate and Delay Variation

Wireless link losses result in poor TCP throughput since losses are perceived as congestion by TCP, resulting in source throttling. In order to mitigate this effect, 3G wireless link designers have augmented their system with extensive local retransmission mechanisms. In addition, in order to increase throughput, intelligent channel state based scheduling have also been introduced. While these mechanisms have reduced the impact of losses on TCP throughput and improved the channel utilization, these gains have come at the expense of increased delay and rate variability. In this paper, we comprehensively evaluate the impact of variable rate and variable delay on long-lived TCP performance. We propose a model to explain and predict TCP's throughput over a link with variable rate and/or delay. We also propose a network-based solution called Ack Regulator that mitigates the effect of variable rate and/or delay without significantly increasing the round trip time, while improving TCP performance by up to 40%.

INTRODUCTION Third generation wide-area wireless networks are currently being deployed in the United States in the form of 3G1X technology [10] with speeds up to 144Kbps. Data-only enhancements to 3G1X have already been standardized in the 3G1X-EVDO standard (also called High Data Rate or HDR) with speeds up to 2Mbps [6]. UMTS [24] is the third generation wireless technology in Europe and Asia with deploy-ments planned this year. As these 3G networks provide pervasive internet access, good performance of TCP over these wireless links will be critical for end user satisfaction. While the performance of TCP has been studied extensively over wireless links [3, 4, 15, 20], most attention has been paid to the impact of wireless channel losses on TCP. Losses are perceived as congestion by TCP, resulting in source throttling and very low net throughput. In order to mitigate the effects of losses, 3G wireless link designers have augmented their system with extensive local retransmission mechanisms. For example, link layer retransmission protocols such as RLP and RLC are used in 3G1X [22] and UMTS [21], respectively. These mechanisms ensure packet loss probability of less than 1% on the wireless link, thereby mitigating the adverse impact of loss on TCP. While these mechanisms mitigate losses, they also increase delay variability. For example, as we shall see in Section 3, ping latencies vary between 179ms to over 1 second in a 3G1X system. In addition, in order to increase throughput, intelligent channel state based scheduling have also been introduced. Channel state based scheduling [7] refers to scheduling techniques which take the quality of wireless channel into account while scheduling data packets of different users at the base station. The intuition behind this approach is that since the channel quality varies asynchronously with time due to fading, it is preferable to give priority to a user with better channel quality at each scheduling epoch. While strict priority could lead to starvation of users with inferior channel quality, a scheduling algorithm such as proportional fair [6] can provide long-term fairness among different users. However, while channel-state based scheduling improves overall throughput, it also increases rate variability. Thus, while the impact of losses on TCP throughput have been significantly reduced by local link layer mechanisms and higher raw throughput achieved by channel-state based scheduling mechanisms, these gains have come at the expense of increased delay and rate variability. This rate and delay variability translates to bursty ack arrivals (also called ack compression) at the TCP source. Since TCP uses ack clocking to probe for more bandwidth, bursty ack arrival leads to release of a burst of packets from the TCP source. When this burst arrives at a link with variable rate or delay , it could result in multiple packet losses. These multiple losses significantly degrade TCPs throughput. In this paper, we make three main contributions. First, 71 we comprehensively evaluate the impact of variable rate and variable delay on long-lived TCP performance. Second, we propose a model to explain and predict TCP's throughput over a link with variable rate and/or delay. Third, we propose a network-based solution called Ack Regulator that mitigates the effect of variable rate and/or delay without significantly increasing the round trip time, thereby improving TCP performance. The remaining sections are organized as follows. In Section 2, we discuss related work. In Section 3, we present the motivation for our work using traces from a 3G1X system. In Section 4, we describe a model for computing the throughput of a long-lived TCP flow over links with variable rate and variable delay. We then present a simple network-based solution, called Ack Regulator, to mitigate the effect of variable rate/delay in Section 5. In Section 6, we present extensive simulation results that compare TCP performance with and without Ack Regulator, highlighting the performance gains using the Ack Regulator when TCP is subjected to variable rate and delay. Finally, in Section 7, we present our conclusions. RELATED WORK In this section, we review prior work on improving TCP performance over wireless networks. Related work on the modeling of TCP performance is presented in Section 4. A lot of prior work has focused on avoiding the case of a TCP source misinterpreting packet losses in the wireless link as congestion signals. In [4], a snoop agent is introduced inside the network to perform duplicate ack suppression and local retransmissions on the wireless link to enhance TCP performance. In [3], the TCP connection is split into two separate connections, one over the fixed network and the second over the wireless link. The second connection can recover from losses quickly, resulting in better throughput. Link-layer enhancements for reducing wireless link losses including retransmission and forward error correction have been proposed in [20]. Link layer retransmission is now part of both the CDMA2000 and UMTS standards [10, 24]. In order to handle disconnections (a case of long-lived loss), M-TCP has been proposed [8]. The idea is to send the last ack with a zero-sized receiver window so that the sender can be in persist mode during the disconnection. Link failures are also common in Ad Hoc networks and techniques to improve TCP performance in the presence of link failures have been proposed in [11]. Note that none of these approaches address specifically the impact of delay and rate variation on TCP, which is the focus of this paper. Several generic TCP enhancements with special applica-bility to wireless links are detailed in [12, 13]. These include enabling the Time Stamp option, use of large window size and window scale option, disabling Van Jacobson header compression, and the use of Selective Acknowledgments (Sack). Large window size and window scaling are necessary because of the large delay of wireless link while Sack could help TCP recover from multiple losses without the expensive timeout recovery mechanism. Another issue with large delay variation in wireless links is spurious timeouts where TCP unnecessarily retransmits a packet (and lowers its congestion window to a minimum) after a timeout, when the packet is merely delayed. In [13], the authors refer to rate variability due to periodic allocation and de-allocation of high-speed channels in 3G networks as Bandwidth Oscillation. Bandwidth Oscillation can also lead to spurious timeouts in TCP because as the rate changes from high to low, the rtt value increases and a low Retransmission Timeout (RTO) value causes a spurious retransmission and unnecessarily forces TCP into slow start. In [15], the authors conduct experiments of TCP over GSM circuit channels and show that spurious timeouts are extremely rare. However, 3G wireless links can have larger variations than GSM due to processing delays and rate variations due to channel state based scheduling. Given the increased variability on 3G packet channels, the use of TCP time stamp option for finer tracking of TCP round trip times and possibly the use of Eifel retransmission timer [16] instead of the conventional TCP timer can help avoid spurious timeouts. As mentioned earlier, the effect of delay and rate variability is ack compression and this results in increased burstiness at the source. Ack compression can also be caused by bidirectional flows over regular wired networks or single flow over networks with large asymmetry. This phenomenon has been studied and several techniques have been proposed to tackle the burstiness of ack compression. In order to tackle burstiness, the authors in [18] propose several schemes that withholds acks such that there is no packet loss at the bottleneck router, resulting in full throughput. However, the round trip time is unbounded and can be very large. In [23], the authors implement an ack pacing technique at the bottleneck router to reduce burstiness and ensure fairness among different flows. In the case of asymmetric channels , solutions proposed [5] include ack congestion control and ack filtering (dropping of acks), reducing source burstiness by sender adaptation and giving priority to acks when scheduling inside the network. However, the magnitude of asymmetry in 3G networks is not large enough and can be tolerated by TCP without ack congestion control or ack filtering according to [12]. Note that, in our case, ack compression occurs because of link variation and not due to asymmetry or bidirectional flows. Thus, we require a solution that specifically adapts to link variation. Moreover, the node at the edge of the 3G wireless access network is very likely to be the bottleneck router (given rates of 144Kbps to 2Mbps on the wireless link) and is the element that is exposed to varying delays and service rates. Thus, this node is the ideal place to regulate the acks in order to improve TCP performance. This is discussed in more detail in the next section. MOTIVATION MD RNC RNC BS BS PDSN/ SGSN BS: Base Station MD: Mobile Device HA/ GGSN HA: Home Agent RNC: Radio Network Controller PDSN: Packet Data Service Node SGSN: Serving GPRS Service Node GGSN: Gateway GPRS Service Node (RLP/RLC) Link Layer Retransmission INTER NET Figure 1: 3G network architecture A simplified architecture of a 3G wireless network is shown 72 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 200 400 600 800 1000 1200 1400 Prob. Ping latency (ms) Figure 2: CDF of Ping Latencies in Figure 1. The base stations are connected to a node called the Radio Network Controller (RNC). The RNC performs CDMA specific functions such as soft handoffs, encryption, power control etc. It also performs link layer retransmission using RLP(RLC) in 3G1X(UMTS) system. In the 3G1X system, the RNC is connected to a PDSN using a GRE tunnel (one form of IP in IP tunnel) and the PDSN terminates PPP with the mobile device. If Mobile IP service is enabled, the PDSN also acts as a Foreign Agent and connects to a Home Agent. In the UMTS system, the RNC is connected to a SGSN using a GTP tunnel (another form of IP in IP tunnel); the SGSN is connected to a GGSN, again through a GTP tunnel. Note that the tunneling between the various nodes allows for these nodes to be connected directly or through IP/ATM networks. In this architecture, the RNC receives a PPP/IP packet through the GRE/GTP tunnel from the PDSN/SGSN. The RNC fragments this packet into a number of radio frames and then performs transmission and local retransmission of these radio frames using the RLP(RLC) protocol. The base station (BS) receives the radio frames from the RNC and then schedules the transmission of the radio frames on the wireless link using a scheduling algorithm that takes the wireless channel state into account. The mobile device receives the radio frames and if it discovers loss of radio frames, it requests local retransmission using the RLP(RLC) protocol . Note that, in order to implement RLP(RLC), the RNC needs to keep a per-user queue of radio frames. The RNC can typically scale up to tens of base stations and thousands of active users. In order to illustrate the variability seen in a 3G system, we obtained some traces from a 3G1X system. The system consisted of an integrated BS/RNC, a server connected to the RNC using a 10Mbps Ethernet and a mobile device connected to the BS using a 3G1X link with 144Kbps downlink in infinite burst mode and 8Kbps uplink. The infinite burst mode implies that the rate is fixed and so the system only had delay variability. Figure 2 plots the cumulative distribution function (cdf) of ping latencies from a set of 1000 pings from the server to the mobile device (with no observed loss). While about 75% of the latency values are below 200ms, the latency values go all the way to over 1s with about 3% of the values higher than 500ms. In the second experiment, a TCP source at the server using Sack with timestamp option transferred a 2MB file to the mobile device. The MTU was 1500 bytes with user data size of 1448 bytes. The buffer at the RNC was larger than the TCP window size 1 . and thus, the transfer resulted in no TCP packet loss and a maximal throughput of about 1 We did not have control over the buffer size at the RNC in 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Prob. Interack time Time (s) 0 0.5 1 1.5 2 2.5 3 3.5 0 20 40 60 80 100 120 Rtt (s) Time (s) (a) TCP Ack Inter-arrival (b) TCP rtt value Figure 3: 3G Link Delay Variability 135Kb/s. The transmission time at the bottleneck link is 1.448 8/135 = 86ms. If the wireless link delay were constant , the TCP acks arriving at the source would be evenly spaced with a duration of 172ms because of the delayed ack feature of TCP (every 2 packets are acked rather than every packet). Figure 3(a) plots the cdf of TCP ack inter-arrival time (time between two consecutive acks) at the server. As can be seen, there is significant ack compression with over 10% of the acks arriving within 50ms of the previous ack. Note that the ack packet size is 52 bytes (40 + timestamp) and ack transmission time on the uplink is 52 8/8=52ms; an interack spacing of less then 52ms is a result of uplink delay variation. Note that the delay variability and the resulting ack compression did not cause any throughput degradation in our system. This was due to the fact that the buffering in the system was greater than the TCP window size resulting in no buffer overflow loss. Figure 3(b) depicts the TCP round trip time (rtt) values over time. Since the buffer at the RNC is able to accommodate the whole TCP window, the rtt increases to over 3s representing a case of over 30 packets in the buffer at the RNC (30 0.086 = 2.5s). Given an average ping latency of 215ms and a transmission time of 86ms for a 1500 byte packet, the bandwidth delay product of the link is approximately (0.215 + 0.86) 135=5KB or about 3 packets. Thus, the system had a buffer of over 10 times the bandwidth delay product. Given that we had only one TCP flow in the system, a buffer of over 64KB is not a problem. But, if every TCP flow is allocated a buffer of 64KB, the buffer requirements at the RNC would be very expensive, since the RNC supports thousands of active users. Even discounting the cost of large buffers, the inflated rtt value due to the excessive buffering has several negative consequences as identified in [15]. First, an inflated rtt implies a large retransmission timeout value (rto). In the case of multiple packet losses (either on the wireless link or in a router elsewhere in the network), a timeout-based recovery would cause excessive delay, especially if exponential backoff gets invoked. Second, if the timestamp option is not used, the rtt sampling rate is reduced and this can cause spurious timeouts. Third, there is a higher probability that the data in the queue becomes obsolete (for e.g., due to user aborting the transmission), but the queue will still have to be drained resulting in wasted bandwidth. Thus, while excessive buffering at the RNC can absorb the variability of the wireless links without causing TCP throughput degradation, it has significant negative side effects , making it an undesirable solution. our system. 73 MODEL In this section, we model the performance of a single long-lived TCP flow over a network with a single bottleneck server that exhibits rate variation based on a given general distribution and a single wireless link attached to the bottleneck server that exhibits delay variation based on another given distribution. We use a general distribution of rate and delay values for the discussion in this section since we would like to capture the inherent variation in rate and delay that is a characteristic of the 3G wireless data environment. Given that the wireless standards are constantly evolving, the actual rate and delay distribution will vary from one standard or implementation to another and is outside the scope of this paper. Later, in Section 6, we will evaluate TCP performance over a specific wireless link, the 3G1X-EVDO (HDR) system, using simulation. We would like to model TCP performance in the case of variable rate and delay for two reasons. One, we would like to understand the dynamics so that we can design an appropriate mechanism to improve TCP performance. Two, we would like to have a more accurate model that specifically takes the burstiness caused by ack compression due to rate/delay variability into account. TCP performance modeling has been extensively studied in the literature [1, 2, 9, 14, 17, 19]. Most of these models assume constant delay and service rate at the bottleneck router and calculate TCP throughput in terms of packet loss probability and round trip time. In [19], the authors model TCP performance assuming deterministic time between congestion events [1]. In [17], the authors improve the throughput prediction of [19] assuming exponential time between congestion events (loss indications as Poisson). In our case, ack compressions and link variation causes bursty losses and the deterministic or Poisson loss models are not likely to be as accurate. In [9], the authors model an UMTS wireless network by extending the model from [19] and inflating the rtt value to account for the average additional delay incurred on the wireless link. However, we believe this will not result in an accurate model because 1) the rtt value in [19] is already an end-to-end measured value and 2) the loss process is much more bursty than the deterministic loss assumption in [19]. In [2], the authors observe that mean values are not sufficient to predict the throughput when routers have varying bandwidth and show that increasing variance for the same mean service rate decreases TCP throughput. However , the approach is numerical, and provides little intuition in the case of delay variance. Our approach starts with the model in [14] which describes how TCP functions in an "ideal" environment with constant round trip time, constant service rate and suffers loss only through buffer overflow. A brief summary of the result from [14] is presented here before we proceed to our model, which can be seen as an extension. We chose to extend the model in [14] since it makes no assumption about the nature of loss event process (which is highly bursty in our case) and explicitly accounts for link delay and service rate (which are variable in our case). For simplicity, we will only discuss the analysis of TCP Reno. TCP Sack can be analyzed similarly. We also assume that the sender is not limited by the maximum receiver window; simple modifications can be made to the analysis for handling this case. Figure 4(a) shows how the TCP congestion window varies 0 5 10 15 20 25 30 35 40 45 0 50 100 150 200 Ideal TCP 0 5 10 15 20 25 30 35 40 45 50 0 50 100 150 200 TCP with Variable Delay (a) Constant delay (b) Variable delay Figure 4: TCP Congestion Window Evolution over time in a constant rate and delay setting. The initial phase where TCP tries to probe for available bandwidth is the Slow Start phase. After slow start, TCP goes to Congestion avoidance phase. In the case of long-lived TCP flow, one can focus only on the congestion avoidance phase. Let be the constant service rate, the constant propagation delay, T the minimum round trip time ( + 1/) and B the buffer size. The congestion window follows a regular saw-tooth pattern, going from W 0 to W max , where W 0 = W max /2 and W max = + B + 1. Due to the regularity of each of the saw-tooth, consider one such saw-tooth. Within a single saw-tooth, the congestion avoidance phase is divided into two epochs. In the first epoch, say epoch A, the congestion window increases from W 0 to T , in time t A with number of packets sent n A . In the second epoch, say epoch B, the congestion window increases from T to W max , in time t B with number of packets sent n B . TCP throughput (ignoring slow start) is simply given by (n A + n B )/(t A + t B ) where t A = T (T - W 0 ) (1) n A = (W 0 t A + t 2 A /(2T ))/T (2) t B = (W 2 max - (T ) 2 )/(2) (3) n B = t B (4) This model, while very accurate for constant and T , breaks down when the constant propagation and service rate assumptions are not valid. Figure 4(b) shows how the congestion window becomes much more irregular when there is substantial variation in the wireless link delay. This is because the delay variation and ack compression result in multiple packet losses. There are three main differences in the TCP congestion window behavior under variable rate/delay from the traditional saw-tooth behavior. First, while the traditional saw-tooth behavior always results in one packet loss due to buffer overflow, we have possibilities for multiple packet losses due to link variation. To account for this, we augment our model with parameters p1, p2, p3 representing respectively the conditional probability of a single packet loss, double packet loss, and three or more packet losses. Note that, p1 + p2 + p3 = 1 by this definition. Second, while the loss in the traditional saw-tooth model always occurs when window size reaches W max = + B + 1, in our model losses can occur at different values of window size, since and are now both variables instead of constants. We capture this by a parameter W f = N i=1 W 2 max i /N , that is the square root of the second moment of the W max values of each cycle. The reason we do this instead of obtaining a simple mean of W max values is because throughput is related to W f quadratically (since it is the area under the curve in the 74 0 5 10 15 20 25 128 130 132 134 136 138 140 142 cwnd (packets) Time (s) 0 5 10 15 20 25 100 105 110 115 cwnd (packets) Time (s) (a) Two packet loss (b) Three packet loss Figure 5: Congestion Window with multiple losses congestion window graph). Third, due to the fact that we have multiple packet losses in our model, we need to consider timeouts and slow starts in our throughput calculation. We represent the timeout duration by the T 0 parameter which represents the average timeout value, similar to the timeout parameter in [19]. We now model the highly variable congestion window behavior of a TCP source under rate/delay variation. We first use W f instead of W max . We approximate (the propagation delay) by ^ , the average link delay in the presence of delay variability. We replace (the service rate) by ^ , the average service rate in the presence of rate variability. Thus, T becomes ^ T = (^ + 1 / ^ ). Now consider three different congestion window patterns: with probability p1, single loss followed by congestion avoidance, with probability p2, double loss followed by congestion avoidance, and with probability p3, triple loss and timeout followed by slow start and congestion avoidance 2 . First, consider the single loss event in the congestion avoidance phase. This is the classic saw-tooth pattern with two epochs as identified in [14]. Lets call these A1 and B1 epochs. In epoch A1, window size grows from W 01 to ^ ^ T in time, t A1 , with number of packets transmitted, n A1 . In epoch B1, window size grows from ^ ^ T to W f in time, t B1 , with number of packets transmitted, n B1 . Thus, with probability p1, n A1 +n B1 packets are transmitted in time t A1 +t B1 where W 01 = (int)W f /2 (5) t A1 = ^ T (^ ^ T - W 01 ) (6) n A1 = (W 01 t A1 + t 2 A1 /(2 ^ T ))/ ^ T (7) t B1 = (W 2 f - (^ ^ T ) 2 )/(2^ ) (8) n B1 = ^ t B1 (9) Next, consider the two loss event. An example of this event is shown in Figure 5(a). The trace is obtained using ns-2 simulation described in Section 6. In this case, after the first fast retransmit (around 130s), the source receives another set of duplicate acks to trigger the second fast retransmit (around 131s). This fixes the two losses and the congestion window starts growing from W 02 . The second retransmit is triggered by the new set of duplicate acks in response to the first retransmission. Thus, the duration between the first and second fast retransmit is the time re-quired for the first retransmission to reach the receiver (with a full buffer) plus the time for the duplicate ack to return 2 We assume that three or more packet losses result in a timeout; this is almost always true if the source is TCP reno. to the sender. In other words, this duration can be approximated by the average link delay with a full buffer, ^ T + B/^ =t R . We have three epochs now, epoch t R (time 130-131s )with one retransmission and zero new packet, epoch A2 (131-137s) with window size growing from W 02 to ^ ^ T in time, t A2 , with number of packets transmitted, n A2 , and epoch B1 (137-143s) as before. Thus, with probability p2, n A2 +n B1 packets are transmitted in time t R +t A2 +t B1 where W 02 = (int)W 01 /2 (10) t R = ^ T + B/^ (11) t A2 = ^ T (^ ^ T - W 02 ) (12) n A2 = (W 02 t A2 + t 2 A2 /(2 ^ T ))/ ^ T (13) Finally, consider the three loss event. An example of this event is shown in Figure 5(b). In this case, after the first fast retransmit, we receive another set of duplicate acks to trigger the second fast retransmit. This does not fixthe three losses and TCP times out. Thus, we now have five epochs: first is the retransmission epoch (100-101s) with time t R and zero new packet, second is the timeout epoch (101-103s) with time T 0 and zero new packet, third is the slow start epoch (103-106s) where the window grows exponentially up to previous ssthresh value of W 03 in time t ss (Eqn. 15) with number of packets transmitted n ss (Eqn. 16) 3 , fourth is epoch A3 (106-111s) where the window size grows from W 03 to ^ ^ T in time t A3 (Eqn. 17) with number of packets transmitted n A3 (Eqn. 18), and fifth is epoch B1 (111-118s) as before. Thus, with probability p3, n ss +n A3 +n B1 packets are transmitted in time t R +T 0 +t ss +t A2 +t B1 where W 03 = (int)W 02 /2 (14) t ss = ^ T log 2 (W 03 ) (15) n ss = W 03 / ^ T (16) t A3 = ^ T (^ ^ T - W 03 ) (17) n A3 = (W 03 t A3 + t 2 A3 /(2 ^ T ))/ ^ T (18) Given that the different types of packet loss events are independent and using p1+p2+p3=1, the average TCP throughput can now be approximated by a weighted combination of the three types of loss events to be p3 (n ss + n A3 ) + p2 n A2 + p1 n A1 + n B1 p3 (t R + T 0 + t ss + t A3 ) + p2 (t R + t A2 ) + p1 t A1 + t B1 (19) If any of t are less than 0, those respective epochs do not occur and we can use the above equation while setting the respective n , t to zero. In this paper, we infer parameters such as p1, p2, p3, W f , and T 0 from the traces. Models such as [19] also infer the loss probability, round trip time, and timeout durations from traces. Table 1 lists the various parameters used by the different models for simulations with rate and delay variability. We use a packet size of 1000 bytes, a buffer of 10 which represents the product of the average bandwidth times average delay and we ensure that the source is not window limited. T D and T O denote the number of loss events that are of the triple duplicate and timeout type respectively and these values are used by models in [19] and [17]. The simulation 3 using analysis similar to [14] and assuming adequate buffer so that there is no loss in slow start. 75 Item Rate(Kb/s) Delay(ms) pkts TD TO T 0 rtt p1 p2 W f ^ T ^ 1 200 400 89713 401 1 1.76 616.2 0.998 0.000 22.00 440 25.0 2 200 380+e(20) 83426 498 1 1.71 579.3 0.639 0.357 21.38 442 25.0 3 200 350+e(50) 78827 489 12 1.79 595.8 0.599 0.367 21.24 461 25.0 4 200 300+e(100) 58348 496 114 1.92 606.0 0.339 0.279 18.95 517 25.0 5 u(200,20) 400 82180 504 1 1.75 578.1 0.535 0.460 21.61 400 24.74 6 u(200,50) 400 74840 517 29 1.80 579.9 0.510 0.403 20.52 400 23.34 7 u(200,75) 400 62674 516 81 1.86 585.9 0.398 0.348 19.05 400 20.93 8 u(200,50) 350+e(50) 70489 507 43 1.81 595.7 0.496 0.377 20.15 459 23.34 9 u(200,75) 300+e(100) 53357 497 93 2.03 635.7 0.404 0.298 17.78 511 20.93 Table 1: Simulation and Model parameters Item Simulator Goodput Model 1 [19] (accu.) Model 2 [17] (accu.) Model 3[Eqn. 19] (accu.) 1 199.8 228.5(0.86) 201.9(0.99) 199.8(1.0) 2 185.4 208.0(0.88) 186.0(1.0) 186.0(1.0) 3 175.1 195.5(0.88) 177.2(0.99) 180.9(0.97) 4 129.4 145.3(0.88) 153.7(0.81) 137.0(0.94) 5 182.5 205.2(0.88) 184.6(0.99) 181.3(0.99) 6 166.2 186.0(0.88) 174.6(0.95) 165.2(0.99) 7 139.2 158.4(0.86) 163.4(0.83) 137.2(0.99) 8 156.5 174.6(0.88) 166.5(0.94) 160.2(0.97) 9 118.4 134.0(0.87) 142.6(0.80) 125.0(0.94) Table 2: Simulation and Model throughput values is run for 3600 seconds. We simulate delay and rate variability with exponential and uniform distributions respectively (u(a, b) in the table represents uniform distribution with mean a and standard deviation b while e(a) represents an exponential distribution with mean a). The details of the simulation are presented in Section 6. Table 2 compares the throughput of simulation of different distributions for rate and delay variability at the server and the throughput predicted by the exact equation of the model in [19], the Poisson model in [17] and by equation 19. The accuracy of the prediction, defined as 1 minus the ratio of the difference between the model and simulation throughput value over the simulation throughput value, is listed in the parenthesis. As the last column shows, the match between our model and simulation is extremely accurate when the delay/rate variation is small and the match is still well over 90% even when the variation is large. The Poisson loss model used in [17] performs very well when the variability is low but, understandably, does not predict well when variability increases. The deterministic loss model seems to consistently overestimate the throughput. From Table 1, one can clearly see the impact of delay and rate variability. As the variability increases, the probability of double loss, p2, and three or more losses, p3=(1-p2-p1), start increasing while the goodput of the TCP flow starts decreasing. For example, comparing case 1 to case 4, p1 decreases from 0.998 to 0.339 while p3 increases. Increases in p2 and p3 come about because when the product ^ T ^ decreases, a pipe that used to accommodate more packets suddenly becomes smaller causing additional packet losses. Given that n A1 /t A1 &gt; n A2 /(t R + t A2 ) &gt; (n ss + n A3 )/(t R + T 0 +t ss +t A3 ), any solution that improves TCP performance must reduce the occurrence of multiple packet losses, p2 and p3. We present a solution that tries to achieve this in the next section. ACK REGULATOR In this section, we present our network-based solution for improving TCP performance in the presence of varying bandwidth and delay. The solution is designed for improving the performance of TCP flows towards the mobile host (for downloading-type applications) since links like HDR are designed for such applications. The solution is implemented at the wireless edge, specifically at the RNC, at the layer just above RLP/RLC. Note that, in order to implement the standard-based RLP/RLC, the RNC already needs to maintain a per-user queue. Our solution requires a per-TCP-flow queue, which should not result in significant additional overhead given the low bandwidth nature of the wireless environment . We also assume that the data and ack packets go through the same RNC; this is true in the case of 3G networks where the TCP flow is anchored at the RNC because of the presence of soft handoff and RLP. We desire a solution that is simple to implement and remains robust across different implementations of TCP. To this end, we focus only on the congestion avoidance phase of TCP and aim to achieve the classic saw-tooth congestion window behavior even in the presence of varying rates and delays by controlling the buffer overflow process in the bottleneck link. We also assume for this discussion that every packet is acknowledged (the discussion can be easily modified to account for delayed acks where single ack packets acknowledge multiple data packets). Our solution is called the Ack Regulator since it regulates the flow of acks back to the TCP source. The intuition behind the regulation algorithm is to avoid any buffer overflow loss until the congestion window at the TCP source reaches a pre-determined threshold and beyond that, allow only a single buffer overflow loss. This ensures that the TCP source operates mainly in the congestion avoidance phase with congestion window exhibiting the classic saw-tooth behavior. Before we present our solution, we describe two variables that will aid in the presentation of our solution. ConservativeMode: Mode of operation during which 76 DataSeqNoLast (DL): Largest Sequence # of Last Data Packet Received DataSeqNoFirst (DF): Largest Sequence # of Last Data Packet Sent AckSeqNoLast (AL): Largest Sequence # of Last Ack Packet Received AckSeqNoFirst (AF): Largest Sequence # of Last Ack Packet Sent DL DF AF AL Per-Flow Data and Ack Queue on RNC Wireless Network Wireline Network Data Queue Ack Queue QueueLength QueueLim Figure 6: Ack Regulator Implementation each time an ack is sent back towards the TCP source, there is buffer space for at least two data packets from the source. Note that if TCP operates in the congestion avoidance phase, there would be no buffer overflow loss as long as the algorithm operates in conservative mode. This follows from the fact that, during congestion avoidance phase, TCP increases its window size by at most one on reception of an ack. This implies that on reception of an ack, TCP source sends either one packet (no window increase) or two packets (window increase). Therefore, if there is space for at least two packets in the buffer at the time of an ack being sent back, there can be no packet loss. AckReleaseCount: The sum of total number of acks sent back towards the source and the total number of data packets from the source in transit towards the RNC due to previous acks released, assuming TCP source window is constant. AckReleaseCount represents the number of packets that can be expected to arrive in the buffer at the RNC assuming that the source window size remains constant. Thus, buffer space equal to AckReleaseCount must be reserved whenever a new ack is sent back to the source if buffer overflow is to be avoided. On Enque of Ack/Deque of data packet: 1. AcksSent=0; 2. BufferAvail=QueueLim-QueueLength; 3. BufferAvail-=(AckReleaseCount+ConservativeMode); 4. if (BufferAvail&gt;=1) 5. if (AckSeqNoLast-AckSeqNoFirst&lt;BufferAvail) 5.1 AcksSent+=(AckSeqNoLast-AckSeqNoFirst); 5.2 AckSeqNoFirst=AckSeqNoLast; else 5.3 AckSeqNoFirst+=BufferAvail; 5.4 AcksSent+=BufferAvail; 5.5 Send acks up to AckSeqNoFirst; Figure 7: Ack Regulator processing at the RNC Figure 6 shows the data and ack flow and the queue variables involved in the Ack Regulator algorithm, which is presented in Figure 7. We assume for now that the AckReleaseCount and ConservativeMode variables are as defined earlier. We later discuss how these variables are updated. The Ack Regulator algorithm runs on every transmission of a data packet (deque) and every arrival of an ack packet (enque). The instantaneous buffer availability in the data queue is maintained by the BufferAvail variable (line 2). BufferAvail is then reduced by the AckReleaseCount and the ConservativeMode variables (line 3). Depending on the value of the ConservativeMode variable (1 or 0), the algorithm operates in two modes, a conservative mode or a non-conservative, respectively. In the conservative mode, an extra buffer space is reserved in the data queue to ensure that there is no loss even if TCP congestion window is increased by 1, while, in the non-conservative mode, a single packet loss occurs if TCP increases its congestion window by 1. Now, after taking AckReleaseCount and ConservativeMode variables into account, if there is at least one buffer space available (line 4) and, if the number of acks present in the ack queue (AckSeqNoLast - AckSeqNoFirst) is lesser than BufferAvail, all those acks are sent to the source (lines 5.1,5.2); otherwise only BufferAvail number of acks are sent to the source (lines 5.3,5.4). Note that the actual transmission of acks (line 5.5) is not presented here. The transmission of AcksSent acks can be performed one ack at a time or acks can be bunched together due to the cumulative nature of TCP acks. However, care must be taken to preserve the duplicate acks since the TCP source relies on the number of duplicate acks to adjust its congestion window. Also, whenever three or more duplicate acks are sent back, it is important that one extra buffer space be reserved to account for the fast retransmission algorithm. Additional buffer reservations of two packets to account for the Limited Transmit algorithm [12] can also be provided for, if necessary. 1. Initialize ConservativeMode=1; = 2 2. On Enque of ack packet: if ((DataSeqNoLast-AckSeqNoFirst)&gt;*QueueLim) ConservativeMode=0; 3. On Enque and Drop of data packet: Conservative Mode=1; 4. On Enque/Deque of data packet: if (((DataSeqNoLast-AckSeqNoFirst)&lt;*QueueLim/2) OR (DataQueueLength==0)) ConservativeMode=1; Figure 8: ConservativeMode updates We now present the algorithm (Figure 8) for updating the ConservativeMode variable which controls the switching of the Ack Regulator algorithm between the conservative and the non-conservative modes. The algorithm starts in conservative mode (line 1). Whenever a targeted TCP window size is reached (in this case, 2*QueueLim) , the algorithm is switched into non-conservative mode (line 2). TCP Window Size is approximated here by the difference between the largest sequence number in the data queue and the sequence number in the ack queue. This is a reasonable approximation in our case since the wireless link is likely the bottleneck and most (if not all) of the queuing is done at the RNC. When operating in the non-conservative mode, no additional buffer space is reserved. This implies that there will be single loss the next time the TCP source increases it window size. At the detection of the packet loss, the algorithm again switches back to the conservative mode (line 3). This ensures that losses are of the single loss variety as long as the estimate of AckReleaseCount is conservative. Line 4 in the algorithm results in a switch back into conservative mode 77 whenever the data queue length goes to zero or whenever the TCP window size is halved. This handles the case when TCP reacts to losses elsewhere in the network and the Ack Regulator can go back to being conservative. Note that, if the TCP source is ECN capable, instead of switching to non-conservative mode, the Ack Regulator can simply mark the ECN bit to signal the source to reduce its congestion window, resulting in no packet loss. 1.Initialize AckReleaseCount=0; 2. On Enque of Ack/Deque of data packet: (after processing in Fig 7) AckReleaseCount+=AcksSent; 3. On Enque of data packet: if (AckReleaseCount&gt;0) AckReleaseCount; 4. On Deque of data packet: if (DataQueueLength==0) AckReleaseCount=0; Figure 9: AckReleaseCount updates We finally present the algorithm for updating the AckReleaseCount variable in Figure 9. Since AckReleaseCount estimates the expected number of data packets that are arriving and reserves buffer space for them, it is important to get an accurate estimate. An overestimate of AckReleaseCount would result in unnecessary reservation of buffers that won't be occupied, while an underestimate of AckReleaseCount can lead to buffer overflow loss(es) even in conservative mode due to inadequate reservation. With the knowledge of the exact version of the TCP source and the round trip time from the RNC to the source, it is possible to compute an exact estimate of AckReleaseCount. However, since we would like to be agnostic to TCP version as far as possible and also be robust against varying round trip times on the wired network, our algorithm tries to maintain a conservative estimate of AckReleaseCount. Whenever we send acks back to the source, we update AckReleaseCount by that many acks (line 2). Likewise, whenever a data packet arrives into the RNC from the source, we decrement the variable while ensuring that it does not go below zero (line 3). While maintaining a non-negative AckReleaseCount in this manner avoids underestimation, it also can result in unbounded growth of AckReleaseCount leading to significant overestimation as errors accumulate. For example, we increase AckReleaseCount whenever we send acks back to the source; however, if TCP is reducing its window size due to loss, we cannot expect any data packets in response to the acks being released. Thus, over time, AckReleaseCount can grow in an unbounded manner. In order to avoid this scenario, we reset AckReleaseCount to zero (line 4) whenever the data queue is empty. Thus, while this reset operation is necessary for synchronizing the real and estimated AckReleaseCount after a loss, it is not a conservative mechanism in general since a AckReleaseCount of zero implies that no buffer space is currently reserved for any incoming data packets that are unaccounted for. However, by doing the reset only when the data queue is empty, we significantly reduce the chance of the unaccounted data packets causing a buffer overflow loss. We discuss the impact of this estimation algorithm of AckReleaseCount in Section 6.6. Finally, we assume that there is enough buffer space for RNC S1 Sn M1 Mn V 100Mb/s 1ms 100Mb/s 1ms L Mb/s D ms L Mb/s D ms FR Mb/s FD ms RR Mb/s RD ms Figure 10: Simulation Topology the ack packets in the RNC. The maximum number of ack packets is the maximum window size achieved by the TCP flow (*QueueLim in our algorithm). Ack packets do not have to be buffered as is, since storing the sequence numbers is sufficient (however, care should be taken to preserve duplicate ack sequence numbers as is). Thus, memory requirement for ack storage is very minimal. SIMULATION RESULTS In this section, we present detailed simulation results comparing the performance of TCP Reno and TCP Sack, in the presence and absence of the Ack Regulator. First, we study the effect of variable bandwidth and variable delay using different distributions on the throughput of a single long-lived TCP flow. Next, we present a model for 3G1X-EVDO (HDR) system (which exhibits both variable rate and variable delay), and evaluate the performance of a single TCP flow in the HDR environment. Then, we present the performance of multiple TCP flows sharing a single HDR wireless link. Finally, we briefly discuss the impact of different parameters affecting the behavior of Ack Regulator. All simulations are performed using ns-2. The simulation topology used is shown in Figure 10. S i , i = 1..n corresponds to the set of TCP source nodes sending packets to a set of the TCP sink nodes M i , i = 1..n. Each set of S i , M i nodes form a TCP pair. The RNC is connected to the M i nodes through a V (virtual) node for simulation purposes. L, the bandwidth between S i and the RNC, is set to 100Mb/s and D is set to 1ms except in cases where D is explicitly varied. The forward wireless channel is simulated as having rate F R and delay F D, and the reverse wireless channel has rate RR and delay RD. Each simulation run lasts for 3600s (1hr) unless otherwise specified and all simulations use packet size of 1KB. TCP maximum window size is set to 500KB. Using such a large window size ensures that TCP is never window limited in all experiments except in cases where the window size is explicitly varied. 6.1 Variable Delay In this section, the effect of delay variation is illustrated by varying F D, the forward link delay. Without modification, the use of a random link delay in the simulation will result in out-of-order packets since packet transmitted later with lower delay can overtake packets transmitted earlier with higher delay. However, since delay variability in our model is caused by factors that will not result in packet reordering (e.g. processing time variation) and RLP delivers packet in sequence, the simulation code is modified such that packets cannot reach the next hop until the packet transmitted earlier has arrived. This modification applies to all simulations with variable link delay. Figure 11(a) shows throughput for a single TCP flow (n = 1) for F R = 200Kb/s and RR = 64Kb/s. F D has an ex-78 100 120 140 160 180 200 20 30 40 50 60 70 80 90 100 TCP Throughput (kb/s) Delay Variance Reno Reno, w/AR Sack Sack, w/AR 100 120 140 160 180 200 2 4 6 8 10 12 14 16 18 Throughput (kb/s) Buffer Size (packet) Reno Reno, w/AR Sack Sack, w/AR BDP=10 (a) Delay Variability (b) Different Buffer Size Figure 11: Throughput with Variable Delay e(x)+400-x ponential distribution with a mean that varies from 20ms to 100ms, and RD = 400ms - mean(F D) so that average F D+RD is maintained at 400ms. The buffer size on the bottleneck link for each run is set to 10, the product of the mean throughput of (200Kb/s or 25pkt/s) and mean link delay (0.4s). This product will be referred to as the bandwidth-delay product (BDP) in later sections. Additional delay distributions like uniform, normal, lognormal, and Poisson were also experimented with. Since the results are similar, only plots for an exponential delay distribution are shown. As expected, when the delay variation increases, throughput decreases for both TCP Reno and TCP Sack. By increasing the delay variance from 20 to 100, throughput of TCP Reno decreases by 30% and TCP Sack decreases by 19%. On the other hand, TCP Reno and TCP Sack flows which are Ack Regulated are much more robust and its throughput decreases by only 8%. Relatively to one another , Ack Regulator performs up to 43% better than TCP Reno and 19% better than TCP Sack. Another interesting result is that Ack Regulator delivers the same throughput irrespective of whether the TCP source is Reno or Sack. This is understandable given the fact that the Ack Regulator tries to ensure that only single buffer overflow loss occurs and in this regime, Reno and Sack are known to behave similarly. This property of Ack Regulator is extremely useful since for a flow to use TCP Sack, both the sender and receiver needs to be upgraded. Given that there are still significant number of web servers that have not yet been upgraded to TCP Sack [28], deployment of Ack Regulator would ensure excellent performance irrespective of the TCP version running. Figure 11(b) shows how throughput varies with buffer size with the same set of parameters except for F D, which is now fixed with a mean of 50ms (exponentially distributed). Even with a very small buffer of 5 packets (0.5 BDP), Ack Regulator is able to maintain a throughput of over 80% of the maximum throughput of 200Kb/s. Thus, Ack Regulator delivers robust throughput performance across different buffer sizes. This property is very important in a varying rate and delay environment of a wireless system, since it is difficult to size the system with an optimal buffer size, given that the BDP also varies with time. For a buffer of 4 packets, the improvement over TCP Reno and Sack is about 50% and 24% respectively. As buffer size increases, the throughput difference decreases. With buffer size close to 20 packets (2 BDP), TCP Sack performs close to Ack Regulated flows, while improvement over TCP Reno is about 4%. Finally, in Table 3, we list parameter values from the simulation for delay variance of 100. First, consider Reno and Item Rate, TD TO p1 p2 p3 W f Kb/s Reno 129 496 114 0.34 0.3 0.38 19 Reno+AR 184 302 8 0.98 0.0 0.02 24 Sack 160 434 4 0.99 0.0 0.01 19 Sack+AR 184 302 8 0.97 0.0 0.03 24 Table 3: Parameters from simulation for variance=100 140 150 160 170 180 190 200 0 10 20 30 40 50 60 70 80 Throughput (kb/s) Rate Variance Reno Reno, w/AR Sack Sack, w/AR 80 100 120 140 160 180 0 5 10 15 20 25 30 35 40 45 Throughput (kb/s) Buffer Size (packet) Reno, No AR Reno, w/AR Sack, No AR Sack, w/ AR BDP=9 (a) Bandwidth Variability (b) Different Buffer Size Figure 12: Throughput with Variable Bandwidth u(200,x) Reno with Ack Regulator (first two rows). It is clear that Ack Regulator is able to significantly reduce the conditional probability of multiple losses p2 and p3 as well as absolute number of loss events (T D and T O) resulting in substantial gains over Reno. Next, consider Sack and Sack with Ack Regulator (last two rows). In this case, we can see that Sack is very effective in eliminating most of the timeout occurrences . However, Ack Regulator is still able to reduce the absolute number of loss events by allowing the congestion window to grow to higher values (24 vs 19), resulting in throughput gains. 6.2 Variable Bandwidth In this section, we vary the link bandwidth, F R. Figure 12(a) shows throughput for a single TCP flow. F R is uniformly distributed with a mean of 200 Kb/s and the variance is varied from 20 to 75. F D = 200ms, RR = 64Kb/s and RD = 200ms. The buffer size on the bottleneck link for each run is 10. Again, we have experimented with other bandwidth distributions, but, due to lack of space, only uniform distribution is shown. Note that, with variable rate, the maximum throughput achievable is different from the mean rate. For uniform distribution, a simple closed form formula for the throughput is simply 1/ b a 1/xdx = 1/(ln b-ln a) where b is the maximum rate and a is the minimum rate. When the rate variance increases, throughput of TCP Reno decreases as expected. Compared to TCP Reno, Ack Regulator improves the throughput by up to 15%. However , TCP Sack performs very well and has almost the same throughput as Ack Regulated flows. Based on the calculations for maximum throughput discussed before, it can be shown that all flows except Reno achieve maximum throughput . This shows that if rate variation is not large enough, TCP Sack is able to handle the variability. However, for very large rate variations (e.g. rate with lognormal distribution and a large variance), the performance of TCP Sack is worse than when Ack Regulator is present. Figure 12(b) shows how the throughput varies with buffer size. Note that with a lower throughput, bandwidth delay product is smaller than 10 packets. Again, Ack Regulated 79 140 145 150 155 160 165 170 175 180 185 190 6 8 10 12 14 16 18 Throughput (kb/s) Buffer Size (packet) Reno, No AR Reno, w/AR Sack, No AR Sack, w/ AR BDP=9 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 6 8 10 12 14 16 18 RTT (sec) Buffer Size (packet) Reno, No AR Reno, w/AR Sack, No AR Sack, w/ AR BDP=9 (a) Throughput vs Buffer Size (b) rtt vs Buffer Size Figure 13: Throughput and rtt for u(200,50),350+e(50) TCP flows perform particularly well when the buffer size is small. With buffer size of 5, the improvement over TCP Sack is 40%. 6.3 Variable Delay and Bandwidth In this section, we vary both the bandwidth and delay of the wireless link. F R is uniformly distributed with a mean of 200 Kb/s and variance of 50, DR is exponentially distributed with a mean of 50ms, RR = 64Kb/s and RD = 350ms. The maximum achievable throughput is 186.7 Kb/s. The BDP is therefore about 9 packets. Figure 13(a) shows the throughput for a single TCP flow with the buffer size ranging from 7 to 20. The combination of variable rate and delay has a large negative impact on the performance of TCP Reno and it is only able to achieve 70% to 80% of the bandwidth of Ack Regulated flows when the buffer size is 6 packets. Even with a buffer size of 18 packets, the throughput difference is more than 5%. Throughput of TCP Sack is about 5% to 10% lower than Ack Regulator, until the buffer size reaches 18 packets (about 2 BDP). One of the cost of using the Ack Regulator is the increase in average round trip time (rtt). The average rtt values for all 4 types of flows are shown in 13(b) for different buffer sizes. TCP Reno has the lowest rtt followed by TCP Sack and the rate of rtt increase with buffer size is comparable. With Ack Regulator, rtt increase is comparable with unreg-ulated flows for buffer size less than 9 (1 BDP). For larger buffer sizes, since Ack Regulator uses = 2 times buffer size to regulate the acks in conservative mode, rtt increases faster with buffer size than regular TCP, where only the data packet buffer size contributes to rtt. For example, with buffer size of 9, Ack Regulated flows have a rtt 15% larger and with buffer size of 18, the rtt is 48% larger compared to TCP Sack. This effect can be controlled by varying the parameter of the Ack Regulator. 6.4 Simulation with High Data Rate High Data Rate (HDR) [6] is a Qualcomm proposed CDMA air interface standard (3G1x-EVDO) for supporting high speed asymmetrical data services. One of the main ideas behind HDR is the use of channel-state based scheduling which transmits packets to the user with the best signal-to-noise ratio. The actual rate available to the selected user depends on the current signal-to-noise ratio experienced by the user. The higher the ratio, the higher the rate available to the user. In addition, in order to provide some form of fairness, a Proportional Fair scheduler is used which provides long-term fairness to flows from different users. We use Qualcomm's Proportional Fair scheduler in our simulation with an averaging window of 1000 time slots, where each Rate(Kb/s) Prob. Rate(Kb/s) Prob. 38.4 0.033 614.4 0.172 76.8 0.015 921.6 0.145 102.6 0.043 1228.8 0.260 153.6 0.023 1843.2 0.042 204.8 0.060 2457.6 0.011 307.2 0.168 Table 4: HDR Data Rates for a one user system 250 300 350 400 450 500 550 600 0 5 10 15 20 25 30 35 40 Throughput (Kb/s) Buffer Size (packet) Reno, No AR Reno, w/AR Sack, No AR Sack, w/ AR BDP=15 200 300 400 500 600 700 800 0 5 10 15 20 25 30 35 40 Average RTT (ms) Buffer Size (Packet) Reno Reno, w/AR Sack Sack, w/AR BDP=15 (a) Throughput (b) rtt Figure 14: Throughput/rtt with HDR, Single Flow slot is 1.67 ms. While the HDR system results in higher raw throughput, the rate and delay variation seen is substantial. In this section, we model a simplified HDR environment in ns-2, focusing on the layer 3 scheduling and packet fragmentation . The fading model for the wireless link used is based on Jake's Rayleigh fading channel model [25]. This gives us the instantaneous signal-to-noise ratio. Using Table 2 in [6] which lists the rate achievable for a given signal-to-noise ratio assuming a frame error rate of less than 1%, the achievable bandwidth distribution (with one user) for our simulation is shown in Table 4. The simulation settings are as follows. F R is a variable that has a bandwidth distribution of Table 4, due to the variations of the fading conditions of the channel. Based on the guidelines from [26], F D is modeled as having a uniform distribution with mean 75ms and variance 30 and RD is modeled as having a uniform distribution with mean 125ms and variance 15. These are conservative estimates. We expect delay variations in actual systems to be higher (for example, note the ping latencies from our experiment in Section 3). The uplink in a HDR system is circuit-based and RR is set to be 64Kb/s. Figure 14(a) shows how throughput for a single TCP flow varies with buffer size. Assuming an average bandwidth of 600Kb/s and a link delay of 200ms, BDP is 15 packets. Again, the performance of TCP Reno flows that are Ack Regulated is significantly better than plain TCP Reno over the range of buffer size experimented, with improvements from 4% to 25%. TCP Sack flows also performs worse than Ack Regulated flows up to buffer size of 20. The improvement of Ack Regulator over TCP Sack ranges from 0.5% to 18%. As mentioned earlier, one of the costs of using the Ack Regulator is increase in average rtt. The average rtt for all 4 types of flows are shown in 14(b) with buffer size varying from 5 to 40. The effect is similar to the rtt variation with buffer size seen in Section 6.3. 6.5 Multiple TCP Flows In this simulation, the number of flows (n) sharing the bottleneck link is increased to 4 and 8. Per-flow buffering is 80 300 400 500 600 700 800 2 4 6 8 10 12 14 16 18 Total Throughput (Kb/s) Per Flow Buffer Size (Packet) 4 Reno Flows 4 Reno Flows, w/AR 4 Sack Flows 4 Sack Flows, w/AR BDP=5 450 500 550 600 650 700 750 800 850 900 950 2 4 6 8 10 12 14 16 18 Total Throughput (kb/s) Per Flow Buffer Size (Packet) 8 Reno Flows 8 Reno Flows, w/AR 8 Sack Flows 8 Sack Flows, w/AR BDP=3 (a) 4 TCP Flows (b) 8 TCP Flows Figure 15: Throughput with HDR, Multiple Flows provided for each TCP flow. For 4 flows, using mean rate of 200Kb/s, 1KB packet and rtt of 0.2s, BDP is 5 packets per flow. For 8 flows, using mean rate of 120Kb/s, 1KB packet and rtt of 0.2s, BDP is 3 packets per flow. As the number of TCP flows increases, the expected rate and delay variation seen by individual flows also increases. Thus, even though the total throughout of the system increases with more users due to channel-state based scheduling , the improvement is reduced by the channel variability. Figure 15(a) shows the throughput for 4 TCP flows. The improvement of Ack Regulator over TCP Sack increases compared to the single TCP case. For example, the gain is 17% with per-flow buffer size of 5 (BDP). For Reno the gain is even greater. With per-flow buffer size of 5, the improvement is 33%. Similar result can also be observed for the case of 8 TCP flows as shown in Figure 15(b). For both TCP Reno and Sack, the gain is about 31% and 29% respectively for per-flow buffer size of 3. From the figure, it can seen that, for TCP Sack and Reno to achieve close to maximum throughput without Ack Regulator, at least three times the buffer requirements of Ack Regulator is necessary (buffer requirements for acks in the Ack Regulator is negligible compared to the 1KB packet buffer since only the sequence number needs to be stored for the acks). This not only increases the cost of the RNC, which needs to support thousands of active flows, it also has the undesirable side-effects of large rtt's that was noted in Section 3. With multiple TCP flows, the issue of throughput fairness naturally arises. One way to quantify how bandwidth is shared among flows is to use the fairness indexdescribed in [27]. This indexis computed as the ratio of the square of the total throughput to n times the square of the individual flow throughput. If all flows get the same allocation, then the fairness indexis 1. As the differences in allocation increases, fairness decreases. A scheme which allocates bandwidth to only a few selected users has a fairness indexnear 0. Computation of this indexis performed for all multiple flows simulation and the indexis greater than 0.99 in all cases. This result is expected since with per flow buffering, and proportional fair scheduling, the long term throughput of many TCP long-lived flows sharing the same link should be fair. 6.6 Parameters affecting the performance of Ack Regulator Due to lack of space, we will only briefly present the results of varying parameters such as wired network latency and . As the network latency is varied from 20ms to 100ms, throughput decreases by 1.63% and 2.62% for Reno and Reno with Ack Regulator flows, respectively. Most of the decrease can be accounted for by the impact of increase in latency on TCP throughput. The result shows that the AckReleaseCount estimation algorithm is effective and hence the Ack Regulator is able to reserve the appropriate amount of buffer for expected packet arrivals even with substantial wireline delay. In another experiment, the parameter in an Ack Regulated TCP flow is varied from 1 to 4. When is increased from 1 to 3, the TCP flow is able to achieve its maximum throughput at a smaller buffer size. As increases, the rtt also increases and when is increased to 4, throughput decreases for larger buffer sizes (&gt; 15). The decrease in throughput is caused by the accumulation of sufficiently large amount of duplicate acks that are sent to the TCP sender. A value of = 2 appears to be a good choice, balancing throughput and rtt for reasonable buffer sizes. 6.7 Summary of Results and Discussion In this section, we first summarize the results from the simulation experiments and then briefly touch upon other issues. We first started with experiments using a wireless link with variable delay. We showed that Ack Regulator delivers performance up to 43% better than TCP Reno and 19% better than TCP Sack when the buffer size was set to one BDP. We then examined the impact of a wireless link with variable rate. We saw that when the rate variance increases, throughput of TCP Reno decreases as expected. Compared to TCP Reno, Ack Regulator improves the throughput by up to 15%. However, TCP Sack performs very well and has almost the same throughput as Ack Regulated flows as long as the rate variation is not extremely large. We next considered the impact of a wireless link with variable delay and variable rate. We found that this combination had a large negative impact on the performance of both TCP Reno and Sack (up to 22% and 10% improvement respectively for Ack Regulated flows). We then considered a specific wireless link standard called HDR which exhibits both variable delay and variable rate. The results were as expected, with Ack Regulator improving TCP Reno performance by 5% to 33% and TCP Sack by 0.5% to 24%. We then evaluated the impact of multiple TCP flows sharing the HDR link. The gains of Ack Regulator over normal TCP flows were even greater in this case (with 32% to 36% improvements) when the buffer size is set to one BDP. In general, we showed that Ack Regulator delivers the same high throughput irrespective of whether the TCP flow is Reno or Sack. We further showed that Ack Regulator delivers robust throughput performance across different buffer sizes with the performance improvement of Ack Regulator increasing as buffer size is reduced. We only considered TCP flows towards the mobile host (for downloading-type applications) since links like HDR are designed for such applications. In the case of TCP flows in the other direction (from the mobile host), Ack Regulator can be implemented, if necessary, at the mobile host to optimize the use of buffer on the wireless interface card. Finally, Ack Regulator cannot be used if the flow uses end-to-end IPSEC. This is also true for all performance enhancing proxies. However, we believe that proxies for performance improvement are critical in current wireless networks. In order to allow for these proxies without compromising security, a split security model can be adopted where the 81 RNC, under the control of the network provider, becomes a trusted element. In this model, a VPN approach to security (say, using IPSEC) is used on the wireline network between the RNC and the correspondent host and 3G authentication and link-layer encryption mechanisms are used between the RNC and mobile host. This allows the RNC to support proxies such as the Ack Regulator to improve performance without compromising security. CONCLUSION In this paper, we comprehensively evaluated the impact of variable rate and variable delay on TCP performance. We first proposed a model to explain and predict TCP's throughput over a link with variable rate and delay. Our model was able to accurately (better than 90%) predict throughput of TCP flows even in the case of large delay and rate variation. Based on our TCP model, we proposed a network based solution called Ack Regulator to mitigate the effect of rate and delay variability. The performance of Ack Regulator was evaluated extensively using both general models for rate and delay variability as well as a simplified model of a 3 rd Generation high speed wireless data air interface . Ack Regulator was able to improve the performance of TCP Reno and TCP Sack by up to 40% without significantly increasingly the round trip time. We also showed that Ack Regulator delivers the same high throughput irrespective of whether the TCP source is Reno or Sack. Furthermore , Ack Regulator also delivered robust throughput performance across different buffer sizes. Given the difficulties in knowing in advance the achievable throughput and delay (and hence the correct BDP value), a scheme, like Ack Regulator, which works well for both large and small buffers is essential. In summary, Ack Regulator is an effective network-based solution that significantly improves TCP performance over wireless links with variable rate and delay. Acknowledgements The authors would like to thank Lijun Qian for providing the fading code used in the HDR simulation, Clement Lee and Girish Chandranmenon for providing the 3G1xtrace and Sandy Thuel for comments on earlier versions of this paper. REFERENCES [1] E. Altman, K. Avrachenkov and C. Barakat, "A Stochastic Model of TCP/IP with Stationary Random Loss," in Proceedings of SIGCOMM 2000. [2] F. Baccelli and D. Hong,"TCP is Max-Plus Linear," in Proceedings of SIGCOMM 2000. [3] A. Bakre and B.R. Badrinath, "Handoff and System Support for Indirect TCP/IP," in proceedings of Second UsenixSymposium on Mobile and Location-Independent Computing, Apr 1995. [4] H. Balakrishnan et al., "Improving TCP/IP Performance over Wireless Networks," in proceedings of ACM Mobicom, Nov 1995. [5] H. Balakrishnan, V.N. Padmanabhan, R.H. Katz, "The Effects of Asymmetry on TCP Performance," Proc. ACM/IEEE Mobicom, Sep. 1997. [6] P. Bender et al., "A Bandwidth Efficient High Speed Wireless Data Service for Nomadic Users," in IEEE Communications Magazine, Jul 2000. [7] P. Bhagwat at al, "Enhancing Throughput over Wireless LANs Using Channel State Dependent Packet Scheduling," in Proc. IEEE INFOCOM'96. [8] K. Brown and S.Singh, "M-TCP: TCP for Mobile Cellular Networks," ACM Computer Communications Review Vol. 27(5), 1997. [9] A. Canton and T. Chahed, "End-to-end reliability in UMTS: TCP over ARQ," in proceedings of Globecomm 2001. [10] TIA/EIA/cdma2000, "Mobile Station - Base Station Compatibility Standard for Dual-Mode Wideband Spread Spectrum Cellular Systems", Washington: Telecommunication Industry Association, 1999. [11] G. Holland and N. H. Vaidya, "Analysis of TCP Performance over Mobile Ad Hoc Networks," in Proceedings of ACM Mobicom'99. [12] H. Inamura et al., "TCP over 2.5G and 3G Wireless Networks," draft-ietf-pilc-2.5g3g-07, Aug. 2002. [13] F. Khafizov and M. Yavuz, "TCP over CDMA2000 networks," Internet Draft, draft-khafizov-pilc-cdma2000-00.txt. [14] T. V. Lakshman and U. Madhow, "The Performance of Networks with High Bandwidth-delay Products and Random Loss," in IEEE/ACM Transactions on Networking, Jun. 1997. [15] R. Ludwig et al., "Multi-layer Tracing of TCP over a Reliable Wireless Link," in Proceedings of ACM SIGMETRICS 1999. [16] Reiner Ludwig and Randy H. Katz "The Eifel Algorithm: Making TCP Robust Against Spurious Retransmissions," in ACM Computer Communications Review, Vol. 30, No. 1, 2000. [17] V. Misra, W. Gong and D. Towsley, "Stochastic Differential Equation Modeling and Analysis of TCP Windowsize Behavior," in Proceedings of Performance'99. [18] P. Narvaez and K.-Y. Siu, "New Techniques for Regulating TCP Flow over Heterogeneous Networks," in LCN'98. [19] "Modeling TCP Throughput: a Simple Model and its Empirical Validation," in Proceedings of SIGCOMM 1998. [20] S. Paul et al., "An Asymmetric Link-Layer Protocol for Digital Cellular Communications," in proceedings of INFOCOM 1995. [21] Third Generation Partnership Project, "RLC Protocol Specification (3G TS 25.322:)", 1999. [22] TIA/EIA/IS-707-A-2.10, "Data Service Options for Spread Spectrum Systems: Radio Link Protocol Type 3", January 2000. [23] S. Karandikar et al., "TCP rate control," in ACM Computer Communication Review, Jan 2000. [24] 3G Partnership Project, Release 99. [25] "Microwave mobile communications," edited by W. C. Jakes, Wiley, 1974. [26] "Delays in the HDR System," QUALCOMM, Jun. 2000. [27] R. Jain, "The Art of Computer Systems Performance Analysis," Wiley, 1991. [28] J. Padhye and S. Floyd, "On Inferring TCP Behavior," in Proceedings of SIGCOMM'2001. 82

algorithm;architecture;TCP;wireless communication;performance evaluation;3G wireless links;prediction model;design;Link and Rate Variation;3G Wireless;simulation result;congestion solution;Network

188

The Forest and the Trees: Using Oracle and SQL Server Together to Teach ANSI-Standard SQL

Students in a sophomore-level database fundamentals course were taught SQL and database concepts using both Oracle and SQL Server. Previous offerings of the class had used one or the other database. Classroom experiences suggest that students were able to handle learning SQL in the dual environment, and, in fact, benefited from this approach by better understanding ANSI-standard versus database-specific SQL and implementation differences in the two database systems.

INTRODUCTION A problem arises in many technology classes. The instructor wishes to teach principles and concepts of a technology. To give students hands-on experience putting those theories to work, a specific product that implements that technology is selected for a lab component. Suddenly the students are learning more about the specific product than they are about the technology concepts. They may or may not realize what is specific to that product and what is general to the technology. Students may even start referring to the course as a VB course, a PHP course, or an Oracle course when what you wanted to teach was programming, web scripting, or database principles. This paper presents the experiences from a database fundamentals course that used both Oracle and SQL Server so that students would better understand ANSI-standard SQL. Though each database is ANSI SQL compliant, there are definite differences in implementation (Gorman, 2001; Gulutzan, 2002). By learning each implementation and how each departs from ANSI-standard SQL, students can be better prepared to work with any database and better understand general concepts of databases and SQL. The paper discusses the observed results from this approach and how well the approach met learning objectives. COURSE CONTEXT AND LEARNING OBJECTIVES CPT 272, Database Fundamentals, is a sophomore-level database programming and design class taught primarily to computer technology majors in their fourth semester. Students will have previously taken a freshman-level course that introduces them to databases as a tool for learning general information system development terms and concepts. The freshman-level course uses Microsoft Access because it is easy to use for quickly developing a small personal information system. That course also introduces both SQL and Query By Example methods for querying a database as well basic database design concepts, which are applied for simple data models. Students then move into two programming courses, the second of which uses single-table SQL statements for providing data to (formerly) Visual Basic or (currently) web-programming applications. So by the time the students take the Database Fundamentals course they have a concept of what a database is and how it is used as the back-end for programming. The Database Fundamentals course is the course where students learn SQL in depth, database concepts, and basic database design. It does not teach stored procedure programming, triggers, or enterprise or distributed database design, which are covered in more advanced courses. The learning objectives for the Database Fundamentals course are: To understand the fundamentals of a relational database. To understand the fundamentals of client-server and multi-tiered applications. To understand the principles and characteristics of good relational database design. To design entity relationship models for a business problem domain verified by the rules of normalization (through third normalized form). To build simple to moderately complex data models. To write simple to moderately complex SQL to query a multiple-table database. To write data manipulation language (DML) SQL to insert rows, delete rows, and update rows. To understand the concept of database transactions and demonstrate the proper use of commits and rollbacks. To write data definition language (DDL) SQL to create and drop tables, indexes, and constraints. To understand and be able to implement the fundamentals of security and permissions in a database. To explain the benefits of using views and write SQL statements to create views. To create and use SQL scripts and use SQL to build scripts. To gain a working knowledge of query optimization, performance tuning, and database administration. To apply team skills to build a client-server database application. CONSIDERATIONS FOR CROSS-ENGINE SQL EDUCATION CPT 272, Database Fundamentals, is taught in a multi-campus university. It was initially taught on the main campus using Oracle. When the course was rolled out to the regional campuses, SQL Server was first used because of administration considerations involved with Oracle. Now WAN connections have been established that allow the use of either database engine or both. During the spring 2003 semester one regional campus experimented with the use of both databases. The reasons for doing this were: To accomplish the course learning objectives in SQL necessitates going beyond ANSI-standard SQL into database-specific functions, sub-queries, and other aspects of SQL that are implemented differently in different databases. If the students learn only Oracle or SQL Server (or any other database) they are likely to confuse ANSI-standard SQL with the database-specific implementation, which can hinder them when they enter the job market. By using both databases, it was hoped that students would learn and understand the differences among ANSI-standard SQL, Oracle SQL, SQL Server T-SQL. Some design considerations, such as Identities/Sequences and datatypes, are implemented differently in different databases. Again, students will enter the job market with a stronger understanding if they understand the difference between the concept and how it is implemented. Neither Oracle nor SQL Server commands a majority of market share. However, the two together make up about fifty percent of the current market share, positioning students well for the job market (Wong, 2002). Studying two databases together opens the door for discussing the pros and cons of these and other databases, including DB2, MySQL, and Sybase. Finally, students often want to install a database engine on their personal computer and work on lab assignments at home. Both Oracle and SQL Server have licensing and hardware requirement issues that on any given computer may preclude one or the other. Using both allowed most students to do at least some of their work at home. To implement a cross-engine approach, an SQL text would be needed that taught both databases. A special textbook was created by two of the instructors, a draft of which was used in CD format. In addition to covering SQL essentials in Oracle and SQL Server, it also covered Microsoft Access and MySQL in hopes that it might also be used in the freshman course and be a good reference for real world web programming. The text has since been picked up by a publisher. COURSE DESIGN AND ASSIGNMENTS CPT 272, Database Fundamentals, consists of a both a lecture and a lab component. The lectures cover fundamental database concepts, including SQL concepts, query optimization, and database design and normalization. The lab component focuses on mastery of SQL. Table 1 shows the labs that were assigned and the database used for each. Table 1. Course Lab Schedule Lab Lab DBMS Single Table Select Oracle Aggregates & Sub Queries SQL Server Joining Tables Both Oracle & SQL Server DBMS Specific Functions Both Oracle & SQL Server Advanced Queries Oracle Data Manipulation Student Choice Database Definition Student Choice Privileges Student Choice The first three labs concentrated as much as possible on ANSI-standard SQL. Of course, implementation of sub queries and joins is in some cases different between Oracle and SQL Server. These differences were taught and discussed. However, those labs avoided all DBMS specific functionality, such as concatenation, date manipulation, and datatype conversion. These things were covered in the DBMS Specific Functions lab. This approach limited some of what could be done in the first labs but provided a solid distinction between what was ANSI-standard SQL and what was database-specific. Later labs used these database-specific functions in various lab exercises so that these functions, which are crucial in the real world, were mastered. In addition to the use of both databases in labs, lecture material constantly referred to how various design, administration, and optimization concepts would be applied in both databases and in other databases, such as MySQL. In addition, exam questions asked students to compare the capabilities of each database. Other 235 assignments led students into an exploration of the pros and cons of various database engines. Two of these were for students to write short papers on the following: 1. Research one of the following databases: DB2, Sybase, Informix, MySQL, PostgreSQL, or SQLWindows Solo. Write a 2-3 page paper comparing it to Oracle and SQL Server. Include your recommendations regarding the circumstances in which this database should be used. 2. Research what people on the Internet say comparing SQL Server and Oracle. Based on their perspectives and your own experience with these two databases, write 1-2 pages comparing them. DISCUSSION With all the course objectives listed above, the Database Fundamentals course makes for a full semester. Using two databases instead of one definitely adds to the challenge of fitting in all the material. This first attempt led to the realization of several "kinks" that would need to be worked out before it could be attempted again. These are listed below. In most cases any given SQL lecture had barely enough time for covering the SQL concepts and implementation in both databases, forcing the instructor to scrimp on in-class examples. This meant that students had a more shaky foundation going into the lab assignment. However, it should be noted that students did about as well on lab assignments as prior classes. One possible solution would be to focus each SQL lecture on only one database, but to revisit that concept in the following lecture when the other database is used. In past semesters using only one database, it was possible to include a discussion of datatypes along with the DDL lecture. Using two databases the DDL lecture had to be expanded, which did not leave enough time for thorough coverage of datatypes in both databases. A solution would be to move the datatype material to one of the design lectures, emphasizing datatyping as a design step. By switching between databases, database-specific syntax never clicked in students' minds. They seemed less able than prior classes to apply concatenation and datatype conversions without looking up the syntax in reference material. While this is regrettable, it may be a worthwhile trade-off for gaining an understanding of what is ANSI-standard vs. database-specific SQL. It is likely that when students move to the real world and settle in with one database, they will quickly be able to internalize the syntax. Some students expressed that they would have preferred going through all labs with one database and then looking at differences with a second database. However, other students liked working with both databases side by side. This suggests that the alternating structure may need to be tweaked. But whatever one does, it will probably not work with every learning style. After using both databases, almost all students, when given a choice, used SQL Server. This was solely because of a perceived superiority in the user interface of Query Analyzer versus SQL Plus. In itself this is not a bad thing because one of the goals was to understand the differences in the two databases. But in future semesters the instructor may want to force a choice more often to insure that students will be exposed more equally to both. These "kinks" aside, both instructor and students considered the experiment a success. Their papers indicated a mature appreciation of the differences between Oracle and SQL Server and how they compared to other databases. Their comments in class indicated that they understood what was ANSI-standard SQL and what was not. In the SQL lab exam these students performed as well as previous classes of students, indicating that the cross-engine approach did not hinder their learning. Students also indicated enthusiasm for being able to list experience in both databases on their resume. Compared to prior semesters, students left the course more comfortable and able to use either of these major database engines to accomplish the goals of a given information system. REFERENCES [1] Gorman, Michael M. (2001). Is SQL a real standard anymore? The Data Administration Newsletter (TDAN.com) http://www.tdan.com/i016hy01.htm (17 Apr. 2003). [2] Gulutzan, Peter. (2002). Standard SQL. http://www.dbazine.com/gulutzan3.html (17 Apr. 2003). [3] Wong, Wylie. (2002). IBM passes Oracle in database market. http://techupdate.zdnet.com/techupdate/stories/main/0,14179 ,2864350,00.html (19 June 2003). 236

SQL;SQL Server;training vs. education;database systems;Oracle;database;educational fundamentals;student feedbacks;ANSI-Standard SQL;teaching in IT;dual environment;SQL Language;course design;practical results

189

The Maximum Entropy Method for Analyzing Retrieval Measures

We present a model, based on the maximum entropy method, for analyzing various measures of retrieval performance such as average precision, R-precision, and precision-at-cutoffs. Our methodology treats the value of such a measure as a constraint on the distribution of relevant documents in an unknown list, and the maximum entropy distribution can be determined subject to these constraints. For good measures of overall performance (such as average precision), the resulting maximum entropy distributions are highly correlated with actual distributions of relevant documents in lists as demonstrated through TREC data; for poor measures of overall performance, the correlation is weaker. As such, the maximum entropy method can be used to quantify the overall quality of a retrieval measure. Furthermore, for good measures of overall performance (such as average precision), we show that the corresponding maximum entropy distributions can be used to accurately infer precision-recall curves and the values of other measures of performance, and we demonstrate that the quality of these inferences far exceeds that predicted by simple retrieval measure correlation, as demonstrated through TREC data.

INTRODUCTION The efficacy of retrieval systems is evaluated by a number of performance measures such as average precision, R-precision , and precisions at standard cutoffs. Broadly speaking , these measures can be classified as either system-oriented measures of overall performance (e.g., average precision and R-precision) or user-oriented measures of specific performance (e.g., precision-at-cutoff 10) [3, 12, 5]. Different measures evaluate different aspects of retrieval performance, and much thought and analysis has been devoted to analyzing the quality of various different performance measures [10, 2, 17]. We consider the problem of analyzing the quality of various measures of retrieval performance and propose a model based on the maximum entropy method for evaluating the quality of a performance measure. While measures such as average precision at relevant documents, R-precision, and 11pt average precision are known to be good measures of overall performance, other measures such as precisions at specific cutoffs are not. Our goal in this work is to develop a model within which one can numerically assess the overall quality of a given measure based on the reduction in uncertainty of a system's performance one gains by learning the value of the measure. As such, our evaluation model is primarily concerned with assessing the relative merits of system-oriented measures, but it can be applied to other classes of measures as well. We begin with the premise that the quality of a list of documents retrieved in response to a given query is strictly a function of the sequence of relevant and non-relevant documents retrieved within that list (as well as R, the total number of relevant documents for the given query). Most standard measures of retrieval performance satisfy this premise. Our thesis is then that given the assessed value of a "good" overall measure of performance, one's uncertainty about the sequence of relevant and non-relevant documents in an unknown list should be greatly reduced. Suppose, for example , one were told that a list of 1,000 documents retrieved in response to a query with 200 total relevant documents contained 100 relevant documents. What could one reasonably infer about the sequence of relevant and non-relevant documents in the unknown list? From this information alone, one could only reasonably conclude that the likelihood of seeing a relevant document at any rank level is uniformly 1/10. Now suppose that one were additionally told that the average precision of the list was 0.4 (the maximum possi-27 ble in this circumstance is 0.5). Now one could reasonably conclude that the likelihood of seeing relevant documents at low numerical ranks is much greater than the likelihood of seeing relevant documents at high numerical ranks. One's uncertainty about the sequence of relevant and non-relevant documents in the unknown list is greatly reduced as a consequence of the strong constraint that such an average precision places on lists in this situation. Thus, average precision is highly informative. On the other hand, suppose that one were instead told that the precision of the documents in the rank range [100, 110] was 0.4. One's uncertainty about the sequence of relevant and non-relevant documents in the unknown list is not appreciably reduced as a consequence of the relatively weak constraint that such a measurement places on lists. Thus, precision in the range [100, 110] is not a highly informative measure. In what follows, we develop a model within which one can quantify how informative a measure is. We consider two questions: (1) What can reasonably be inferred about an unknown list given the value of a measurement taken over this list? (2) How accurately do these inferences reflect reality? We argue that the former question is properly answered by considering the maximum entropy distributions subject to the measured value as a constraint, and we demonstrate that such maximum entropy models corresponding to good overall measures of performance such as average precision yield accurate inferences about underlying lists seen in practice (as demonstrated through TREC data). More specifically, we develop a framework based on the maximum entropy method which allows one to infer the most "reasonable" model for the sequence of relevant and non-relevant documents in a list given a measured constraint. From this model, we show how one can infer the most "reasonable" model for the unknown list's entire precision-recall curve. We demonstrate through the use of TREC data that for "good" overall measures of performance (such as average precision), these inferred precision-recall curves are accurate approximations of actual precision-recall curves; however, for "poor" overall measures of performance, these inferred precision-recall curves do not accurately approximate actual precision-recall curves. Thus, maximum entropy modeling can be used to quantify the quality of a measure of overall performance. We further demonstrate through the use of TREC data that the maximum entropy models corresponding to "good" measures of overall performance can be used to make accurate predictions of other measurements. While it is well known that "good" overall measures such as average precision are well correlated with other measures of performance, and thus average precision could be used to reasonably predict other measures of performance, we demonstrate that the maximum entropy models corresponding to average precision yield inferences of other measures even more highly correlated with their actual values, thus validating both average precision and maximum entropy modeling. In the sections that follow, we first describe the maximum entropy method and discuss how maximum entropy modeling can be used to analyze measures of retrieval performance . We then describe the results of applying our methodology using TREC data, and we conclude with a summary and future work. THE MAXIMUM ENTROPY METHOD The concept of entropy as a measure of information was first introduced by Shannon [20], and the Principle of Maximum Entropy was introduced by Jaynes [7, 8, 9]. Since its introduction, the Maximum Entropy Method has been applied in many areas of science and technology [21] including natural language processing [1], ambiguity resolution [18], text classification [14], machine learning [15, 16], and information retrieval [6, 11], to name but a few examples. In what follows, we introduce the maximum entropy method through a classic example, and we then describe how the maximum entropy method can be used to evaluate measures of retrieval performance. Suppose you are given an unknown and possibly biased six-sided die and were asked the probability of obtaining any particular die face in a given roll. What would your answer be? This problem is under-constrained and the most seemingly "reasonable" answer is a uniform distribution over all faces. Suppose now you are also given the information that the average die roll is 3.5. The most seemingly "reasonable" answer is still a uniform distribution. What if you are told that the average die roll is 4.5? There are many distributions over the faces such that the average die roll is 4.5; how can you find the most seemingly "reasonable" distribution? Finally, what would your answer be if you were told that the average die roll is 5.5? Clearly, the belief in getting a 6 increases as the expected value of the die rolls increases. But there are many distributions satisfying this constraint; which distribution would you choose? The "Maximum Entropy Method" (MEM) dictates the most "reasonable" distribution satisfying the given constraints. The "Principle of Maximal Ignorance" forms the intuition behind the MEM; it states that one should choose the distribution which is least predictable (most random) subject to the given constraints. Jaynes and others have derived numerous entropy concentration theorems which show that the vast majority of all empirical frequency distributions (e.g., those corresponding to sequences of die rolls) satisfying the given constraints have associated empirical probabilities and entropies very close to those probabilities satisfying the constraints whose associated entropy is maximal [7]. Thus, the MEM dictates the most random distribution satisfying the given constraints, using the entropy of the probability distribution as a measure of randomness. The entropy of a probability distribution p = {p 1 , p 2 , . . . , p n } is a measure of the uncertainty (randomness) inherent in the distribution and is defined as follows H(p) = n X i=1 p i lg p i . Thus, maximum entropy distributions are probability distributions making no additional assumptions apart from the given constraints. In addition to its mathematical justification, the MEM tends to produce solutions one often sees in nature. For example, it is known that given the temperature of a gas, the actual distribution of velocities in the gas is the maximum entropy distribution under the temperature constraint. We can apply the MEM to our die problem as follows. Let the probability distribution over the die faces be p = {p 1 , . . . , p 6 }. Mathematically, finding the maximum entropy distribution over die faces such that the expected die roll is 28 1 2 3 4 5 6 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 die face probability 1 2 3 4 5 6 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 die face probability 1 2 3 4 5 6 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 die face probability Figure 1: Maximum entropy die distributions with mean die rolls of 3.5, 4.5, and 5.5, respectively. d corresponds to the following optimization problem: Maximize: H(p) Subject to: 1. 6 P i=1 p i = 1 2. 6 P i=1 i p i = d The first constraint ensures that the solution forms a distribution over the die faces, and the second constraint ensures that this distribution has the appropriate expectation. This is a constrained optimization problem which can be solved using the method of Lagrange multipliers. Figure 1 shows three different maximum entropy distributions over the die faces such that the expected die roll is 3.5, 4.5, and 5.5, respectively. 2.1 Application of the Maximum Entropy Method to Analyzing Retrieval Measures Suppose that you were given a list of length N corresponding to the output of a retrieval system for a given query, and suppose that you were asked to predict the probability of seeing any one of the 2 N possible patterns of relevant documents in that list. In the absence of any information about the query, any performance information for the system , or any a priori modeling of the behavior of retrieval systems, the most "reasonable" answer you could give would be that all lists of length N are equally likely. Suppose now that you are also given the information that the expected number of relevant documents over all lists of length N is R ret . Your "reasonable" answer might then be a uniform distribution over all ` N R ret different possible lists with R ret relevant documents. But what if apart from the constraint on the number of relevant documents retrieved, you were also given the constraint that the expected value of average precision is ap? If the average precision value is high, then of all the ` N R ret lists with R ret relevant documents, the lists in which the relevant documents are retrieved at low numerical ranks should have higher probabilities. But how can you determine the most "reasonable" such distribution ? The maximum entropy method essentially dictates the most reasonable distribution as a solution to the following constrained optimization problem. Let p(r 1 , ..., r N ) be a probability distribution over the relevances associated with document lists of length N , let rel(r 1 , ..., r N ) be the number of relevant documents in a list, and let ap(r 1 , ..., r N ) be the average precision of a list. Then the maximum entropy method can be mathematically formulated as follows: Maximize: H(p) Subject to: 1. P r 1 ,...,r N p(r 1 , . . . , r N ) = 1 2. P r 1 ,...,r N ap(r 1 , . . . , r N ) p(r 1 , . . . , r N ) = ap 3. P r 1 ,...,r N rel(r 1 , . . . , r N ) p(r 1 , . . . , r N ) = R ret Note that the solution to this optimization problem is a distribution over possible lists, where this distribution effectively gives one's a posteriori belief in any list given the measured constraint. The previous problem can be formulated in a slightly different manner yielding another interpretation of the problem and a mathematical solution. Suppose that you were given a list of length N corresponding to output of a retrieval system for a given a query, and suppose that you were asked to predict the probability of seeing a relevant document at some rank. Since there are no constraints, all possible lists of length N are equally likely, and hence the probability of seeing a relevant document at any rank is 1/2. Suppose now that you are also given the information that the expected number of relevant documents over all lists of length N is R ret . The most natural answer would be a R ret /N uniform probability for each rank. Finally, suppose that you are given the additional constraint that the expected average precision is ap. Under the assumption that our distribution over lists is a product distribution (this is effectively a fairly standard independence assumption), we may solve this problem as follows. Let p(r 1 , . . . , r N ) = p(r 1 ) p(r 2 ) p(r N ) where p(r i ) is the probability that the document at rank i is relevant. We can then solve the problem of calculating the probability of seeing a relevant document at any rank using the MEM. For notational convenience, we will refer to this product distribution as the probability-at-rank distribution and the probability of seeing a relevant document at rank i, p(r i ), as p i . Standard results from information theory [4] dictate that if p(r 1 , . . . , r N ) is a product distribution, then H(p(r 1 , . . . , r N )) = N X i=1 H(p i ) where H(p i ) is the binary entropy H(p i ) = -p i lg p i - (1 - p i ) lg(1 - p i ). Furthermore, it can be shown that given a product distribution p(r 1 , . . . , r N ) over the relevances associated with docu-29 Maximize: P N i=1 H(p i ) Subject to: 1. 1 R N P i=1 ` p i i `1 + i-1 P j=1 p j = ap 2. N P i=1 p i = R ret Figure 2: Maximum entropy setup for average precision. Maximize: P N i=1 H(p i ) Subject to: 1. 1 R R P i=1 p i = rp 2. N P i=1 p i = R ret Figure 3: Maximum entropy setup for R-precision. Maximize: P N i=1 H(p i ) Subject to: 1. 1 k k P i=1 p i = PC (k) 2. N P i=1 p i = R ret Figure 4: Maximum entropy setup for precision-at-cutoff. ment lists of length N , the expected value of average precision is 1 R N X i=1 p i i 1 + i-1 X j=1 p j !! . (1) (The derivation of this formula is omitted due to space constraints .) Furthermore, since p i is the probability of seeing a relevant document at rank i, the expected number of relevant documents retrieved until rank N is P N i=1 p i . Now, if one were given some list of length N , one were told that the expected number of relevant documents is R ret , one were further informed that the expected average precision is ap, and one were asked the probability of seeing a relevant document at any rank under the independence assumption stated, one could apply the MEM as shown in Figure 2. Note that one now solves for the maximum entropy product distribution over lists, which is equivalent to a maximum entropy probability-at-rank distribution. Applying the same ideas to R-precision and precision-at-cutoff k, one obtains analogous formulations as shown in Figures 3 and 4, respectively . All of these formulations are constrained optimization problems , and the method of Lagrange multipliers can be used to find an analytical solution, in principle. When analytical solutions cannot be determined, numerical optimization methods can be employed. The maximum entropy distributions for R-precision and precision-at-cutoff k can be obtained analytically using the method of Lagrange multipliers . However, numerical optimization methods are required to determine the maximum entropy distribution for average precision. In Figure 5, examples of maximum entropy probability-at-rank curves corresponding to the measures average precision, R-precision, and precision-at-cutoff 10 for a run in TREC8 can be seen. Note that the probability-at -rank curves are step functions for the precision-at-cutoff and R-precision constraints; this is as expected since, for example, given a precision-at-cutoff 10 of 0.3, one can only reasonably conclude a uniform probability of 0.3 for seeing a relevant document at any of the first 10 ranks. Note, however, that the probability-at-rank curve corresponding to average precision is smooth and strictly decreasing. Using the maximum entropy probability-at-rank distribution of a list, we can infer the maximum entropy precision-recall curve for the list. Given a probability-at-rank distribution p, the number of relevant documents retrieved until rank i is REL(i) = P i j=1 p j . Therefore, the precision and recall at rank i are PC (i) = REL(i)/i and REC (i) = REL(i)/R. Hence, using the maximum entropy probability-0 100 200 300 400 500 600 700 800 900 1000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 rank probability TREC8 System fub99a Query 435 AP = 0.1433 ap maxent dist. rp maxent dist. pc-10 maxent dist. Figure 5: Probability-at-rank distributions. at-rank distribution for each measure, we can generate the maximum entropy precision-recall curve of the list. If a measure provides a great deal of information about the underlying list, then the maximum entropy precision-recall curve should approximate the precision-recall curve of the actual list. However, if a measure is not particularly informative , then the maximum entropy precision-recall curve need not approximate the actual precision-recall curve. Therefore , noting how closely the maximum entropy precision-recall curve corresponding to a measure approximates the precision-recall curve of the actual list, we can calculate how much information a measure contains about the actual list, and hence how "informative" a measure is. Thus, we have a methodology for evaluating the evaluation measures themselves . Using the maximum entropy precision-recall curve of a measure, we can also predict the values of other measures. For example, using the maximum entropy precision-recall curve corresponding to average precision, we can predict the precision-at-cutoff 10. For highly informative measures, these predictions should be very close to reality. Hence, we have a second way of evaluating evaluation measures. EXPERIMENTAL RESULTS We tested the performance of the evaluation measures average precision, R-precision, and precision-at-cutoffs 5, 10, 15, 20, 30, 100, 200, 500 and 1000 using data from TRECs 3, 5, 6, 7, 8 and 9. For any TREC and any query, we chose those systems whose number of relevant documents retrieved was at least 10 in order to have a sufficient number of points on the precision-recall curve. We then calculated the maximum entropy precision-recall curve subject to the given measured constraint, as described above. The maximum entropy precision-recall curve corresponding to an average precision 30 constraint cannot be determined analytically; therefore, we used numerical optimization 1 to find the maximum entropy distribution corresponding to average precision. We shall refer to the execution of a retrieval system on a particular query as a run. Figure 6 shows examples of maximum entropy precision-recall curves corresponding to average precision, R-precision, and precision-at-cutoff 10 for three different runs, together with the actual precision-recall curves. We focused on these three measures since they are perhaps the most commonly cited measures in IR. We also provide results for precision-at-cutoff 100 in later plots and detailed results for all measures in a later table. As can be seen in Figure 6, using average precision as a constraint, one can generate the actual precision-recall curve of a run with relatively high accuracy. In order to quantify how good an evaluation measure is in generating the precision-recall curve of an actual list, we consider two different error measures: the root mean squared error (RMS) and the mean absolute error (MAE). Let { 1 , 2 , . . . , R ret } be the precisions at the recall levels {1/R, 2/R, . . . , R ret /R} where R ret is the number of relevant documents retrieved by a system and R is the number of documents relevant to the query, and let {m 1 , m 2 , . . . , m R ret } be the estimated precisions at the corresponding recall levels for a maximum entropy distribution corresponding to a measure. Then the MAE and RMS errors are calculated as follows. RMS = v u u t 1 R ret R ret X i=1 ( i - m i ) 2 MAE = 1 R ret R ret X i=1 | i - m i | The points after recall R ret /R on the precision-recall curve are not considered in the evaluation of the MAE and RMS errors since, by TREC convention, the precisions at these recall levels are assumed to be 0. In order to evaluate how good a measure is at inferring actual precision-recall curves, we calculated the MAE and RMS errors of the maximum entropy precision-recall curves corresponding to the measures in question, averaged over all runs for each TREC. Figure 7 shows how the MAE and RMS errors for average precision, R-precision, precision-at-cutoff 10, and precision-at-cutoff 100 compare with each other for each TREC. The MAE and RMS errors follow the same pattern over all TRECs. Both errors are consistently and significantly lower for average precision than for the other measures in question, while the errors for R-precision are consistently lower than for precision-at-cutoffs 10 and 100. Table 1 shows the actual values of the RMS errors for all measures over all TRECs. In our experiments, MAE and RMS errors follow a very similar pattern, and we therefore omit MAE results due to space considerations. From this table, it can be seen that average precision has consistently lower RMS errors when compared to the other measures. The penultimate column of the table shows the average RMS errors per measure averaged over all TRECs. On average, R-precision has the second lowest RMS error after average precision, and precision-at-cutoff 30 is the third best measure in terms of RMS error. The last column of the table 1 We used the TOMLAB Optimization Environment for Matlab. shows the percent increase in the average RMS error of a measure when compared to the RMS error of average precision . As can be seen, the average RMS errors for the other measures are substantially greater than the average RMS error for average precision. We now consider a second method for evaluating how informative a measure is. A highly informative measure should properly reduce one's uncertainty about the distribution of relevant and non-relevant documents in a list; thus, in our maximum entropy formulation, the probability-at-rank distribution should closely correspond to the pattern of relevant and non-relevant documents present in the list. One should then be able to accurately predict the values of other measures from this probability-at-rank distribution. Given a probability-at-rank distribution p 1 , p 2 , . . . , p N , we can predict average precision, R-precision and precision-at-cutoff k values as follows: ap = 1 R N X i=1 p i i 1 + i-1 X j=1 p j !! rp = 1 R R X i=1 p i PC (k) = 1 k k X i=1 p i The plots in the top row of Figures 8 and 9 show how average precision is actually correlated with R-precision, precision-at -cutoff 10, and precision-at-cutoff 100 for TRECs 6 and 8, respectively. Each point in the plot corresponds to a system and the values of the measures are averaged over all queries. Using these plots as a baseline for comparison, the plots in the bottom row of the figures show the correlation between the actual measures and the measures predicted using the average precision maximum entropy probability-at -rank distribution. Consider predicting precision-at-cutoff 10 values using the average precision maximum entropy distributions in TREC 6. Without applying the maximum entropy method, Figure 8 shows that the two measures are correlated with a Kendall's value of 0.671. However, the precision-at-cutoff 10 values inferred from the average precision maximum entropy distribution have a Kendall's value of 0.871 when compared to actual precisions-at-cutoff 10. Hence, the predicted precision-at-cutoff 10 and actual precision-at-cutoff 10 values are much more correlated than the actual average precision and actual precision-at-cutoff 10 values. Using a similar approach for predicting R-precision and precision-at-cutoff 100, it can be seen in Figures 8 and 9 that the measured values predicted by using average precision maximum entropy distributions are highly correlated with actual measured values. We conducted similar experiments using the maximum entropy distributions corresponding to other measures, but since these measures are less informative, we obtained much smaller increases (and sometimes even decreases) in inferred correlations. (These results are omitted due to space considerations .) Table 2 summarizes the correlation improvements possible using the maximum entropy distribution corresponding to average precision. The row labeled act gives the actual Kendall's correlation between average precision and the measure in the corresponding column. The row labeled inf gives the Kendall's correlation between the 31 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 recall precision TREC8 System fub99a Query 435 AP = 0.1433 actual prec-recall ap maxent prec-recall rp maxent prec-recall pc-10 maxent prec-recall 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 recall precision TREC8 System MITSLStd Query 404 AP = 0.2305 actual prec-recall ap maxent prec-recall rp maxent prec-recall pc-10 maxent prec-recall 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 recall precision TREC8 System pir9At0 Query 446 AP = 0.4754 actual prec-recall ap maxent prec-recall rp maxent prec-recall pc-10 maxent prec-recall Figure 6: Inferred precision-recall curves and actual precision-recall curve for three runs in TREC8. TREC3 TREC5 TREC6 TREC7 TREC8 TREC9 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 Mean Absolute Error ap maxent prec-recall rp maxent prec-recall pc-10 maxent prec-recall pc-100 maxent prec-recall TREC3 TREC5 TREC6 TREC7 TREC8 TREC9 0.1 0.15 0.2 0.25 RMS Error ap maxent prec-recall rp maxent prec-recall pc-10 maxent prec-recall pc-100 maxent prec-recall Figure 7: MAE and RMS errors for inferred precision-recall curves over all TRECs. TREC3 TREC5 TREC6 TREC7 TREC8 TREC9 AVERAGE %INC AP 0.1185 0.1220 0.1191 0.1299 0.1390 0.1505 0.1298 RP 0.1767 0.1711 0.1877 0.2016 0.1878 0.1630 0.1813 39.7 PC-5 0.2724 0.2242 0.2451 0.2639 0.2651 0.2029 0.2456 89.2 PC-10 0.2474 0.2029 0.2183 0.2321 0.2318 0.1851 0.2196 69.1 PC-15 0.2320 0.1890 0.2063 0.2132 0.2137 0.1747 0.2048 57.8 PC-20 0.2210 0.1806 0.2005 0.2020 0.2068 0.1701 0.1968 51.6 PC-30 0.2051 0.1711 0.1950 0.1946 0.2032 0.1694 0.1897 46.1 PC-100 0.1787 0.1777 0.2084 0.2239 0.2222 0.1849 0.1993 53.5 PC-200 0.1976 0.2053 0.2435 0.2576 0.2548 0.2057 0.2274 75.2 PC-500 0.2641 0.2488 0.2884 0.3042 0.3027 0.2400 0.2747 111.6 PC-1000 0.3164 0.2763 0.3134 0.3313 0.3323 0.2608 0.3051 135.0 Table 1: RMS error values for each TREC. TREC3 TREC5 TREC6 RP PC-10 PC-100 RP PC-10 PC-100 RP PC-10 PC-100 act 0.921 0.815 0.833 0.939 0.762 0.868 0.913 0.671 0.807 inf 0.941 0.863 0.954 0.948 0.870 0.941 0.927 0.871 0.955 %Inc 2.2 5.9 14.5 1.0 14.2 8.4 1.5 29.8 18.3 TREC7 TREC8 TREC9 RP PC-10 PC-100 RP PC-10 PC-100 RP PC-10 PC-100 act 0.917 0.745 0.891 0.925 0.818 0.873 0.903 0.622 0.836 inf 0.934 0.877 0.926 0.932 0.859 0.944 0.908 0.757 0.881 %Inc 1.9 17.7 3.9 0.8 5.0 8.1 0.6 21.7 5.4 Table 2: Kendall's correlations and percent improvements for all TRECs. 32 0 0.1 0.2 0.3 0.4 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Actual RP Actual AP TREC 6 Actual RP vs Actual AP Kendall's = 0.913 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Actual PC-10 Actual AP TREC 6 Actual PC-10 vs Actual AP Kendall's = 0.671 0 0.1 0.2 0.3 0.4 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Actual PC-100 Actual AP TREC 6 Actual PC-100 vs Actual AP Kendall's = 0.807 0 0.1 0.2 0.3 0.4 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Actual RP Inferred RP TREC 6 Actual RP vs Inferred RP Kendall's = 0.927 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Actual PC-10 Inferred PC-10 TREC 6 Actual PC-10 vs Inferred PC-10 Kendall's = 0.871 0 0.1 0.2 0.3 0.4 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Actual PC-100 Inferred PC-100 TREC 6 Actual PC-100 vs Inferred PC-100 Kendall's = 0.955 Figure 8: Correlation improvements, TREC6. measure inferred from the maximum entropy distribution corresponding to average precision and the measure in the corresponding column. The row labeled %Inc gives the percent increase in correlation due to maximum entropy modeling . As can be seen, maximum entropy modeling yields great improvements in the predictions of precision-at-cutoff values. The improvements in predicting R-precision are no-ticeably smaller, though this is largely due to the fact that average precision and R-precision are quite correlated to begin with. CONCLUSIONS AND FUTURE WORK We have described a methodology for analyzing measures of retrieval performance based on the maximum entropy method, and we have demonstrated that the maximum entropy models corresponding to "good" measures of overall performance such as average precision accurately reflect underlying retrieval performance (as measured by precision-recall curves) and can be used to accurately predict the values of other measures of performance, well beyond the levels dictated by simple correlations. The maximum entropy method can be used to analyze other measures of retrieval performance, and we are presently conducting such studies. More interestingly, the maximum entropy method could perhaps be used to help develop and gain insight into potential new measures of retrieval performance . Finally, the predictive quality of maximum entropy models corresponding to average precision suggest that if one were to estimate some measure of performance using an incomplete judgment set, that measure should be average precision--from the maximum entropy model corresponding to that measure alone, one could accurately infer other measures of performance. Note that the concept of a "good" measure depends on the purpose of evaluation. In this paper, we evaluate measures based on how much information they provide about the overall performance of a system (a system-oriented evaluation ). However, in different contexts, different measures may be more valuable and useful, such as precision-at-cutoff 10 in web search (a user-oriented evaluation). R-precision and average precision are system-oriented measures, whereas precision-at-cutoff k is typically a user-oriented measure. Another important conclusion of our work is that one can accurately infer user-oriented measures from system-oriented measures, but the opposite is not true. Apart from evaluating the information captured by a single measure, we could use the MEM to evaluate the information contained in combinations of measures. How much does knowing the value of precision-at-cutoff 10 increase one's knowledge of a system's performance beyond simply knowing the system's average precision? Which is more informative : knowing R-precision and precision-at-cutoff 30, or knowing average precision and precision-at-cutoff 100? Such questions can be answered, in principle, using the MEM. Adding the values of one or more measures simply adds one or more constraints to the maximum entropy model, and one can then assess the informativeness of the combination. Note that TREC reports many different measures. Using the MEM, one might reasonably be able to conclude which are the most informative combinations of measures. REFERENCES [1] A. L. Berger, V. D. Pietra, and S. D. Pietra. A maximum entropy approach to natural language processing. Comput. Linguist., 22:3971, 1996. [2] C. Buckley and E. Voorhees. Evaluating evaluation measure stability. In SIGIR '00: Proceedings of the 23rd annual international ACM SIGIR conference on Research and development in information retrieval, pages 3340. ACM Press, 2000. [3] W. S. Cooper. On selecting a measure of retrieval effectiveness. part i. In Readings in information retrieval, pages 191204. Morgan Kaufmann Publishers Inc., 1997. 33 0 0.1 0.2 0.3 0.4 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Actual RP Actual AP TREC 8 Actual RP vs Actual AP Kendall's = 0.925 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Actual PC-10 Actual AP TREC 8 Actual PC-10 vs Actual AP Kendall's = 0.818 0 0.1 0.2 0.3 0.4 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Actual PC-100 Actual AP TREC 8 Actual PC-100 vs Actual AP Kendall's = 0.873 0 0.1 0.2 0.3 0.4 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Actual RP Inferred RP TREC 8 Actual RP vs Inferred RP Kendall's = 0.932 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Actual PC-10 Inferred PC-10 TREC 8 Actual PC-10 vs Inferred PC-10 Kendall's = 0.859 0 0.1 0.2 0.3 0.4 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Actual PC-100 Inferred PC-100 TREC 8 Actual PC-100 vs Inferred PC-100 Kendall's = 0.944 Figure 9: Correlation improvements, TREC8. [4] T. M. Cover and J. Thomas. Elements of Information Theory. John Wiley & sons, 1991. [5] B. Dervin and M. S. Nilan. Information needs and use. In Annual Review of Information Science and Technology, volume 21, pages 333, 1986. [6] W. R. Greiff and J. Ponte. The maximum entropy approach and probabilistic ir models. ACM Trans. Inf. Syst., 18(3):246287, 2000. [7] E. Jaynes. On the rationale of maximum entropy methods. In Proc.IEEE, volume 70, pages 939952, 1982. [8] E. T. Jaynes. Information theory and statistical mechanics: Part i. Physical Review 106, pages 620630, 1957a. [9] E. T. Jaynes. Information theory and statistical mechanics: Part ii. Physical Review 108, page 171, 1957b. [10] Y. Kagolovsky and J. R. Moehr. Current status of the evaluation of information retrieval. J. Med. Syst., 27(5):409424, 2003. [11] P. B. Kantor and J. Lee. The maximum entropy principle in information retrieval. In SIGIR '86: Proceedings of the 9th annual international ACM SIGIR conference on Research and development in information retrieval, pages 269274. ACM Press, 1986. [12] D. D. Lewis. Evaluating and optimizing autonomous text classification systems. In SIGIR '95: Proceedings of the 18th annual international ACM SIGIR conference on Research and development in information retrieval, pages 246254. ACM Press, 1995. [13] R. M. Losee. When information retrieval measures agree about the relative quality of document rankings. J. Am. Soc. Inf. Sci., 51(9):834840, 2000. [14] K. Nigam, J. Lafferty, and A. McCallum. Using maximum entropy for text classification. In IJCAI-99 Workshop on Machine Learning for Information Filtering, pages 6167, 1999. [15] D. Pavlov, A. Popescul, D. M. Pennock, and L. H. Ungar. Mixtures of conditional maximum entropy models. In T. Fawcett and N. Mishra, editors, ICML, pages 584591. AAAI Press, 2003. [16] S. J. Phillips, M. Dudik, and R. E. Schapire. A maximum entropy approach to species distribution modeling. In ICML '04: Twenty-first international conference on Machine learning, New York, NY, USA, 2004. ACM Press. [17] V. Raghavan, P. Bollmann, and G. S. Jung. A critical investigation of recall and precision as measures of retrieval system performance. ACM Trans. Inf. Syst., 7(3):205229, 1989. [18] A. Ratnaparkhi and M. P. Marcus. Maximum entropy models for natural language ambiguity resolution, 1998. [19] T. Saracevic. Evaluation of evaluation in information retrieval. In SIGIR '95: Proceedings of the 18th annual international ACM SIGIR conference on Research and development in information retrieval, pages 138146. ACM Press, 1995. [20] C. E. Shannon. A mathematical theory of communication. The Bell System Technical Journal 27, pages 379423 & 623656, 1948. [21] N. Wu. The Maximum Entropy Method. Springer, New York, 1997. 34

Average Precision;Evaluation;Maximum Entropy

19

A Similarity Measure for Motion Stream Segmentation and Recognition

Recognition of motion streams such as data streams generated by different sign languages or various captured human body motions requires a high performance similarity measure . The motion streams have multiple attributes, and motion patterns in the streams can have different lengths from those of isolated motion patterns and different attributes can have different temporal shifts and variations. To address these issues, this paper proposes a similarity measure based on singular value decomposition (SVD) of motion matrices . Eigenvector differences weighed by the corresponding eigenvalues are considered for the proposed similarity measure . Experiments with general hand gestures and human motion streams show that the proposed similarity measure gives good performance for recognizing motion patterns in the motion streams in real time.

INTRODUCTION Motion streams can be generated by continuously performed sign language words [14] or captured human body motions such as various dances. Captured human motions can be applied to the movie and computer game industries by reconstructing various motions from video sequences [10] or images [15] or from motions captured by motion capture systems [4]. Recognizing motion patterns in the streams with unsupervised methods requires no training process, and is very convenient when new motions are expected to be added to the known pattern pools. A similarity measure with good performance is thus necessary for segmenting and recognizing the motion streams. Such a similarity measure needs to address some new challenges posed by real world Work supported partially by the National Science Foundation under Grant No. 0237954 for the project CAREER: Animation Databases. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Copyright 200X ACM X-XXXXX-XX-X/XX/XX ... $ 5.00. motion streams: first, the motion patterns have dozens of attributes , and similar patterns can have different lengths due to different motion durations; second, different attributes of similar motions have different variations and different temporal shifts due to motion variations; and finally, motion streams are continuous, and there are no obvious "pauses" between neighboring motions in a stream. A good similarity measure not only needs to capture the similarity of complete motion patterns, but also needs to capture the differences between complete motion patterns and incomplete motion patterns or sub-patterns in order to segment a stream for motion recognition. As the main contribution of this paper, we propose a similarity measure to address the above issues. The proposed similarity measure is defined based on singular value decomposition of the motion matrices. The first few eigenvectors are compared for capturing the similarity of two matrices, and the inner products of the eigenvectors are given different weights for their different contributions. We propose to use only the eigenvalues corresponding to the involved eigenvectors of the two motion matrices as weights. This simple and intuitive weighing strategy gives the same importance to eigenvalues of the two matrices. We also show that the 95% variance rule for choosing the number of eigenvectors [13] is not sufficient for recognizing both isolated patterns and motion streams. Our experiments demonstrate that at least the first 6 eigenvectors need to be considered for motion streams of either 22 attribute or 54 attributes, and the first 6 eigenvalues accounts for more than 99.5% of the total variance in the motion matrices. RELATED WORK Multi-attribute pattern similarity search, especially in continuous motion streams, has been widely studied for sign language recognition and for motion synthesis in computer animation. The recognition methods usually include template matching by distance measures and hidden Markov models (HMM). Template matching by using similarity/distance measures has been employed for multi-attribute pattern recognition. Joint angles are extracted in [11] as features to represent different human body static poses for the Mahalanobis distance measure of two joint angle features. Similarly, momentum, kinetic energy and force are constructed in [2, 5] as activity measure and prediction of gesture boundaries for various segments of the human body, and the Mahalanobis distance function of two composite features are solved by dynamic programming. 89 Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MDM/KDD 2005 Chicago, August 21, Chicago, Illinois, USA Copyright 2005 ACM -- MDM 2005 - 1-59593-216-X...$5.00. Similarity measures are defined for multi-attribute data in [6, 12, 16] based on principal component analysis (PCA). Inner products or angular differences of principal components (PCs) are considered for similarity measure definitions , with different weighted strategies for different PCs. Equal weights are considered for different combinations of PCs in [6], giving different PCs equal contributions to the similarity measure. The similarity measure in [12] takes the minimum of two weighted sums of PC inner products, and the two sums are respectively weighted by different weights. A global weight vector is obtained by taking into account all available isolated motion patterns in [16], and this weight vector is used for specifying different contributions from different PC inner products to the similarity measure Eros. The dominating first PC and a normalized eigenvalue vector are considered in [7, 8] for pattern recognition. In contrast, this paper propose to consider the first few PCs, and the angular differences or inner products of different PCs are weighted by different weights which depends on the data variances along the corresponding PCs. The HMM technique has been widely used for sign language recognition, and different recognition rates have been reported for different sign languages and different feature selection approaches. Starner et al. [14] achieved 92% and 98% word accuracy respectively for two systems, the first of the systems used a camera mounted on a desk and the second one used a camera in a user's cap for extracting features as the input of HMM. Similarly Liang and Ouhyoung [9] used HMM for postures, orientations and motion primitives as features extracted from continuous Taiwan sign language streams and an average 80.4% recognition rate was achieved. In contrast, the approach proposed in this paper is an unsupervised approach, and no training as required for HMM recognizers is needed. SIMILARITY MEASURE FOR MOTION STREAM RECOGNITION The joint positional coordinates or joint angular values of a subject in motion can be represented by a matrix: the columns or attributes of the matrix are for different joints, and the rows or frames of the matrix are for different time instants. Similarity of two motions is the similarity of the resulting motion matrices, which have the same number of attributes or columns, and yet can have different number of rows due to different motion durations. To capture the similarity of two matrices of different lengths, we propose to apply singular value decomposition (SVD) to the motion matrices in order to capture the similarity of the matrix geometric structures. Hence we briefly present SVD and its associated properties below before proposing the similarity measure based on SVD in this section. 3.1 Singular Value Decomposition The geometric structure of a matrix can be revealed by the SVD of the matrix. As shown in [3], any real m n matrix A can be decomposed into A = U V T , where U = [u 1 , u 2 , . . . , u m ] R mm and V = [v 1 , v 2 , . . . , v n ] R nn are two orthogonal matrices, and is a diagonal matrix with diagonal entries being the singular values of A: 1 2 . . . min(m,n) 0. Column vectors u i and v i are the i th left and right singular vectors of A, respectively. It can be shown that the right singular vectors of the sym-metric n n matrix M = A T A are identical to the corresponding right singular vectors of A, referred to as eigenvectors of M . The singular values of M , or eigenvalues of M , are squares of the corresponding singular values of A. The eigenvector with the largest eigenvalue gives the first principal component. The eigenvector with the second largest eigenvalue is the second principal component and so on. 3.2 Similarity Measure Since SVD exposes the geometric structure of a matrix, it can be used for capturing the similarity of two matrices. We can compute the SVD of M = A T A instead of computing the SVD of A to save computational time. The reasons are that the eigenvectors of M are identical to the corresponding right singular vectors of A, the eigenvalues of M are the squares of the corresponding singular values of A, and SVD takes O(n 3 ) time for the n n M and takes O(mn 2 ) time with a large constant for the m n A, and usually m &gt; n. Ideally, if two motions are similar, their corresponding eigenvectors should be parallel to each other, and their corresponding eigenvalues should also be proportional to each other. This is because the eigenvectors are the corresponding principal components, and the eigenvalues reflect the variances of the matrix data along the corresponding principal components. But due to motion variations, all corresponding eigenvectors cannot be parallel as shown in Figure 1. The parallelness or angular differences of two eigenvectors u and v can be described by the absolute value of their inner products: | cos | = |u v|/(|u||v|) = |u v|, where |u| = |v| = 1. We consider the absolute value of the inner products because eigenvectors can have different signs as shown in [8]. Since eigenvalues are numerically related to the variances of the matrix data along the associated eigenvectors, the importance of the eigenvector parallelness can be described by the corresponding eigenvalues. Hence, eigenvalues are to be used to give different weights to different eigenvector pairs. Figure 2 shows that the first eigenvalues are the dominating components of all the eigenvalues, and other eigenvalues become smaller and smaller and approach zero. As the eigenvalues are close to zero, their corresponding eigenvectors can be very different even if two matrices are similar. Hence not all the eigenvectors need to be incorporated into the similarity measure. Since two matrices have two eigenvalues for the corresponding eigenvector pair, these two eigenvalues should have equal contributions or weights to the eigenvector parallelness . In addition, the similarity measure of two matrices should be independent to other matrices, hence only eigenvectors and eigenvalues of the two matrices should be considered . Based on the above discussions, we propose the following similarity measure for two matrices Q and P : (Q, P ) = 1 2 k i =1 (( i / n i =1 i + i / n i =1 i )|u i v i |) where i and i are the i th eigenvalues corresponding to the i th eigenvectors u i and v i of square matrices of Q and P , respectively, and 1 &lt; k &lt; n. Integer k determines how many eigenvectors are considered and it depends on the number of attributes n of motion matrices. Experiments with hand gesture motions (n = 22) and human body motions (n = 90 2 4 6 8 10 12 14 16 18 20 22 -0.7 -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 Component of First Eigenvector Component Value of First Eigenvector Motion341 Motion342 2 4 6 8 10 12 14 16 18 20 22 -0.4 -0.2 0 0.2 0.4 0.6 Component Value of Second Eigenvector Component of Second Eigenvector Motion341 Motion342 Figure 1: Eigenvectors of similar patterns. The first eigenvectors are similar to each other, while other eigenvectors, such as the second vectors shown in the bottom, can be quite different. 54) in Section 4 show that k = 6 is large enough without loss of pattern recognition accuracy in streams. We refer to this non-metric similarity measure as k Weighted Angular Similarity (kWAS) , which captures the angular similarities of the first k corresponding eigenvector pairs weighted by the corresponding eigenvalues. It can be easily verified that the value of kWAS ranges over [0,1]. When all corresponding eigenvectors are normal to each other, the similarity measure will be zero, and when two matrices are identical, the similarity measure approaches the maximum value one if k approaches n. 3.3 Stream Segmentation Algorithm In order to recognize motion streams, we assume one motion in a stream has a minimum length l and a maximum length L. The following steps can be applied to incremen-tally segment a stream for motion recognition: 1. SVD is applied to all isolated motion patterns P to obtain their eigenvectors and eigenvalues. Let be the incremented stream length for segmentation, and let L be the location for segmentation. Initially L = l. 2. Starting from the beginning of the stream or the end of the previously recognized motion, segment the stream at location L. Compute the eigenvectors and eigenvalues of the motion segment Q. 3. Compute kWAS between Q and all motion patterns P . Update max to be the highest similarity after the previous motion's recognition. 4. If L+ &lt; L, update L = L+ and go to step 2. Otherwise , the segment corresponding to max is recognized to be the motion pattern which gives the highest similarity max , update L = l starting from the end of the last recognized motion pattern and go to step 2. 1 2 3 4 5 6 7 8 85 87 89 91 93 95 97 99 100 Number of Eigenvalues Accumulated Eigenvalue Percentage (%) CyberGlove Data MoCap Data Figure 2: Accumulated eigenvalue percentages in total eigenvalues for CyberGlove data and captured human body motion data. There are 22 eigenvalues for the CyberGlove data and 54 eigenvalues for the captured motion data. The sum of the first 2 eigenvalues is more than 95% of the corresponding total eigenvalues, and the sum of the first 6 eigenvalues is almost 100% of the total eigenvalues. PERFORMANCE EVALUATION This section evaluates experimentally the performances of the similarity measure kWAS proposed in this paper. It has been shown in [16] that Eros [16] outperforms other similarity measures mentioned in Section 2 except MAS [8]. Hence in this section, we compare the performances of the proposed kWAS with Eros and MAS for recognizing similar isolated motion patterns and for segmenting and recognizing motion streams from hand gesture capturing CyberGlove and human body motion capture system. 4.1 Data Generation A similarity measure should be able to be used not only for recognizing isolated patterns with high accuracy, but also for recognizing patterns in continuous motions or motion streams. Recognizing motion streams is more challenging than recognizing isolated patterns. This is because many very similar motion segments or sub-patterns needs to be compared in order to find appropriate segmentation locations , and a similarity measure should capture the difference between a complete motion or pattern and its sub-patterns. Hence, both isolated motion patterns and motion streams were generated for evaluating the performance of kWAS. Two data sources are considered for data generation: a CyberGlove for capturing hand gestures and a Vicon motion capture system for capturing human body motions. 4.1.1 CyberGlove Data A CyberGlove is a fully instrumented data glove that provides 22 sensors for measuring hand joint angular values to capture motions of a hand, such as American Sign Language (ASL) words for hearing impaired. The data for a hand gesture contain 22 angular values for each time instant/frame, one value for a joint of one degree of freedom. The motion data are extracted at around 120 frames per second. Data matrices thus have 22 attributes for the CyberGlove motions. One hundred and ten different isolated motions were generated as motion patterns, and each motion was repeated for three times, resulting in 330 isolated hand gesture motions . Some motions have semantic meanings. For example, 91 the motion for BUS as shown in Table 1 is for the ASL sign "bus". Yet for segmentation and recognition, we only require that each individual motion be different from others, and thus some motions are general motions, and do not have any particular semantic meanings, such as the THUMBUP motion in Table 1. The following 18 motions shown in Table 1 were used to generate continuous motions or streams. Twenty four different motion streams were generated for segmentation and recognition purpose. There are 5 to 10 motions in a stream and 150 motions in total in 24 streams, with 6.25 motions in a stream on average. It should be noted that variable-length transitional noises occur between successive motions in the generated streams. Table 1: Individual motions used for streams 35 60 70 80 90 BUS GOODBYE HALF IDIOM JAR JUICE KENNEL KNEE MILK TV SCISSOR SPREAD THUMBUP 4.1.2 Motion Capture Data The motion capture data come from various motions captured collectively by using 16 Vicon cameras and the Vicon iQ Workstation software. A dancer wears a suit of non-reflective material and 44 markers are attached to the body suit. After system calibration and subject calibration, global coordinates and rotation angles of 19 joints/segments can be obtained at about 120 frames per second for any motion . Similarity of patterns with global 3D positional data can be disguised by different locations, orientations or different paths of motion execution as illustrated in Figure 3(a). Since two patterns are similar to each other because of similar relative positions of corresponding body segments at corresponding time, and the relative positions of different segments are independent of locations or orientations of the body, we can transform the global position data into local position data as follows. Let X p , Y p , Z p be the global coordinates of one point on pelvis, the selected origin of the "moving" local coordinate system, and , , be the rotation angles of the pelvis segment relative to the global coordinate system axes, respectively . The translation matrix is T as follows: T = 1 0 0 0 0 1 0 0 0 0 1 0 -X p -Y p -Z p 1 The rotation matrix R = R x R y R z , where R x = 1 0 0 0 0 cos - sin 0 0 sin cos 0 0 0 0 1 R y = cos 0 sin 0 0 1 0 0 - sin 0 cos 0 0 0 0 1 0 50 100 150 200 250 300 350 400 450 -1500 -1000 -500 0 500 1000 1500 0 50 100 150 200 250 300 350 400 450 0 500 1000 1500 2000 Motion Capture Frames Global Coordinates of Joints(mm) Global Coordinates of Joints(mm) (a) 0 50 100 150 200 250 300 350 400 450 -1000 -500 0 500 1000 Transformed Coordinates of Joints (mm) 0 50 100 150 200 250 300 350 400 450 -1000 -500 0 500 1000 Motion Capture Frames Transformed Coordinates of Joints (mm) (b) Figure 3: 3D motion capture data for similar motions executed at different locations and in different orientations : (a) before transformation; (b) after transformation . R z = cos - sin 0 0 sin cos 0 0 0 0 1 0 0 0 0 1 Let X, Y, Z be the global coordinates of one point on any segments, and x, y, z be the corresponding transformed local coordinates. x, y and z can be computed as follows: [x y z 1] = [X Y Z 1] T R The transformed data are positions of different segments relative to a moving coordinate system with the origin at some fixed point of the body, for example the pelvis. The moving coordinate system is not necessarily aligned with the global system, and it can rotate with the body. So data transformation includes both translation and rotation, and the transformed data would be translation and rotation invariant as shown in Figure 3(b). The coordinates of the origin pelvis are not included, thus the transformed matrices have 54 columns. Sixty two isolated motions including Taiqi, Indian dances, and western dances were performed for generating motion capture data, and each motion was repeated 5 times, yielding 310 isolated human motions. Every repeated motion has a different location and different durations, and can face different orientations. Twenty three motion streams were generated for segmentation. There are 3 to 5 motions in a stream, and 93 motions in total in 23 streams, with 4.0 motions in a stream on average. 4.2 Performance of k WAS for Capturing Similarities and Segmenting Streams We first apply kWAS to isolated motion patterns to show that the proposed similarity measure kWAS can capture the similarities of isolated motion patterns. Then kWAS is applied to motion streams for segmenting streams and recognizing motion patterns in the streams. We experimented with different k values in order to find out the smallest k without loss of good performance. Figure 2 shows the accumulated eigenvalue percentages averaged on 330 hand gestures and 310 human motions, respectively . Although the first two eigenvalues account for 92 1 2 90 91 92 93 94 95 96 97 98 99 100 Number of Nearest Neighbors (Most Similar Patterns) Pattern Recognition Rate (%) kWAS (k = 22) kWAS (k = 5) kWAS (k = 3) kWAS (k = 2) MAS EROS Figure 4: Recognition rate of similar CyberGlove motion patterns. When k is 3, kWAS can find the most similar motions for about 99.7% of 330 motions , and can find the second most similar motions for 97.5% of the them. 1 2 3 4 95 95.5 96 96.5 97 97.5 98 98.5 99 99.5 100 Number of Nearest Neighbors (Most Similar Patterns) Pattern Recognition Rate (%) kWAS (k = 54) kWAS (k = 5) kWAS (k = 4) kWAS (k = 3) MAS EROS Figure 5: Recognition rate of similar captured motion patterns. When k is 5, by using kWAS, the most similar motions of all 310 motions can be found, and the second most similar motions of 99.8% of the 310 motions can also be found. more than 95% of the respective sums of all eigenvalues, considering only the first two eigenvectors for kWAS is not sufficient as shown in Figure 4 and Figure 5. For CyberGlove data with 22 attributes, kWAS with k = 3 gives the same performance as kWAS with k = 22, and for motion capture data with 54 attributes, kWAS with k = 5 gives the same performance as kWAS with k = 54. Figure 4 and Figure 5 illustrate that kWAS can be used for finding similar motion patterns and outperforms MAS and Eros for both hand gesture and human body motion data. The steps in Section 3.3 are used for segmenting streams and recognizing motions in streams. The recognition accuracy as defined in [14] is used for motion stream recognition. The motion recognition accuracies are shown in Table 2. For both CyberGlove motion and captured motion data, k = 6 is used for kWAS, which gives the same accuracy as k = 22 for CyberGlove data and k = 54 for motion capture data, respectively. Figure 6 shows the time taken for updating the candidate segment, including updating the matrix, computing the SVD of the updated matrix, and computing the similarities of the segment and all motion patterns. The code implemented in C++ was run on one 2.70 GHz Intel processor of a GenuineIntel Linux box. There are 22 attributes for the CyberGlove streams, and 54 attributes for the captured CyberGlove Streams Motion Capture Streams 0 2 4 6 8 10 12 14 16 18 20 Time (milliseconds) MAS kWAS (k = 6) EROS Figure 6: Computation time for stream segment update and similarity computation. Table 2: Stream Pattern Recognition Accuracy (%) Similarity CyberGlove Motion Capture Measures Streams Streams Eros 68.7 78.5 MAS 93.3 78.5 kWAS (k=6) 94.0 94.6 motion streams. Hence updating captured motion segments takes longer than updating CyberGlove motion segments as shown in Figure 6. The time required by kWAS is close to the time required by MAS, and is less than half of the time taken by using Eros. 4.3 Discussions k WAS captures the similarity of square matrices of two matrices P and Q, yet the temporal order of pattern execution is not revealed in the square matrices. As shown in [7], two matrices with the identical row vectors in different orders have identical eigenvectors and identical eigenvalues. If different temporal orders of pattern execution yield patterns with different semantic meanings, we need to further consider the temporal execution order, which is not reflected in the eigenvectors and eigenvalues and has not been considered previously in [6, 12, 16]. Since the first eigenvectors are close or parallel for similar patterns, we can project pattern A onto its first eigenvector u 1 by Au 1 . Then similar patterns would have similar projections (called projection vectors hereafter), showing similar temporal execution orders while the projection variations for each pattern can be maximized. The pattern projection vectors can be compared by computing their dynamic time warping (DTW) distances, for DTW can align sequences of different lengths and can be solved easily by dynamic programming [1]. Incorporating temporal order information into the similarity measure can be done as for MAS in [7] if motion temporal execution orders cause motion pattern ambiguity to kWAS. CONCLUSIONS This paper has proposed a similarity measure kWAS for motion stream segmentation and motion pattern recognition . kWAS considers the first few k eigenvectors and computes their angular similarities/differences, and weighs contributions of different eigenvector pairs by their correspond-93 ing eigenvalues. Eigenvalues from two motion matrices are given equal importance to the weights. Experiments with CyberGlove hand gesture streams and captured human body motions such as Taiqi and dances show that kWAS can recognize 100% most similar isolated patterns and can recognize 94% motion patterns in continuous motion streams. REFERENCES [1] D. Berndt and J. Clifford. Using dynamic time warping to find patterns in time series. In AAAI-94 Workshop on Knowledge Discovery in Databases, pages 229248, 1994. [2] V. M. Dyaberi, H. Sundaram, J. James, and G. Qian. Phrase structure detection in dance. In Proceedings of the ACM Multimedia Conference 2004, pages 332335, Oct. 2004. [3] G. H. Golub and C. F. V. Loan. Matrix Computations. The Johns Hopkins University Press, Baltimore,Maryland, 1996. [4] L. Ikemoto and D. A. Forsyth. Enriching a motion collection by transplanting limbs. In Proceedings of the 2004 ACM SIGGRAPH/Eurographics symposium on Computer animation, pages 99 108, 2004. [5] K. Kahol, P. Tripathi, S. Panchanathan, and T. Rikakis. Gesture segmentation in complex motion sequences. In Proceedings of IEEE International Conference on Image Processing, pages II 105108, Sept. 2003. [6] W. Krzanowski. Between-groups comparison of principal components. J. Amer. Stat. Assoc., 74(367):703707, 1979. [7] C. Li, B. Prabhakaran, and S. Zheng. Similarity measure for multi-attribute data. In Proceedings of the 2005 IEEE International Conference on Acoustics, Speach, and Signal Processing (ICASSP), Mar. 2005. [8] C. Li, P. Zhai, S.-Q. Zheng, and B. Prabhakaran. Segmentation and recognition of multi-attribute motion sequences. In Proceedings of the ACM Multimedia Conference 2004, pages 836843, Oct. 2004. [9] R. H. Liang and M. Ouhyoung. A real-time continuous gesture recognition system for sign language. In Proceedings of the 3rd. International Conference on Face and Gesture Recognition, pages 558565, 1998. [10] K. Pullen and C. Bregler. Motion capture assisted animation: texturing and synthesis. In SIGGRAPH, pages 501508, 2002. [11] G. Qian, F. Guo, T. Ingalls, L. Olson, J. James, and T. Rikakis. A gesture-driven multimodal interactive dance system. In Proceedings of IEEE International Conference on Multimedia and Expo, June 2004. [12] C. Shahabi and D. Yan. Real-time pattern isolation and recognition over immersive sensor data streams. In Proceedings of the 9th International Conference on Multi-Media Modeling, pages 93113, Jan 2003. [13] A. Singhal and D. E. Seborg. Clustering of multivariate time-series data. In Proceedings of the American Control Conference, pages 39313936, 2002. [14] T. Starner, J. Weaver, and A. Pentland. Real-time american sign language recognition using desk and wearable computer based video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(12):13711375, 1998. [15] C. J. Taylor. Reconstruction of articulated objects from point correspondences in a single image. Computer Vision and Image Understanding, 80(3):349363, 2000. [16] K. Yang and C. Shahabi. A PCA-based similarity measure for multivariate time series. In Proceedings of the Second ACM International Workshop on Multimedia Databases, pages 6574, Nov. 2004. 94

motion stream;segmentation;data streams;eigenvector;singular value decomposition;gesture;recognition;eigenvalue;similarity measure;Pattern recognition

190

The Model, Formalizing Topic Maps

This paper presents a formalization for Topic Maps (TM). We first simplify TMRM, the current ISO standard proposal for a TM reference model and then characterize topic map instances. After defining a minimal merging operator for maps we propose a formal foundation for a TM query language. This path expression language allows us to navigate through given topic maps and to extract information. We also show how such a language can be the basis for a more industrial version of a query language and how it may serve as foundation for a constraint language to define TM-based ontologies.

Introduction Topic Maps (TM (Pepper 1999)), a knowledge representation technology alternative to RDF (O. Lassila and K. Swick 1993), have seen some industrial adoption since 2001. Concurrently, the TM community is taking various efforts to define a more fundamental , more formal model to capture the essence of what Topic Maps are (Newcomb, Hunting, Algermissen & Durusau 2003, Kipp 2003, Garshol 2004-07-22 , Bogachev n.d.). While the degree of formality and the extent of TM machinery varies, all models tend to abstract away from the sets of concepts defined in (Pepper 2000) and use assertions (and topics) as their primitives. After giving an overview over the current state of affairs, we start with an attempt to conceptually simplify the TMRM (Newcomb et al. 2003) model. From that, a mathematically more rigorous formalization of TMs follows in section 4. Based on maps and elementary map composition we define a path expression language using a postfix notation. While low-level, it forms the basis for querying and constraining topic maps as we point out in section 6. The last section closes with future research directions. Related Work Historically, Topic Maps, being a relatively new technology , had some deficits in rigor in terms of a defining model. This may be due to the fact that it was more Paradoxically, the standardization efforts started out with the syntax (XTM) with only little, informal description of the individual constructs. TMDM (formerly known as SAM) was supposed to fill this role by precisely defining how XTM instances are to be deserialized into a data structure. This is done by mapping the syntax into an infoset model (comparable to DOM) whereby UML diagrams help to illustrate the intended structure as well as the constraints put on it. While such an approach to model definition has a certain appeal for (Java) developers, its given complexity puts it well outside the reach for a more mathematical formalization. In parallel a fraction within the TM community ar-gued that the TM paradigm can be interpreted on a much more fundamental level if one considers assertions as the basic building blocks, abstracting from the TAO-level which mainly sees topics with their names, occurrences and involvements in associations. This group has developed several generations of the TMRM (Newcomb et al. 2003), the reference model. The model therein is mainly based on graph theory mixed with informal descriptions of constraints which cover the resolution of subject identity. Several attempts to suggest an alternative founda-tional model (Garshol 2004-07-22, Bogachev n.d.) or to formalize TMRM have been made. (Kipp 2003) is successfully using a purely set-theoretic approach to define topic map instances. As all TMRM concepts have been faithfully included, this resulted in a significant set of constraints to be used when reasoning about map instances. The contribution of this paper we see threefold: Firstly, we believe that TMRM can be reasonably simplified without any loss of generality by the steps outlined in section 3. This is under the assumption that all questions of subject identity are handled outside the model. Secondly, the assertion model seems to be general enough to host conceptually not only TMRM, but also serve as basis for TMDM. As the TM community now moves to ontology definition languages, retrieval and transformation languages , we contend that the path language which is based on the model can serve as semantic fundament Conceptual Simplification TMRM's main building blocks are properties which are attached to topics and assertions which connect topics in various ways. 3.1 Properties For properties TMRM distinguishes between subject identifying properties and other properties. The for-37 mer can be stand-alone or a combination of other properties; they control--for a given application-under which conditions two topics should be regarded the same. With the assumption that all identity inducing constraints are best covered by a proper ontology language , we drop this distinction. Also conferred properties can be handled much more flexibly with an ontology language, which allows us to let conferred and builtin properties collapse. We abstract further by regarding properties just as a special form of binary assertions where the topic plays a role object and the property forms the other member of the assertion. 3.2 Assertions A TMRM assertion stands for a statement between subjects whereby these subjects play certain roles. Such an assertion consists of the subject it is about and a type. Additionally, the players are cast into their respective roles. To be able to reify the fact that a certain topic plays a certain role in an assertion , also this substatement is represented by a another topic (casting). We observe that any type information for an assertion a can be represented by a second, dedicated assertion b where a plays the instance and that type plays the role class. A similar consideration applies to casting topics: again, a second, dedicated assertion can be used where the role, the assertion and the player are playing appropriate roles. Scoping--the restriction of an assertion to a certain context--is clearly a statement about an assertion , so we can represent scoping relations via a further assertion, one which connects the original assertion with the scope itself, again via some predefined roles. At the end of this process we only have to deal with assertions containing role-player pairs. Assertions have an identity which allows us to use them in other assertions. Topics only exist as focal points and have no explicit property except an identifier. 3.3 Reification The term reification has a long tradition (Sowa 2000) in the knowledge representation community. It has changed its meaning over the years, but it is usu-ally used to describe how humans form concepts and then connect them with the `real world'. To fully capture the term formally, we would have to adopt a philosophical approach, something which we prefer to avoid for obvious reasons. The question, though, is whether any formalization of TMs can completely ignore reification. Whenever a statement S is about another assertion A then one of two things could be intended by the author: either (a) S is a statement about the relationship in the `real world' A is supposed to represent. As an example consider that A is about an employment of a person within an organisation and that we want to qualify in such that "the employment only started in year 2000". Alternatively (b), a statement can be about the assertion within the map itself, such as "this assertion was commented on by user X". In the latter case we treat A as if it were in the `real world' (inverting somehow the notion of reification by pulling something abstract from a concept space and making it `real'). Our--pragmatic--approach is that this distinction can (and should) be indicated by the proper form of identifiers. If a topic is supposed to reify a real world concept, then its identifier should be a URI (a locator or a name), in case that the `real world thing' has one. If that thing is a topic in a topic map, then the author must have a way to address the map as well as the topic within it. If a direct reification is not possible, then the topic's identifier will simply not be a URI. Indirect identification can be achieved via subject indicators attached to the topic or more generally speaking by the context the topic is in. For assertions we assume that they--as a whole-implicitly reify the relationship they describe. If another assertion makes a reference to an assertion then using the assertion's identifier may thus automatically cover case (a) above. Like with topics, case (b) can be handled by using an identifier which addresses the map and then the assertion within it. How eventually maps as `real world' objects are to be addressed is again a matter how identifiers are formed; but this is outside the scope of our model. Formal Maps In this section we first prepare the grounds by defining identifiers, then we build members and assertions and then finally maps. For presentation, the text here has two layers, one for the formal part and an informal one, shaded grey. The latter is to justify design issues or present examples. 4.1 Identifiers The set of identifiers, I, contains two sets of objects : names and literals. Literals may be numbers or quoted strings. The set of names, N , is an enu-merable collection of atomic objects. Atomic means that objects have no other properties than being dis-tinguishable from each other. In practical situations names may be strings such as URIs. They also may be more complex like XLink or even HyTime pointers. The model only uses the property that they are distinct from each other. The reason we chose literals to be numbers or strings is simply one of convenience. First, these two basic data types are the most frequently used, and secondly, both have naturally defined an ordering a b on which we can later base sorting. One issue with selecting a particular set of primitive data types is that of how to represent others, like composite types as one would need for, say, spatial coordinates . We see two approaches: One way is to model the content explicitly with assertions themselves. The other option can be used if the structure of the data is not specifically relevant to a particular application, but has to be kept in a map for archiving purposes. In these situations data can be serialized into a string and treated as such. Further we assume that I also contains a small set of predefined identifiers, id, instance, class, subclass, superclass. By themselves, they are not special. We only single them out to be able to define additional semantics later. 4.2 Members and Assertions As we are mainly interested in expressing associative relations, we first define a member to be a pair r, p (N � I), with r being the role and p the player of the member. An assertion a is a finite (possibly empty) set of members. The set of all assertions is denoted by A. 38 Assertions always have an identity. It is a function id(a) over the set of members of a, whereby we only request that different member sets result in different identities. Obviously, assertions are only equal if they have identical members. To access the components of an assertion a we define the set roles(a) = {r 1 , . . . , r n } with r i being the roles in the individual members of a, and the set players(a) = {p 1 , . . . , p n } with p i being the players in a. Note that in assertions players are not grouped around a role. If several players play one and the same role, then individual members have to exist for every such player. Also note, that assertions do not have a type component; it is up to a further assertion to establish such a relationship whereby the predefined identifiers instance and class can be used as roles. The base model does not impose any restrictions on players and roles. While not necessary for the formalism itself, we might later want to put additional constraints on the form of assertions to only meaningful combinations. Examples of such meaningful constraints are "there may be only one player for a particular role" or "in one and the same assertion a particular identifier cannot be used as role and as player": a A, roles(a) players(a) = . Another useful constraint could avoid that the identifier for an assertion appears in that assertion itself: a A, id(a) / (roles(a) players(a)). This assertion structure proves to be central to the whole model. It is sufficiently flat as there is no distinction between assertions and properties. The focus on assertions alone also reduces topics to identifiers . Still, the chosen structure seems to incorporate enough of the TM paradigm, in that any number of concepts can be bound together into an assertion and topics--as TMRM mandates--can function as the sole aggregation point for information. 4.3 Maps We now consider assertions to be atoms from which maps can be constructed. A map is a finite (possibly empty) set of of assertions. The set of all maps is denoted by M. To build bigger maps, we define the elementary composition, denoted by , of two maps m, m M, is defined as set union m m = m m . We say that m is a submap of m if m m . Note that we have no special merging operation; only exactly identical assertions will be identified. In our setting special-purpose merging, such as TNC (topic name constraint), is split into two phases: first maps are combined using elementary composition and then a second operator is applied to the composite map. That operator will perform a--more or less sophisticated--transformation where all the appropriate merging is done. As an example we consider a network which hosts several servers, organized into clusters (Table 1). At a particular point in time, servers may be &quot;up&quot; or &quot;down&quot; . Accordingly, macy, lacy and stacy are the servers, the first two being in clusterA, the other in clusterB . While lacy is down, clusterA is still functional , not so clusterB as its only machine is down. 4.4 Primitive Navigation Operators To navigate through maps and to extract information out of them, we first need to define basic navigation operations within a given map. In our model we can navigate along roles. One way is to follow a role outwards in a given assertion a m. Given additionally a name r we define the role-out operator a r = {p | r, p a}. It returns all players of a given role in an assertion. Looking at a00 in the above example, the expression a00 class returns the set containing server only. Another option to navigate is to follow a role inwards , seen from an assertion's point of view. Given a map m, a name r and an identifier p, we define the role-in operator p m r = {a m | r, p a}. We omit the reference to m if clear from the context. To find all assertions in which clusterA plays the role whole, we can write clusterA whole which evaluates to {a02, a11}. The role-in operator does not respect the type of assertions . It simply finds all assertions where a particular player plays the given role. However, for practical reasons a refined version of the operator will be defined in section 4.6. 4.5 Subclassing and Instances To describe (and query) topic maps, we need to express relationships between concepts. While the variety of such relations itself is huge, two special relationships stand out as being fundamental: The subclass-superclass relationship is used between classes to form taxonomies (type systems). The instance-class relationship is established between an object and the class (or set) the object can be classified into. Given a map m and names b, c N , we define the predicate subclasses m (b, c) to be true if there exists an a m such that both conditions, a subclass = {b} and a superclass = {c}, hold. As the usual interpretation of subclassing is that it is transitive, we build the transitive closure subclasses m+ and the transitive, reflexive closure subclasses m . Another relationship is instance of, abbreviated as is - a which holds if there exists a m such that a instance = {b} and a class = {c}. Mostly we are interested in an instance-of relationship which includes the transitive version of subclassing above. is - a m (b, c) holds if there exist a m such that for some name c we have a instance = {b}, a class = {c } and subclasses m (c , c). According to our cluster map the relations subclasses m (server, machine), is - a m (macy, server) and is - a m (macy, machine) are all true. The difference between is - a m (b, c) and is - a m (b, c) is that the former only reiterates the information which is already explicit in the map. When querying a map, though, queries should be built more robust: If we ask for "all machines" in a map, then most likely one is also interested in instances of all (direct and indirect) subclasses of "machine". 39 Table 1: An example map about a computer network a00 = { &lt; instance, macy &gt;, &lt; class, server &gt; } a01 = { &lt; instance, a00 &gt;, &lt; class, isInstance &gt; } a02 = { &lt; part, macy &gt;, &lt; whole, clusterA &gt; } a03 = { &lt; instance, a02 &gt;, &lt; class, isPartOf &gt; } a04 = { &lt; object, macy &gt;, &lt; status, &quot;up&quot; &gt; } a05 = { &lt; instance, a04 &gt;, &lt; class, hasStatus &gt; } ... a10 = { &lt; instance, lacy &gt;, &lt; class, server &gt; } a11 = { &lt; part, lacy &gt;, &lt; whole, clusterA &gt; } a12 = { &lt; object, lacy &gt;, &lt; status, &quot;down&quot; &gt; } ... a20 = { &lt; instance, stacy &gt;, &lt; class, server &gt; } a21 = { &lt; part, stacy &gt;, &lt; whole, clusterB &gt; } a22 = { &lt; object, stacy &gt;, &lt; status, &quot;down&quot; &gt; } a30 = { &lt; subclass, server &gt;, &lt; superclass, machine &gt; } a40 = { &lt; instance, clusterA &gt;, &lt; class, cluster &gt; } a41 = { &lt; instance, clusterB &gt;, &lt; class, cluster &gt; } 4.6 Typed Navigation We can use the relation is - a m (b, c) to specialize the role-in navigation. Given a map m, names r and t and an identifier p the typed role-in operator honors additionally an assertion type: p m r [t] = {a p m r | is - a m (id(a), t)} (1) The obvious difference to the original role-in navigation is that we now only consider assertions of the given type to be part of the resulting set. The expression clusterA m whole [hasStatus] is supposed to find all assertions of type hasStatus in which clusterA is the whole. Since there is no such assertion, the result is empty. A further way to generalize the navigation is to allow as role also all subclasses: a m r = {p | r , p a : subclasses m (r , r)} (2) p m r = {a m | r , p a : subclasses m (r , r)} (3) Map Path Language The topic map path language can be used to extract information out of given map. The language will be defined via postfix operators which are applied to (sets of) assertions (or identifiers). Before we can formally define the individual postfixes and chains of postfixes (path expressions) we have to characterize the results of applying postfixes to a set of assertions, such as a map. This is done with a simple algebra based on tuples. 5.1 Tuple Algebra Our final result of applying a path expression will be a bag of tuples. The advantage of tuples are that they can hold composite results. Every tuple represents then one possible result, all of them are organized into a bag. Bags are like sets except that a particular element may appear any number of times. This is convenient if we later want to sort or count the tuples. Otherwise all the usual set operations can be used on bags. Assertion tuples are elements from the cartesian product A n with A being the set of assertions. Simi-larily , identifier tuples are elements from I n . We call n the dimension of the tuple. When we organize tuples t 1 , . . . , t n into a bag, then we denote this as [t 1 , . . . , t n ]. A map m = {a 1 , . . . , a n } can be represented as the tuple bag [ a 1 , . . . , a n ]. Conversely, we can also interpret a tuple bag as map when the tuples it contains are single assertions. If a bag contains other bags, then the structure can be flattened out: [b 1 , b 2 , . . . , b n ] = [b ij | b ij b i (1 i n)] (4) During application of path expressions also tuples of bags may be created. Also these can be reduced by building tuples of all combinations of bag elements: b 1 , b 2 , . . . , b n = b 1 � b 2 � � � � � b n (5) Finally, if a tuple only contains a single component , then it is equivalent to that component: b = b (6) As we have covered all possible constellations which can occur when evaluating path expressions, we can always reduce every result to a bag of tuples. We call this set B I . 5.2 Postfixes and Path Expressions Individual postfixes (as detailed below) can be combined to form chains. The set of path expressions P M is defined as the smallest set satisfying the following conditions: 1. The projection postfix i is in P M for any non-negative integer i. 2. Every identifier from I is in P M . 3. The role-out and role-in postfixes r and r for a name r are in P M . 40 4. The positive predicate postfix [ p = q ] and the negative predicate postfix [ p != q ] are both in P M for two path expressions p and q. As special cases we also include [ p ] and [ !p ]. 5. For two path expressions p and q also the concatenation p � q is in P M . If - from the context - it is clear that two path expressions are to be concatenated, we omit the infix. 6. For two path expressions p and q the alternation p q is in P M . The application of a path expression p to a map m is denoted by m p. For this process, first we will reinterpret the map as tuple bag. Then each of the postfixes in p is applied to it. Each such step results in a new bag which will be flattened according to the tuple algebra above. The final bag will be the overall result. 5.2.1 Projection and Identifiers For both, assertion and identifier tuples, we will use the projection postfix to extract a particular j: u 1 , . . . , u n j = [ u j ] (7) Projection here plays a similar role like in query languages like SQL, except that we here use an index for selection instead of names. We drop the index 1 in 1 if it is applied to a tuple with only a single component where then obviously it holds that u = u . Such a projection also serves as the empty postfix. In case the path expression is simply an identifier i I, then for any u the result is always this identifier : u i = [ i ] (8) 5.2.2 Concatenation and Alternation We define the concatenation � of path expressions p and q (given any u) as u (p � q) = (u p) q (9) The syntactic structure of path expressions ensures that u is always a structure for which such an evaluation is defined. The alternation of two path expressions p and q is defined as the union of the result tuple bags of the individual evaluations: u (p q) = u p u q (10) 5.2.3 Navigation Postfix Next we define how role-out and role-in navigation postfixes can be applied to an assertion tuple. We simply apply the navigation to every assertion in the tuple: a 1 , . . . , a n r = a 1 r , . . . , a n r (11) p 1 , . . . , p n r = p 1 r , . . . , p n r (12) Note that we have used the typed navigation from section 4.6. While not absolutely necessary, it helps to keep path expressions more concise. Note also, that the individual elements of the resulting tuples are bags. Again, the transformation rules of the tuple algebra have to be used to reduce this into a bag of tuples. 5.2.4 Filtering Postfixes From tuple bags we can filter out specific tuples using predicates. Given a tuple bag B = [t 1 , . . . , t k ] and two path expressions p and q, applying the positive predicate postfix [ p = q ] to B is defined as B [ p = q ] = [t B | t p t q = ] (13) If p and q are identical, then we can abbreviate [ p = p ] with [ p ]. The result of the positive predicate prefix is that sub-bag of B for which elements the evaluation of p and q gives at least one common result. Note that this implements an exists semantics as B [ p = p ] is reducable to [t B | t p = ]. Only those tuples of B will be part of the result tuple bag if there exists at least one result when p is applied to that tuple. By introducing negation in predicate postfixes, we can also implement forall semantics. Given a tuple bag B and two path expressions p and q, we define the negative predicate postfix as B [ p != q ] = [t B | t p t q = ] (14) If p and q are identical, then we can abbreviate [ p != p ] with [ ! p ]. In this case the result tuple bag becomes [t B | t p = ]. A particular tuple will only then be part of the result tuple bag if p applied to it will not render a single value, i.e. all evaluations will return no result. Implicit in the formalism are the logic conjunction and disjunction of predicate postfixes. Obviously, a logical and is provided by concatenating two predicate postfixes ([ .. ] � [ .. ]) as the result of the first postfix will be further tested for the second predicate. The logical or between predicate postfixes is implicitly given by alternating them ([ .. ] [ .. ]). 5.3 Evaluation Example Let us assume that we are looking for the status of the servers in clusterA: [ class = isPartOf] instance [ whole = clusterA] part object &lt; object , status &gt; The first predicate selects out all those assertions in the map which have a class role where one of the players happens to be isPartOf. If we are then looking at these assertions and the player(s) of the role instance , then we have effectively selected the assertions of type isPartOf from the map. We consider each of these assertions (in our case these are a02, a11 and a21) and filter out those of them which have a whole role where one player is clusterA. When we continue with a02 and a11, and then follow the part, this leads to a bag containing only the names macy and lacy. 41 In the next step we investigate where these names are players of the role object, so we find a bag with assertions a04 and a12. Here our path splits into two components: the first one navigates to the name of that object, the other to its status. The result is then [ macy, "up" , lacy, "down" ]. In second example we look at all clusters which are down, i.e. where all machines in that cluster are down. As result we get [ clusterB ]: [ class = cluster ] instance [ whole part object status ! = "up"] Querying, Filtering and Constraining of Maps Maps and path expressions, as presented here, can serve as a basis for more high-level concepts, as they are needed for ontology and knowledge engineering (Fensel, Hendler, Lieberman & Wahlster 2003). The use of path expressions to extract information out of maps leads to the following observations: Obviously, P M is a (primitive) language to query topic maps. Note, though, that P M lacks all facilities to newly create content, such as XML or TM content as described in (Garshol & Barta 2003). A more industrial topic map query language (TMQL) will have to offer content generation language constructs . While it will also provide more concise syntax due to high-level concepts, P M can (and probably will) act as a semantic foundation. More formally, we can identify a subset of P M , the filters F M , which contains all those queries which return maps: F M = {q P M | m M, mq = [ a 1 . . . a n | a i m]} (15) Clearly, the filtered maps are always submaps of the queried map: m f m, for f F M . Interestingly, P M can also be regarded as primitive constraint language: only when the application of a path expression c to a map m renders any result, then the map conforms to the expectations we have set up in c. If, for instance, we had set up a query which asks for all weapons of mass destruction in our running example, then the result would have been the empty bag. Only if the query follows the structure and the vocabulary of the map, then there will be a non-empty result. Equivalently, this is also true the other way round. Consequently, we can define a satisfaction relation |= P M M between a path expression c and a map m, such that c |= m m c = (16) Based on this, logical connectives between constraints can be defined. Future Work While we concentrate in this work on formalizing the structure of topic maps (at least our understanding thereof) and of an expression language to extract information from them, we have not yet studied any properties of P M . Specifically, we are interested how path expressions relate to formulas in description logics (Baader, Calvanese, McGuinness, Nardi & Patel-Schneider 2003, Description Logics Home Page n.d.), especially in the light that both can be used to model an ontology. A related question is how a path language can be used to express identity (apart from the explicit identity given by the topic's identifier). Finally, in a larger picture, we are interested in connecting maps, constraints, queries and even maybe updates for topic maps in an algebra. When connecting maps, merging as defined by the XTM standard is an issue. References Baader, F., Calvanese, D., McGuinness, D., Nardi, D. & Patel-Schneider, P., eds (2003), The Description Logic Handbook. URL: http:// books.cambridge.org/ 0521781760. htm Bogachev, D. (n.d.), `TMAssert'. URL: http:// homepage.mac.com/ dmitryv/ TopicMaps/ TMRM/ TMAssert.pdf Description Logics Home Page (n.d.). URL: http:// dl.kr.org/ Fensel, D., Hendler, J. A., Lieberman, H. & Wahlster, W., eds (2003), Spinning the Semantic Web, The MIT Press. URL: http:// mitpress.mit.edu/ catalog/ item/ default.asp? tid=9182 Garshol, L. M. (2004-07-22), `A proposed founda-tional model for Topic Maps'. URL: http:// www.jtc1sc34.org/ repository/ 0529.htm Garshol, L. M. & Barta, R. (2003), `JTC1/SC34: TMQL requirements'. URL: http:// www.isotopicmaps.org/ tmql/ tmqlreqs.html Kipp, N. A. (2003), `A mathematical formalism for the Topic Maps reference model'. http://www.isotopicmaps.org/tmrm/0441.htm. URL: http:// www.isotopicmaps.org/ tmrm/ 0441.htm Newcomb, S. R., Hunting, S., Algermissen, J. & Durusau , P. (2003), `ISO/IEC JTC1/SC34, Topic Maps - reference model, editor's draft, revision 3.10'. URL: http:// www.isotopicmaps.org/ tmrm/ O. Lassila and K. Swick (1993), Resource Description Framework (RDF) model and syntax specification , Technical report, W3C, Camo AS. URL: http:// www.w3.org/ TR/ 1999/ REC-rdf-syntax-19990222.html Pepper, S. (1999), `Navigating haystacks, discovering needles', Markup Languages: Theory and Practice , Vol. 1 No. 4 . Pepper, S. (2000), `The TAO of Topic Maps'. URL: http:// www.gca.org/ papers/ xmleurope2000/ papers/ s11-01.html Sowa, J. (2000), Knowledge Representation: Logical, Philosophical and Computational Foundations, Brooks-Cole, Pacific Grove. 42

Semantic Web;Topic Maps;Knowledge Engineering